* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download Mapping the Brain
Synaptogenesis wikipedia , lookup
Neuroscience and intelligence wikipedia , lookup
Multielectrode array wikipedia , lookup
Cortical cooling wikipedia , lookup
Molecular neuroscience wikipedia , lookup
Subventricular zone wikipedia , lookup
Donald O. Hebb wikipedia , lookup
Neuromarketing wikipedia , lookup
Biochemistry of Alzheimer's disease wikipedia , lookup
Neuroesthetics wikipedia , lookup
Blood–brain barrier wikipedia , lookup
Single-unit recording wikipedia , lookup
Human multitasking wikipedia , lookup
Neural engineering wikipedia , lookup
Feature detection (nervous system) wikipedia , lookup
Selfish brain theory wikipedia , lookup
Artificial general intelligence wikipedia , lookup
Functional magnetic resonance imaging wikipedia , lookup
Clinical neurochemistry wikipedia , lookup
Development of the nervous system wikipedia , lookup
Synaptic gating wikipedia , lookup
Neuroeconomics wikipedia , lookup
Activity-dependent plasticity wikipedia , lookup
Neurotechnology wikipedia , lookup
Neurolinguistics wikipedia , lookup
Human brain wikipedia , lookup
Neurogenomics wikipedia , lookup
Neural correlates of consciousness wikipedia , lookup
Aging brain wikipedia , lookup
Optogenetics wikipedia , lookup
Brain Rules wikipedia , lookup
Mind uploading wikipedia , lookup
Neuroplasticity wikipedia , lookup
Brain morphometry wikipedia , lookup
Cognitive neuroscience wikipedia , lookup
Neuropsychology wikipedia , lookup
Haemodynamic response wikipedia , lookup
Channelrhodopsin wikipedia , lookup
Neurophilosophy wikipedia , lookup
Neuroinformatics wikipedia , lookup
Holonomic brain theory wikipedia , lookup
Nervous system network models wikipedia , lookup
Neuropsychopharmacology wikipedia , lookup
Metastability in the brain wikipedia , lookup
History of neuroimaging wikipedia , lookup
Focus on Mapping the Brain npg © 2013 Nature America, Inc. All rights reserved. W Visualization of functional connectivity in the human cerebral cortex based on magnetic resonance imaging data. Brain image by Joachim Böttger and Daniel Margulies (Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany) with compositing help by Tobias S. Hoffmann. Cover composition by Erin Dewalt. e are entering a new era in neuroscience in which technological development will allow us to obtain full anatomical, high-resolution renderings of entire brain circuits and to map the activity of ever larger cellular populations as an animal performs specific behaviors. Assembling anatomical, molecular and functional maps has the potential to greatly advance our understanding of how brains work. In this Focus, experts outline the technologies needed to obtain these maps and discuss what will be needed beyond them to understand brain function. In a Historical Perspective, Cornelia Bargmann and Eve Marder discuss what has been learned from invertebrate circuits whose connectivity patterns are known and what will be needed beyond anatomical maps to understand brain function in other organisms. In a Commentary, Jeff Lichtman and Joshua Morgan express their views about why obtaining detailed, high-resolution structural maps should be an essential part of this endeavor. To understand the deluge of data these maps will engender once generated, Olaf Sporns argues in another Commentary that data representation and modeling will be critical. Other experts discuss the newest technologies available for obtaining brain maps. Moritz Helmstaedter presents the state of the art and current challenges of electron microscopy–based circuit reconstruction. In three papers, the potential of using light to unveil the function and anatomy of brain circuits is presented. Karl Deisseroth and Kwanghun Chung discuss their newly developed method named CLARITY for rendering mammalian brains permeable to visible photons and molecules. Pavel Osten and Troy Margrie review light-microscopy methods available for large-scale anatomical tracing and discuss ways to integrate molecular identity, activity recording and anatomical information. In a Resource, Josh Sanes and colleagues present improved tools for mapping the mouse brain using the Brainbow technology. Finally, Michael Milham, Stan Colcombe and their colleagues review methods for functional and anatomical analysis of human brains at the macroscale. We are pleased to acknowledge the financial support of Carl Zeiss Microscopy, Hamamatsu Corporation, LaVision BioTec, TissueVision, Inc. and Chroma Technology Corp. Nature Methods carries sole responsibility for all editorial content and peer review. Contents 483 From the connectome to brain function C I Bargmann & E Marder 491 Making sense of brain network data O Sporns 494 Why not connectomics? J L Morgan & J W Lichtman 501Cellular-resolution connectomics: challenges of dense neural circuit reconstruction M Helmstaedter 508 CLARITY for mapping the nervous system K Chung & K Deisseroth 515 Mapping brain circuitry with a light microscope P Osten & T W Margrie 524 Imaging human connectomes at the macroscale R C Craddock, S Jbabdi, C-G Yan, J T Vogelstein, F X Castellanos, A Di Martino, C Kelly, K Heberlein, S Colcombe & M P Milham 540 Improved tools for the Brainbow toolbox Dawen Cai, K B Cohen, T Luo, J W Lichtman & J R Sanes Erika Pastrana Editor, Nature Methods Daniel Evanko Copy Editor Odelia Ghodsizadeh Carol Evangelista, Ivy Robles Focus Editor Erika Pastrana Publisher Stephanie Diment Managing Production Editor Renee Lucas Sponsorship David Bagshaw, Yvette Smith, Reya Silao Senior Copy Editor Irene Kaganman Production Editors Brandy Cafarella, Marketing Nazly De La Rosa nature methods | VOL.10 NO.6 | JUNE 2013 | 481 focus on mapping the brain historical perspective From the connectome to brain function npg © 2013 Nature America, Inc. All rights reserved. Cornelia I Bargmann1 & Eve Marder2,3 In this Historical Perspective, we ask what information is needed beyond connectivity diagrams to understand the function of nervous systems. Informed by invertebrate circuits whose connectivities are known, we highlight the importance of neuronal dynamics and neuromodulation, and the existence of parallel circuits. The vertebrate retina has these features in common with invertebrate circuits, suggesting that they are general across animals. Comparisons across these systems suggest approaches to study the functional organization of large circuits based on existing knowledge of small circuits. An animal’s behavior arises from the coordinated activity of many interconnected neurons—“many” meaning 302 for Caenorhabditis elegans, 20,000 for a mollusc, several hundred thousand for an insect or billions for humans. Determining the connectivity of these neurons, via combined anatomical and electrophysiological methods, has always been a part of neuroscience. As we were writing this, these ideas were being revisited from the perspective of massively parallel methods for dense reconstruction, or ‘connectomics’. One thread of this analysis involves the detailed, high-density mapping of point-to-point connections between neurons at synapses1–4. The specialized membrane structures and synaptic vesicles of synapses can be visualized with an electron microscope, and consequently dense reconstructions of nervous-system connectomes rely on electron microscopy of serial brain sections. In a complementary approach, detailed electrophysiological analysis shows how synapses and circuits function at high resolution, and is increasingly being applied to large numbers of interconnected neurons. The first approaches used to map complete circuits came from studies of the smaller nervous systems of invertebrates. In the 1960s and 1970s, systematic electrophysiological recordings from neurons in discrete ganglia enabled the identification of neuronal components of circuits that generate specific behaviors5–7. In association with the recordings of these individually recognizable, identified neurons, the cells were filled with dye to visualize their structures and projection patterns via light microscopy8–10. In some cases, electron microscopy was used to observe the anatomical synapses in these small circuits11–13. But until the publication of the heroic electron micros copy reconstruction of the full nervous system of C. elegans14 in the mid-1980s, it was unimaginable that the electron microscope could be used to determine circuit connectivity rather than providing ultrastructural detail to connectivity determined either with physiological or light microscopy–based anatomical methods. Recent advances in electron microscopy and image analysis have made it possible to scale up this ultrastructural approach: to serially section and reconstruct pieces of both vertebrate and invertebrate nervous systems, with the stated purpose of using detailed connectomes to reveal how these circuits work4,15–18. Such largescale projects will provide new anatomical data that will offer invaluable insights into the functional organization of the structures studied. An unbiased approach to data acquisition always reveals surprises and new insights. Moreover, because of the scope and size of these projects, such efforts will generate unprecedented amounts of data to be analyzed and understood. Here we ask what additional information is needed beyond connectivity diagrams to understand circuit function, informed by the invertebrate circuits whose connectivity is known. For the prototypical case, the complete C. elegans nervous system, the anatomical connectome was largely established over 25 years ago14. In a variety of other invertebrate preparations, connectivity was established using combinations of electrophysiological recordings and neuronal tracing 30– 40 years ago, which enabled researchers to generate a wiring map that incorporates activity information. Despite their different starting points from anatomy and electrophysiology, these two approaches have uncovered 1Howard Hughes Medical Institute, The Rockefeller University, New York, New York, USA. 2Volen Center, Brandeis University, Waltham, Massachusetts, USA. 3Department of Biology, Brandeis University, Waltham, Massachusetts, USA. Correspondence should be addressed to C.I.B. ([email protected]). Received 27 February; accepted 5 April; published online 30 may 2013; doi:10.1038/nmeth.2451 nature methods | VOL.10 NO.6 | JUNE 2013 | 483 npg © 2013 Nature America, Inc. All rights reserved. historical perspectiveFOCUS ON MAPPING THE BRAIN Figure 1 | Connectivity of two well-studied invertebrate circuits. (a) Connectivity diagram of the crab STG based on electrophysiological recordings. Red and blue background shading indicates neurons that are primarily part of the pyloric and gastric circuits, respectively. Purple shading indicates that some neurons switch between firing in pyloric and gastric time, and that there is no fixed boundary between the pyloric and gastric circuits. Yellow highlights two neurons that are both electrically coupled and reciprocally inhibitory. Green highlights one of many examples of neurons that are coupled both monosynaptically and polysynaptically. (b) The connectome of C. elegans, showing all 302 neurons and their chemical synapses but not their gap junctions. Each neuron has a three-letter name, often followed by a spatial designator. This topto-bottom arrangement (signal flow view) is arranged to reflect dominant information flow, which goes from sensory neurons (red) to interneurons (blue) to motor neurons (green). Reprinted from ref. 50. a AB PD LPG LP IC LG MG Electrical synapse GM Chemical inhibitory synapses PY Int1 VD DG AM b ASJL AIMR AINL ASIR ALMR AIML ASIL AWAR ASGL AWALAINR IL2R IL2L IL2VL IL2VR IL2DL IL2DR RIH ASEL PHAR ASJR AWBL AWBR ADFL CEPDR CEPVR ASERASGR AWCR AWCL ADFR BAGL CEPDL CEPVL URBR URXL PLNR PLNL ADEL ASKL ALNL PVQL ADLR ADLL HSNL similar principles and similar puzzles as to how circuit function arises from the component neurons and their interactions. What do functional and anatomical maps reveal? We begin with the connectivity diagram of the stomatogastric ganglion (STG) of the crab, Cancer borealis (Fig. 1a) and a graph of the connectome of C. elegans (Fig. 1b). In each case, the number of neurons is small, ~27 neurons or 302 neurons, respectively, but the number of synaptic connections is much larger; the neurons are extensively interconnected. The basic function of each circuit is known: to generate rhythmic stomach movements for the crab STG and to control locomotion behavior in response to sensory inputs for C. elegans. The intellectual strength of the STG system is the ability to relate neuronal connectivity to neuronal activity patterns; the complementary strength of C. elegans is the ability to relate neuronal connectivity to wholeanimal behavior. The STG contains motor neurons and interneurons that generate two rhythmic motor patterns19. The pyloric rhythm is an oscillating, triphasic motor pattern that is continuously active and depends on a set of electrically coupled pacemaker neurons. The gastric mill rhythm is episodically active and depends on descending modulatory inputs activated by sensory neurons for its generation19,20. Although these rhythms are easily studied separately, a close look at the STG connectivity diagram reveals that the neurons that conventionally are thought to be part of the pyloric circuit (neurons AB, PD, LP, PY, VD and IC) are highly interconnected with those conventionally thought part of the gastric mill circuit (neurons DG, GM, LPG, MG, LG and Int1) (AM is part of a third circuit that we will not discuss here). Indeed, many 484 | VOL.10 NO.6 | JUNE 2013 | nature methods PHCR ALML FLPR VC05AVM SDQL PDEL ASHR PVPR PLMR ASHL AVHR BDUL BDUR AVHL FLPL PVNL PVQR ADER PDER ASKR AVG RIFR SDQR RIFL HSNR ADAL AFDR RIR AFDL OLLL AUALBAGR URXRAUAR URYVL ALNR URYVR URBL PVT AIYL RMGR RIS AIAL RMGL ALA PVR URAVR ADAR AIAR AIZL URADR OLLR AIZR URAVL IL1VR RICL IL1R URADL AIYR URYDR IL1L SAAVL OLQVL IL1VL RICR AIBL OLQVR URYDL RIVR SMBVR RMFL RIAR RIAL OLQDR SAAVR IL1DL RIVL RIGL RIGR AIBR OLQDL SAADR AVKR SMBVL SMBDR IL1DR AVKL RMFR RMHR RMHL SAADL RIBL AVEL RMEL AVER RIBR RIPL SMBDL RIPR SIADR RIMR SIBDR RMER RIML SIADL RMEV SIBVL SIAVR RMDL RMED SIAVLSIBDL SIBVR RMDDR SMDVL RMDDL RMDR RMDVR SMDVR SMDDL RMDVL SMDDR PHAL PVM AVFL LUAR PVDR PHBR PVDL PHBL VC04 PQR PLMLPHCL LUAL PVWR PVNR AVFRAVJR PVPL AQR DVA PVWL PVCR DVC AVJL VB01 AVDR VA12 AVDL PDA VD11 PVCL DVB VC02 AVL AVBL AVBR VB08 VB11 AVAL AVAR VC03 DB01 VB09 RID VA07 VB07 VB02 VB06 VA02 DB02 VC01 AS11 DA09 DB03 VA09 DB04 PDB VA08 VB04 VA01 VA04 AS01 AS02 VD10VA03 VD13 VA11 AS09 AS06 DA01 DA03 DA04 VA06 AS03 DB05 DB06 AS05 DA05 DA08 DA02 VA05 AS04 VB10 DB07 SABD SABVR SABVL VD01 DA06 AS07 AS10 VD12 DA07 VA10 AS08 VD09 DD05 DD06 DD01 VD02 VD08 VD07 VD05 DD04DD02 DD03 VD03 VD04 VD06 STG neurons switch their activity between the two rhythms19, and the separation of the STG’s connectivity into two discrete circuits, although convenient for those who study the network, does not really capture the highly interconnected reality of the ganglion’s architecture. Like all nervous systems, the circuit has many chemical synapses, in which a presynaptic neuron releases a chemical neurotransmitter to activate receptors on the postsynaptic neuron. Chemical synapses can be inhibitory or excitatory depending on the nature of the receptor and associated ion channels; the chemical synapses among STG neurons are inhibitory. Additional connections are created by the widespread electrical synapses, mediated by direct cytoplasmic communication through gap junctions, through which current flows depending on the voltages of the coupled neurons. In the STG circuit, there are many instances of neurons that are connected by electrical synapses as well as by chemical inhibitory synapses (Fig. 1a). There are also many instances of neurons connected by reciprocal inhibition. These wiring motifs contribute to circuit properties that are not easily predictable. In addition, there are many ‘parallel pathways’ in which two neurons are connected via two or more synaptic routes, one direct route and additional indirect routes (Fig. 1a). The complexity of this connection map poses the essential question: are all synapses important, or are some only important under certain conditions (as appears to be the case)21? How do we understand the importance of synaptic connectivity patterns that seem to oppose each other, such as the common motif of electrical coupling between neurons that also inhibit each other? The C. elegans wiring diagram was assembled in the nearcomplete absence of prior functional information. It allowed an FOCUS ON MAPPING THE BRAIN Anterior touch ALM 9 classes of neurons Posterior touch AVM 9 classes of neurons historical perspective PLM AVD AVB PVC 6 classes of neurons VA VB Gap junction DA DB Chemical synapse Sensory neuron Interneuron Motor neuron Reversal Acceleration npg © 2013 Nature America, Inc. All rights reserved. Figure 2 | C. elegans neurons essential for avoidance of light touch. Inferred connections necessary for anterior and posterior touch avoidance are in purple and orange, respectively; other synapses are in black. The ‘essential’ synapses here shown in orange and purple comprise less than 10% of the output synapses of the mechanosensory neurons. Image based on ablation data from ref. 22. immediate classification of neurons into large classes: sensory neurons (with distinctive sensory dendrites and cilia), motor neurons (with neuromuscular junctions) and interneurons (a term that is used in C. elegans to describe any neuron that is not evidently sensory or motor, encompassing projection neurons and local neurons)14. In each group, neurons were subdivided into unique types with similar morphologies and connections, collapsing the wiring diagram from 302 neurons to 119 neuronal types. The flow of information through chemical synapses is predominantly from sensory to interneuron to motor neuron, with many parallel pathways linking neurons both directly and indirectly (as in the STG), as well as gap junctions that may form electrical synapses (~10% of all synapses). Most neurons are separated from each other by no more than two or three synaptic connections. The C. elegans map was immediately used to define neurons required for the touch-avoidance response, which is still the most completely characterized of the animal’s behaviors22. Light touch to the head elicits a reversal, and light touch to the tail elicits a forward acceleration. The neurons required for the touchavoidance response were identified by killing cells with a laser microbeam and assessing the behavioral repertoire of the worms. Guided by the wiring diagram, this analysis revealed essential mechanosensory neurons in the head and tail, key interneurons required to propagate information, and motor neurons required for forward and backward movement (Fig. 2). The success of this approach inspired similar analyses of chemosensory behaviors, foraging, egg-laying, feeding and more. At this point, over 60% of C. elegans neuron types have defined functions in one or more behaviors. This notable success, however, hides a surprising failure. For C. elegans, although we know what most of the neurons do, we do not know what most of the connections do, we do not know which chemical connections are excitatory or inhibitory, and we cannot easily predict which connections will be important from the wiring diagram. The problem is illustrated most simply by the classical touch-avoidance circuit22 (Fig. 2). The PLM sensory neurons in the tail are solely responsible for tail touch avoidance. PLM forms 31 synapses with 11 classes of neurons, but only one of those targets is essential for the behavior—an interneuron called PVC that is connected to PLM by just two gap junctions and two chemical synapses. An even greater mismatch between the number of synapses and their importance in behavior is seen in the avoidance of head touch, where just two of 58 synapses (again representing gap junctions) are the key link between the sensory neurons (ALM and AVM) and the essential interneuron (AVD). This general mismatch between the number of synapses and apparent functional importance has applied wherever C. elegans circuits have been defined. As a result, early guesses about how information might flow through the wiring diagram were largely incorrect. Clearly, the wiring diagram could generate hypotheses to test, but solving a circuit by anatomical inspection alone was not successful. We believe that anatomical inspection fails because each wiring diagram encodes many possible circuit outcomes. Parallel and antagonistic pathways complicate circuits Both of the wiring diagrams shown in Figure 1 are richly connected. In the STG, a large fraction of the synapses are electrical synapses. In some cases, the electrical synapses connect multiple copies of the same neuron, such as the two PD neurons in the STG. Notably, many electrical synapses connect neurons with different functions. Almost invariably, the combination of electrical and chemical synapses create ‘parallel pathways’, that is to say, multiple pathways by which neuron 1 can influence neuron 2 (Fig. 1a). For example, in the STG, the PD neuron inhibits the IC neuron through chemical synapses but also can influence the IC neuron via the electrical synapse from LP to IC. Parallel pathways such as those in the STG can be viewed as degenerate, as they create multiple mechanisms by which the network output can be switched between states23 (Fig. 3). A simulation study23 shows a simplified five-cell network of oscillating neurons coupled with electrical synapses and chemical inhibitory synapses. The f1 and f2 neurons are connected reciprocally by chemical inhibitory synapses, as are the s1 and s2 neurons. This type of wiring configuration, called a half-center oscillator, often but not universally causes the neurons to be rhythmically active in alternation 24. In this example, two different oscillating rhythms are generated, one fast and one slow. The hub neuron at the center of the network can be switched between firing in time with the fast f1 and f2 neurons to firing in time with the slow s1 and s2 neurons by three entirely different circuit mechanisms: changing the strength of the electrical synapses, changing the strength of the synapses between f1 and s1 onto the hub neuron, and changing the strength of the reciprocal inhibitory synapses linking f1 to f2 and s1 to s2 in the half-center oscillators. An example from the C. elegans connectome illustrates another twist of circuit logic: divergent circuits that start at a common point but result in different outcomes. In this example, gap junctions and chemical synapses from ADL sensory neurons generate opposite behavioral responses to a C. elegans pheromone (Fig. 4a). The chemical synapses drive avoidance of the pheromone, whereas the gap junctions stimulate a pheromone-regulated aggregation behavior25. Differing use of the chemical synapse subcircuit versus the gap junction subcircuit allows ADL to switch between these two opposing behaviors in different contexts. ADL illustrates the point that is not possible to ‘read’ a connectome if it is intrinsically ambiguous, encoding two different behaviors. Parallel and divergent systems of synapses are widespread features of invertebrate and vertebrate networks alike, and can be composed of sets of chemical synapses as well as sets of chemical nature methods | VOL.10 NO.6 | JUNE 2013 | 485 historical perspectiveFOCUS ON MAPPING THE BRAIN npg © 2013 Nature America, Inc. All rights reserved. Figure 3 | Similar changes in a b c d circuit dynamics can arise from f1 f1 f1 gsynB f1 three entirely different circuit gsynA g mechanisms. (a) Circuit diagram el s2 f2 hn s2 f2 hn s2 f2 hn s2 f2 hn (top) shows the ‘control’ condition gel in which f1 and f2 are firing gsynA gsynB s1 s1 s1 s1 in a fast rhythm (indicated by red shading) and the remaining f1 neurons are firing in a slow f2 rhythm (shaded in blue). hn is the hn hub neuron. In these diagrams, s2 electrical synapses are shown as s1 resistor symbols and chemical 100 mV 1 s inhibitory synapses with filled circuits. Traces (bottom) show the voltage waveforms of the five neurons. (b–d) Responses when the strength of the chemical synapses to the hub neuron (gsynA) was decreased (b), when the strength of the electrical synapses (gel) was decreased (c) and when the strength of the chemical synapses between f1 and f2 and between s1 and s2 (gsynB) was decreased (d). Image modified from ref. 23. and electrical synapses. To understand information flow, there will be no substitute for recording activity. The methods for monitoring neuronal activity have improved dramatically in recent years, with development of new multi-electrode recording techniques and a suite of genetically encoded indicators that can be used to measure calcium, voltage and synaptic release at cellular and subcellular levels. However, improved methods are needed to detect electrical synapses, which can also be difficult to see in electron micrographs. The regulation of electrical synapses by voltage, neuromodulation, phosphorylation and small mole cules is understudied26,27. A chemical method for measuring gap junctions, local activation of molecular fluorescent probes, is a promising new direction that should spawn innovation28. Neuromodulation reconfigures circuit properties Superimposed on the fast chemical synapses and electrical synapses in the wiring diagram are the neuromodulators—biogenic amines (serotonin, dopamine, norepinephrine and histamine) and neuropeptides (dozens to hundreds, depending on species)29. These molecules are often released together with a fast chemical transmitter near a synapse, but they can diffuse over a greater distance. Modulators also can be released from neuroendocrine cells that do not make defined synaptic contacts or can be delivered as hormones through the circulation. As a result, the targets of neuromodulation are invisible to the electron microscope. Signaling primarily through G protein–regulated biochemical processes rather than through ionotropic receptors, neuromodulators change neuronal functions over seconds to minutes, or even hours. Many years of work on the effects of neuromodulators on the STG have revealed that the functional connections that give rise to a specific circuit output are specified, or in fact ‘configured’, by the neuromodulatory environment29. Every synapse and every neuron in the STG is subject to modulation; the connectivity diagram by itself only establishes potential circuit configurations, whose availability and properties depend critically on which of many neuromodulators are present at a given moment29. Under some modulatory conditions, anatomically ‘present’ synaptic connections may be functionally silent, only to be strengthened under other modulatory conditions. Likewise, modulators can qualitatively alter the neurons’ intrinsic properties, transforming neurons from tonic spiking to those generating plateau potentials or bursts29. These effects of neuromodulators can activate or silence an entire circuit, change its frequency and/or the phase relationships of the motor patterns generated. C. elegans has over 100 different neuropeptides as well as biogenic amine neuromodulators. The integration of neuromodulation into its fast circuits appears to selectively enhance the use of particular connections at the expense of others. For example, a ‘hub-and-spoke’ circuit drives aggregation of C. elegans by coupling multiple sensory inputs through gap junctions with a common target neuron, RMG (Fig. 4b). Neuromodulation of RMG by the neuropeptide receptor NPR-1 effectively silences this gapjunction circuit, while sparing other functions of the input sensory neurons that are mediated through chemical synapses30. Neuromodulators are prominent in all nervous systems, and act as key mediators of motivational and emotional states such as Figure 4 | Two views of a multifunctional a b C9 pheromone C. elegans circuit. (a) Ambiguous circuitry Pheromone Oxygen of the ADL sensory neurons, which drive ADL ASK URX attraction avoidance AVA NPR-1 avoidance of the ascaroside pheromone C9 through chemical synapses onto multiple AIA Nociceptive Pheromone RMG RMG ADL ASH interneurons (right) but can also promote AVD avoidance avoidance aggregation (attraction toward pheromones) AIB through gap junctions with RMG (left). Image IL2 AWB modified from ref. 25. (b) Neuromodulation Pheromone Aggregation avoidance Aggregation separates overlapping circuits. Multiple sensory neurons form gap junctions with the RMG hub Gap junction Chemical synapse Sensory neuron Interneuron or motor neuron neurons and promote aggregation through this circuit, but each sensory neuron also has chemical synapses that can drive RMG-independent behaviors. The neuropeptide receptor NPR-1 inhibits RMG to suppress aggregation. Image modified from ref. 30. 486 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN npg © 2013 Nature America, Inc. All rights reserved. Neuronal dynamics shape the activity of circuits The existence of parallel circuits and neuromodulation means that connectivity alone does not provide adequate information to predict the physiological output of circuits. Even without these factors, the behavior of neurons over time is unpredictable from anatomy because neuronal behavior is sensitive to intrinsic channels and electrical properties that vary within and between cell types. Channels, synapses and biochemical processes interact to generate explicitly time-delimited features, or dynamics, in neurons and circuits. The importance of neuronal dynamics in circuit function can be seen most simply in a two-cell circuit (Fig. 5). Two isolated neurons from the STG that are not normally synaptically coupled were connected using the dynamic clamp, a computer-neuronal interface that allows a user to manipulate biological neurons with conductances that imitate ion channels and synaptic connections32. The neurons are connected reciprocally by dynamic clamp-created inhibitory synapses so that the neurons rhythmically alternate their activity24. The dynamic clamp allows the investigator to change the strength of the synapses as well as the amount of one of the membrane currents, hyperpolarizationactivated inward current (Ih)—either of which dramatically alters the period of the circuit oscillation (Fig. 5). Thus, a given wiring diagram can produce widely different dynamics with different sets of circuit parameters, and conversely, different circuit mechanisms can give rise to similar oscillation dynamics. Without knowing the strength and time course of the synaptic connections as well as the numbers and kinds of membrane currents in each of the neurons, it would not be possible to simply go from the wiring diagram to the dynamics of even two neurons. Synaptic connectivity alone does not sufficiently constrain a system. Understanding neuron-specific and circuit-specific dynamics will be essential to understanding mammalian circuits as well as invertebrate circuits. In some cases, unique dynamic properties are characteristics of particular cell types—for example, different classes of inhibitory cortical interneurons are distinguished as much by their dynamics as by their connectivity33. In other cases, neuronal dynamics are variable among similar cells or even within one cell type. For example, pyramidal neurons in specific areas of the cortex exhibit persistent activity associated with working memory34, and neurons in brainstem modulatory systems switch their properties between tonic and phasic firing modes depending 25 1 Period (s) sleep, arousal, stress, mood and pain. To understand their release and their effects on circuits, new methods are needed to monitor neuromodulation in vivo. Electrophysiology remains the best tool for characterizing the functional effects of neuromodulators but is low-throughput. Biochemical methods can be used to reveal the presence of neuromodulators in tissue or in bulk extracellular fluid but are less effective for detecting them near a particular synapse or release site. A new genetic method can be used to read out the neuromodulatory state directly by monitoring receptor activation but with a timeframe of hours, whereas endogenous modulation can change within minutes or seconds31. Progress is needed in all of these domains and beyond: there is a need to move from individual neurons and modulators to physiologically relevant modulatory states, which are likely to include multiple neuromodulators acting at many sites. historical perspective 2 Synaptic Ih 20 15 10 5 0 0 20 40 60 80 100 Conductance (nS) Synaptic conductance 60 nS 50 nS 80 nS 2 22 22 1 11 11 10 mV 2s lh conductance 80 nS 50 nS 30 nS 2 22 22 1 11 11 10 mV 2s Figure 5 | Changing either intrinsic neuronal properties or synaptic properties can alter network function. The dynamic clamp, a computerneuron interface (top) was used to vary either the strength of synaptic connections between two neurons (synaptic) or the amount of an intrinsic hyperpolarization-activated inward current (Ih), in one neuron as graphed (right). Traces (bottom) show the action potentials generated by the alternating, oscillating neuron pair as those properties were varied. Image modified from ref. 24. on behavioral states35. Finally, synaptic plasticity can occur on rapid timescales to strengthen and weaken synapses based on use, adding complexity to circuit-level dynamics36. Analyzing neuronal dynamics often requires the circuit to be simultaneously monitored and manipulated, as shown in the example of the dynamic clamp. Emerging techniques of optogenetics and pharmacogenetics can be combined with recording as well, but a limitation of all of these methods is that they act at the level of neurons or groups of neurons. To understand functional connectivity, it will be useful to develop methods to silence or activate specific channels and specific synaptic connections between two specified neurons, without affecting all other functions of the same cells. Vertebrate retina also has complex circuit properties What lessons will emerge as connectomes are scaled up from small-scale to large-scale circuits? Many features will be common to small and large circuits. Vertebrate circuits, like invertebrate circuits, have multiple cell types with nonuniform intrinsic properties, extensive and massively parallel synaptic connectivity, and neuromodulation. The balance of these components varies between animals and brain regions (the STG has more electrical synapses than most vertebrate brain regions; C. elegans uses mostly graded potentials instead of all-or-none action potentials), but in reality, the diversity of circuits in the vertebrate brain is at least as great as the difference between any one vertebrate region and any invertebrate circuit. The essential distinction we see in vertebrate brains is not a particular microcircuit property but their repeating structure (for example, the many cortical columns) and their enormous scale compared to the worm brain and the STG. nature methods | VOL.10 NO.6 | JUNE 2013 | 487 npg © 2013 Nature America, Inc. All rights reserved. historical perspectiveFOCUS ON MAPPING THE BRAIN Examination of the vertebrate retina has begun to reveal the relationships between performance of a large circuit and properties of a small circuit. The special power of studying the isolated retina is the ability to experimentally control visual input while simultaneously recording output—the spikes from retinal ganglion cells that project to the brain. The retina represents an intermediate degree of complexity with features of both a small circuit and a large circuit, and has been subject to the most complete anatomical and electrophysiological characterizations of any vertebrate brain region. Current connectomics studies of the retina, for example, include dense reconstruction of serialsection electron micrographs accompanied by analysis of the neurotransmitter phenotype and activity patterns of the reconstructed neurons3,17. The combination of structure and function, and a rich history of elegant experiments, make this the ideal system for understanding neural computations in detail. The retina contains millions of neurons that fall into five major neuronal classes (photoreceptors, horizontal cells, bipolar cells, amacrine cells and retinal ganglion cells), which are subdivided into about 60 discrete cell types37. Each of the 60 cell types is arrayed in a near-crystalline two-dimensional array, so that any pixel viewed by the retina is covered by at least one neuron of each cell type. Ultimately, information leaves the retina through the 20 classes of retinal ganglion cells, each of which is considered to be a parallel but partially overlapping processing stream. The first views of the retinal connectome show all of the properties that we highlight in small circuits: cellular complexity, extensive interconnectivity, parallel circuits with chemical and electrical synapses, and neuromodulation. The heterogeneity of the 60 retinal cell types is substantial, echoing the heterogeneity of individual neurons in the STG or C. elegans. Anatomically, some dendritic arbors cover only a tiny area of the visual field, but others arborize much more broadly. Their intrinsic physiological properties are also extremely diverse, with some neurons that spike (such as retinal ganglion cells), and many neurons that do not spike (such as photoreceptors and bipolar cells)37. There are even amacrine neurons that perform independent computations in different parts of their complex arbors38. Synaptic connections in the retina are extensive and diverse, and electron microscopy reconstructions have revealed many classes of synaptic connections that had not been observed in physiological studies3,17. There is a great variety of excitatory and inhibitory chemical synapses, and there are many electrical synapses, that all vary in their strength and their modification by experience. Both anatomical and physiological studies demonstrate that the retina, like small circuits, consists of many partly parallel circuits with overlapping elements. In particular, the retina operates over many orders of magnitude of light intensity, and the properties of its circuits change with its visual inputs. Within a few seconds in a new visual environment, retinal ganglion cells shift their properties to encode relevant features of light intensity, contrast and motion, drawing on different features of the network39. Subsets of retinal ganglion cells change their weighting of center and surround inputs in a switch-like fashion as light levels change40. A brief period of visual stimulation can even reverse the apparent direction-selectivity of retinal ganglion cells41. Finally, neuromodulation has a role in retinal processing that reshapes visual circuits. Dopamine is released from a subset of 488 | VOL.10 NO.6 | JUNE 2013 | nature methods amacrine cells around dawn, under the control of acute light stimuli and circadian rhythm42, and acts on cells throughout the retina to switch them from properties appropriate to night vision to day vision. The photoreceptors themselves are modulated, and their coupling through gap junctions decreases to increase their resolution but reduce their sensitivity. Downstream of the rod photoreceptors, which dominate night vision, dopamine closes the gap junctions between rod bipolar cells and AII amacrine cells, effectively diminishing rod input to the retinal ganglion cell output of the retina. What differences are there between small and large circuits? The sheer size of the retina shows a sharp transition compared to the size of the STG and the worm brain, and the level of analysis moves from single cells to cell classes. Understanding a single pixel is not sufficient to understand the retina, and here the properties of simple and complex circuits diverge. For example, long-range communication allows groups of retinal cells to perform computations that a single cell cannot. Wide-field cells such as starburst amacrine cells can make judgments about motion that no single-pixel neuron could make but can then feed that information into narrow-field single-pixel neurons to bias their properties. The scaling from fine resolution to broad resolution and back again emerges from the diversity of spatial scales across the structure of the retina. Circuits interact to generate behavior The entire nervous system is connected, but reductionist neuroscientists invariably focus on pieces of nervous systems. The value of these simplified systems should not let us forget that behavior emerges from the nervous system as a whole. At the moment, obtaining the connectomes of even small parts of the vertebrate nervous system is a heroic task. However, establishing the detailed pattern of connectivity for a small part of the nervous system may not be sufficient to understand how that piece functions in its full context. By parceling out small regions, one invariably loses information about the long-range connections to and from that area. The extent to which long-range connectivity clouds our understanding of connectomes will vary. For example, the vertebrate retina is anatomically isolated, functionally coherent and lacks recurrent feedback synapses from other brain areas that are prominent in most other parts of the central nervous system. We might imagine the retina as a two-dimensional circuit, whereas most vertebrate circuits are three-dimensional; new principles will certainly arise from connectomes that include recurrent inputs. In the amygdala, for example, the intermixing of multiple cell types with different long-range inputs and outputs would preclude a meaningful understanding based on local anatomy alone43. Choosing well among brain regions, and combining connectomes with molecular and functional information about the same cells, as is being done in the vertebrate retina3,17, will lead to the most informative results. How can we ‘solve’ the brain? As we look to ways that other neural systems may be characterized with similar power to the three described here, we can draw certain lessons. One is that precise circuit mapping and specific neuron identification have had great importance for unifying structural and functional data from different laboratories. Extending this idea, other systems may not have individually npg © 2013 Nature America, Inc. All rights reserved. FOCUS ON MAPPING THE BRAIN named cells, but all nervous systems have cell types distinguished by anatomy, connectivity and molecular profile that can serve as the basis of a common vocabulary. Improvements in molecular staining methods will only increase the power of a connectome anchored in cellular identity. We can also see that connectomic endeavors will need to be supplemented by experiments that monitor, manipulate and model circuit activity. Monitoring and manipulation of circuit function have been considered above. To complement and inform these experimental approaches, the third step is to develop models that describe how a system’s output results from the interactions of its components. There is a tension between the desire to study abstract models that are amenable to precise mathematical analysis and the desire to study models with sufficient biological realism to represent the system’s underlying structures and functions. In small circuits, it is now possible to construct models and families of models that can be quite instructive44. In C. elegans, a few testable models emerged directly from analyses of anatomy. One was the concept of a motif, a set of connection patterns between three or four neurons that are over-represented in the wiring diagram compared to the statistical expectation based on individual connections45. Perhaps these motifs perform a canonical computation, or a few canonical computations, so that solving a few of them effectively solves a larger piece of the diagram. But how should we approach building models of large networks without generating models that are as difficult to understand as the biological systems that motivated and inspired them? The beauty of the connectome is its precision and specificity, but it is hard to imagine useful network models that implement all of the details of cell-to-cell connectivity obtained with the electron microscope, when building such models would require enormous numbers of assumptions about other circuit parameters, and these parameters are likely to change in different modulatory states. So we face a conundrum: the new anatomical data will be instructive, but it is not yet obvious what kinds of models will best reveal the implications of these data for how circuits actually work. We are in the midst of a fascinating international debate about whether it is the right time to embark on a ‘big science’ project to monitor and model large brain regions. There are those who argue that we are now at the point at which investments in large-scale projects will considerably advance the field in ways not possible by a distributed small-lab approach46–48. Big science works best when the goals of a project are well-defined and when the outcomes can be easily recognized. Both were true about the human genome project, but neither is true, yet, about large-scale attempts to understand the brain. Moreover, this is well-recognized, and all of the proponents of large-scale initiatives are acutely aware of the necessity to develop new technology48 and of the extraordinary complexity of biological systems49. That said, the largest challenge we face in future attempts to understand the dynamics of large circuits is not in collecting the data: what is most needed are new methods that allow our human brains to understand what we find. Humans are notoriously bad at understanding multiple nonlinear processes, although we excel at pattern recognition. Somehow, we have to turn the enormous data sets that are already starting to be generated into a form we can analyze and think about. Otherwise, we will be doomed to creating a machine that will understand the human brain better than we can! historical perspective Acknowledgments We thank M. Meister for sharing his knowledge of the retina. C.I.B. is funded by the Howard Hughes Medical Institute. Research in the Marder laboratory relevant to this piece is funded by the US National Institutes of Health (NS17813, NS 81013 and MH46742). COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests. Reprints and permissions information is available online at http://www.nature. com/reprints/index.html. 1. Bock, D.D. et al. Network anatomy and in vivo physiology of visual cortical neurons. Nature 471, 177–182 (2011). 2. Briggman, K.L. & Bock, D.D. Volume electron microscopy for neuronal circuit reconstruction. Curr. Opin. Neurobiol. 22, 154–161 (2012). 3. Briggman, K.L., Helmstaedter, M. & Denk, W. Wiring specificity in the direction-selectivity circuit of the retina. Nature 471, 183–188 (2011). 4. Kleinfeld, D. et al. Large-scale automated histology in the pursuit of connectomes. J. Neurosci. 31, 16125–16138 (2011). 5. Burrows, M. Monosynaptic connexions between wing stretch receptors and flight motoneurones of the locust. J. Exp. Biol. 62, 189–219 (1975). 6. Fentress, J.C. Simpler Networks and Behavior (Sinauer Associates, 1976). 7. Getting, P.A., Lennard, P.R. & Hume, R.I. Central pattern generator mediating swimming in Tritonia. I. Identification and synaptic interactions. J. Neurophysiol. 44, 151–164 (1980). 8. Stretton, A.O. & Kravitz, E.A. Neuronal geometry: determination with a technique of intracellular dye injection. Science 162, 132–134 (1968). 9. Remler, M., Selverston, A. & Kennedy, D. Lateral giant fibers of cray fish: location of somata by dye injection. Science 162, 281–283 (1968). 10. Goodman, C.S. Isogenic grasshoppers: genetic variability in the morphology of identified neurons. J. Comp. Neurol. 182, 681–705 (1978). 11. Maynard, E.A. Electron microscopy of the stomatogastric ganglion in the lobster, Homarus americanus. Tissue Cell 3, 137–160 (1971). 12. King, D.G. Organization of crustacean neuropil. I. Patterns of synaptic connections in lobster stomatogastric ganglion. J. Neurocytol. 5, 207–237 (1976). 13. King, D.G. Organization of crustacean neuropil. II. Distribution of synaptic contacts on identified motor neurons in lobster stomatogastric ganglion. J. Neurocytol. 5, 239–266 (1976). 14. White, J.G., Southgate, E., Thomson, J.N. & Brenner, S. The structure of the nervous system of the nematode Caenorhabditis elegans. Phil. Trans. R. Soc. Lond. B 314, 1–340 (1986). 15. Seung, H.S. Reading the book of memory: sparse sampling versus dense mapping of connectomes. Neuron 62, 17–29 (2009). 16. Meinertzhagen, I.A. & Lee, C.H. The genetic analysis of functional connectomics in Drosophila. Adv. Genet. 80, 99–151 (2012). 17. Anderson, J.R. et al. Exploring the retinal connectome. Mol. Vis. 17, 355–379 (2011). 18. Lu, J., Tapia, J.C., White, O.L. & Lichtman, J.W. The interscutularis muscle connectome. PLoS Biol. 7, e32 (2009). 19. Marder, E. & Bucher, D. Understanding circuit dynamics using the stomatogastric nervous system of lobsters and crabs. Annu. Rev. Physiol. 69, 291–316 (2007). 20. Nusbaum, M.P. & Beenhakker, M.P. A small-systems approach to motor pattern generation. Nature 417, 343–350 (2002). 21. Thirumalai, V., Prinz, A.A., Johnson, C.D. & Marder, E. Red pigment concentrating hormone strongly enhances the strength of the feedback to the pyloric rhythm oscillator but has little effect on pyloric rhythm period. J. Neurophysiol. 95, 1762–1770 (2006). 22. Chalfie, M. et al. The neural circuit for touch sensitivity in Caenorhabditis elegans. J. Neurosci. 5, 956–964 (1985). 23. Gutierrez, G.J., O’Leary, T. & Marder, E. Multiple mechanisms switch an electrically coupled, synaptically inhibited neuron between competing rhythmic oscillators. Neuron 77, 845–858 (2013). 24. Sharp, A.A., Skinner, F.K. & Marder, E. Mechanisms of oscillation in dynamic clamp constructed two-cell half-center circuits. J. Neurophysiol. 76, 867–883 (1996). 25. Jang, H. et al. Neuromodulatory state and sex specify alternative behaviors through antagonistic synaptic pathways in C. elegans. Neuron 75, 585–592 (2012). 26. Pereda, A.E. et al. Gap junction-mediated electrical transmission: regulatory mechanisms and plasticity. Biochim. Biophys. Acta 1828, 134–146 (2013). 27. Neyton, J. & Trautmann, A. Physiological modulation of gap junction permeability. J. Exp. Biol. 124, 93–114 (1986). nature methods | VOL.10 NO.6 | JUNE 2013 | 489 28. Guo, Y.M. et al. Imaging dynamic cell-cell junctional coupling in vivo using Trojan-LAMP. Nat. Methods 5, 835–841 (2008). 29. Marder, E. Neuromodulation of neuronal circuits: back to the future. Neuron 76, 1–11 (2012). 30. Macosko, E.Z. et al. A hub-and-spoke circuit drives pheromone attraction and social behaviour in C. elegans. Nature 458, 1171–1175 (2009). 31. Inagaki, H.K. et al. Visualizing neuromodulation in vivo: TANGO-mapping of dopamine signaling reveals appetite control of sugar sensing. Cell 148, 583–595 (2012). 32. Sharp, A.A., O’Neil, M.B., Abbott, L.F. & Marder, E. Dynamic clamp: computergenerated conductances in real neurons. J. Neurophysiol. 69, 992–995 (1993). 33. Kawaguchi, Y. & Kondo, S. Parvalbumin, somatostatin and cholecystokinin as chemical markers for specific GABAergic interneuron types in the rat frontal cortex. J. Neurocytol. 31, 277–287 (2002). 34. Egorov, A.V., Hamam, B.N., Fransen, E., Hasselmo, M.E. & Alonso, A.A. Graded persistent activity in entorhinal cortex neurons. Nature 420, 173–178 (2002). 35. Aston-Jones, G. & Cohen, J.D. An integrative theory of locus coeruleusnorepinephrine function: adaptive gain and optimal performance. Annu. Rev. Neurosci. 28, 403–450 (2005). 36. Zucker, R.S. & Regehr, W.G. Short-term synaptic plasticity. Annu. Rev. Physiol. 64, 355–405 (2002). 37. Masland, R.H. The neuronal organization of the retina. Neuron 76, 266–280 (2012). 38. Grimes, W.N., Zhang, J., Graydon, C.W., Kachar, B. & Diamond, J.S. Retinal parallel processors: more than 100 independent microcircuits operate within a single interneuron. Neuron 65, 873–885 (2010). npg © 2013 Nature America, Inc. All rights reserved. historical perspectiveFOCUS ON MAPPING THE BRAIN 490 | VOL.10 NO.6 | JUNE 2013 | nature methods 39. Hosoya, T., Baccus, S.A. & Meister, M. Dynamic predictive coding by the retina. Nature 436, 71–77 (2005). 40. Farrow, K. et al. Ambient illumination toggles a neuronal circuit switch in the retina and visual perception at cone threshold. Neuron 78, 325–338 (2013). 41. Rivlin-Etzion, M., Wei, W. & Feller, M.B. Visual stimulation reverses the directional preference of direction-selective retinal ganglion cells. Neuron 76, 518–525 (2012). 42. Witkovsky, P. Dopamine and retinal function. Doc. Ophthalmol. 108, 17–40 (2004). 43. Ciocchi, S. et al. Encoding of conditioned fear in central amygdala inhibitory circuits. Nature 468, 277–282 (2010). 44. Marder, E. & Taylor, A.L. Multiple models to capture the variability in biological neurons and networks. Nat. Neurosci. 14, 133–138 (2011). 45. Milo, R. et al. Network motifs: simple building blocks of complex networks. Science 298, 824–827 (2002). 46. Markram, H. The blue brain project. Nat. Rev. Neurosci. 7, 153–160 (2006). 47. Koch, C. & Reid, R.C. Neuroscience: Observatories of the mind. Nature 483, 397–398 (2012). 48. Alivisatos, A.P. et al. The brain activity map project and the challenge of functional connectomics. Neuron 74, 970–974 (2012). 49. Koch, C. Systems biology. Modular biological complexity. Science 337, 531–532 (2012). 50. Varshney, L.R., Chen, B.L., Paniagua, E., Hall, D.H. & Chklovskii, D.B. Structural properties of the Caenorhabditis elegans neuronal network. PLoS Comput. Biol. 7, e1001066 (2011). focus on mapping the brain commentary Making sense of brain network data Olaf Sporns npg © 2013 Nature America, Inc. All rights reserved. New methods for mapping synaptic connections and recording neural signals generate rich and complex data on the structure and dynamics of brain networks. Making sense of these data will require a concerted effort directed at data analysis and reduction as well as computational modeling. New empirical methodologies deliver an ever increasing amount of data on patterns of synaptic connectivity. For any given nervous system, the complete map of its neural components and their synaptic interconnections corresponds to the connectome1. The connectome delivers a description of structural brain connectivity, in popular parlance often referred to as a ‘wiring diagram’. This wiring diagram may describe brain connectivity at different scales, capturing synaptic connections among single neurons or axonal projections among brain regions. An important distinction must be made between such connectome data and the highly variable and dynamic patterns of neural activity and interactions in the connectome, also called ‘functional connectivity’. Dynamic patterns of functional connectivity are distinct from connectome maps as they are not based on synaptic connections but on statistical measures derived from neuronal recordings, ranging from simple correlation to sophisticated inferences of causal dependence. Despite these important differences between structural and functional connectivity, both can be represented as networks comprising a set of discrete nodes and edges (Fig. 1), corresponding to neural elements and their pairwise associations. Creating accurate maps of brain networks presents many methodological challenges. For example, the construction of connectome data from observations of micrometer-scale anatomical patterns among individual neurons2 or from imaging of millimeter-scale projections among neural populations or regions 3 poses challenges regarding the Olaf Sporns is in the Department of Psychological and Brain Sciences, Indiana University, Bloomington, Indiana, USA. e-mail: [email protected] precision, sensitivity and reproducibility of network maps, partly because of various methodological biases. Consequently, much attention in the area of brain networks has been devoted to addressing potential problems and pitfalls in data acquisition, regarding the reconstruction and inference of connectional anatomy as well as the measurement of dynamic neural interactions. Clearly, progress critically depends on methodological developments that increase the completeness and accuracy of empirically measured network maps. But the scientific challenges with regard to the study of brain networks do not end once data are acquired. The sheer volume and complexity of these data requires sophisticated new approaches to enable efficient data integration and sharing via public neuroinformatics resources4,5. And another key question looms on the horizon: how can we use brain network data to gain new insight into fundamental neurobiological structures and processes? In this Commentary, I argue that the shift toward an explicitly networkoriented approach to studying brain function requires the development of new strategies for the analysis and modeling of brain network data. Shift toward networks The rise of genomics at the end of the 20th century ushered in a period of profound and still ongoing change in the biomedical sciences. New technological developments such as the capability to sequence whole genomes and to create comprehensive inventories of biomolecules and their chemical interactions have not only transformed the empirical study of biological systems, but also radically altered our conceptual understanding of complex biological processes. Complex Nodes Hubs Modules Edges Figure 1 | Simple network. Schematic illustrates nodes and edges of the network, and its community structure, with network modules and hubs highlighted. biological functions—once thought to depend on the expression of single genes or on the action of single molecules—are now more commonly thought to depend on large ensembles of biological components interacting in complex signaling, metabolic and gene-regulatory networks. The genomic revolution has given rise to a new discipline, systems biology, dedicated to unraveling the structure and dynamics of molecular and cellular networks that underpin all aspects of biological function6. Accomplishing this task requires tackling ‘big data’ with new computational tools and theoretical models. With the arrival of new experimental methods for the comprehensive mapping of neural structures and for large-scale recording of neural activity both in single cells and in populations, neuroscience appears poised to embark on its own ‘omics’ revolution7. Connectomics has emerged as a major focus of a new ‘systems biology of the brain’, centering on the study of anatomical networks across multiple scales as well as the complex nature methods | VOL.10 NO.6 | JUNE 2013 | 491 npg © 2013 Nature America, Inc. All rights reserved. focus on mapping the brain dynamics these networks generate when becoming active. In parallel to recent developments driven by modern genomics and systems biology, the aim of connectomics is to disclose the architecture of structural brain networks and to explain the mechanisms by which these structural networks shape and constrain brain function. New methods for mapping the layout and operation of brain networks deliver huge volumes of ‘big data’ that are shared in public repositories and neuroinformatics databases. These data pose fundamental challenges for analysis and modeling, and will require extensive data integration across multiple domains of brain and behavior. Network theory has become vitally important in this endeavor. By relying on networks, connectomics follows a path similar to the one already taken in systems biology, harnessing some of the same opportunities as well as encountering some of the same challenges. Researchers in both fields continue to grapple with methodological problems concerning the accurate and complete mapping of system interactions, the efficient representation and sharing of big data sets, and the application of large-scale computational models. Analysis of brain network data Brain networks are built from empirical data that are extraordinarily rich and complex. There is ongoing debate concerning the proper definition of network edges and nodes. Structural brain network data usually comprise multiple measures of connectional anatomy, the neurobiological interpretation of which is not straightforward, as is the case for inferences of connection weight or strength from synaptic morphology8, or from axonal microstructure or myelination status9. Functional brain network data derived from time-series recordings of neuronal activity can be expressed in many ways, from simply recording linear statistical dependencies (for example, Pearson correlations) to sophisticated inferences on dynamic causal interactions10. Which of these structural and dynamic network measures best represents the underlying neurobiological reality remains unclear11,12. Progress in this area is important because the initial definition of nodes and edges can have profound consequences for how brain networks are configured and interpreted. Of particular importance for revealing the organization of complex networks are methods for detection of network communities or modules13 (Fig. 1). Generally speaking, 492 | VOL.10 NO.6 | JUNE 2013 | nature methods modules correspond to sets of network elements that are densely linked with each other and less so with other elements in the network (that is, other modules). Module detection is an appealing strategy for analyses of brain network data sets for a host of reasons: (i) many studies have shown that brain network modules correspond to prominent functional subdivisions, for example, so-called resting-state networks reliably encountered in spontaneous and task-evoked brain dynamics14; (ii) the concept of modularity has validity and can be productively applied across both structural and functional networks, thus enabling comparisons across the two domains; (iii) the topography and relations among modules provide important means for characterizing individual differences or developmental patterns in network organization; (iv) modular decomposition offers a low-dimensional description of complex network data, which is useful for purposes of data reduction and compression; (v) modules help to define the roles of nodes and edges in the global network topology, for example, as connectors or bridges that crosslink different communities15; and (vi) modularity is thought to promote robust network function16. Modularity may also become an important tool for joining brain network data across multiple scales. For example, modularity should allow the identification of structurally and functionally coherent populations of neurons in cellular-scale data, which may then be aggregated into larger-scale network communities at regional and systems levels. Establishing links between the connectome on the one side and the dynamic patterns of neural activity and the interactions on the other requires relating data across the two different domains of structural and functional networks. To reiterate the difference between structural and functional network data, edges in structural networks refer to aspects of the physical infrastructure of brain connectivity, that is, synaptic connections jointly comprising the ‘wiring diagram’, whereas edges in functional networks reflect aspects of statistical dependencies among neuronal time series, corresponding to simple correlation or covariance or more sophisticated measures of nonlinear coupling or causal dependence. These differences in edge definition entail clear differences in the way structural and functional network data should be analyzed and interpreted. For example, communication processes governing the exchange of information along connections do not commentary occur in functional networks; instead functional networks are a manifestation of these processes as they unfold in the structural connectome. Therefore, studies aiming to identify important network nodes (so-called ‘hubs’) or to probe the robustness of a network after lesion (involving deletion of nodes and/or edges) are more appropriately carried out on structural networks than on functional networks. Whereas the identification of hubs in structural networks is relatively straightforward, it remains somewhat ambiguous in the context of functional networks17. Brain networks are dynamic and change across time, on both long time scales (for example, plasticity and remodeling during development and learning) as well as short time scales (for example, fast remodeling of dendritic spines in structural connectivity or rapid edge dynamics in functional connectivity, the latter in part resulting from neuromodulation18). Hence, capturing characteristic features of how networks change through time will be as important as characterizing the structure of nodes and edges at each instant in time. In this area, there is currently a dearth of data-driven approaches, in part because network studies in other disciplines (social networks, protein networks and others) are only just beginning to explicitly take network dynamics into account. The problem is particularly pressing in the analysis of functional networks where the topology and strengths of dynamic interactions are continually and rapidly modulated by endogenous changes in state and exogenous changes in input and task. Studies of the human brain with electrophysiological or imaging methods consistently show that functional brain networks rapidly reconfigure with changes in the momentary demands of the environment19 and exhibit nonstationarity during ‘rest’20. Development of methods to characterize these network dynamics is becoming critical in the area of data representation and analysis. Recently, far-ranging proposals for cataloging all functional activity maps a given brain can generate (referred to as the brain’s ‘functional connectome’) have been made 21 , culminating in the Brain Research through Advancing Innovative Neurotechnologies (BRAIN) initiative announced in April 2013 by President Barack Obama. Brain networks are central to achieving the goals of this ambitious plan. Dynamic recordings of functional activity maps can naturally and efficiently be represented in the form of functional npg © 2013 Nature America, Inc. All rights reserved. commentary networks. Such networks, especially if recorded at single-cell and millisecond resolution, will be highly dynamic and variable, thus posing difficult computational and analytic challenges. To address these challenges, the BRAIN initiative must include intense efforts directed not only at data analysis and modeling but also toward extending fundamental network science and theory to the brain. Such efforts have proven productive in other areas of neuroscience. Catalogs of functional activation maps recorded with noninvasive neuroimaging methods in the human brain are available as public repositories that can be data-mined and analyzed online22,23. These databases have enabled important insights regarding the link between functional activations, networks of coactivation patterns and functional connectivity. And these functional network maps have been linked to various domains of behavior and cognition. Growing role of computational models One of the key rationales behind studies of protein networks is that patterns of proteinprotein interactions can be predictive of a protein’s functional roles as well as its involvement in disease states24. The notion that structure can predict function also informs one of the chief goals of connectomics, which is to furnish network models that can bridge brain structure and function1,7. A large body of evidence suggests that neuronal structure, and especially connectional anatomy, is crucial for endowing neural elements with specific functional properties. This structurefunction relationship holds across scales and species, and it can even predict individual variations in brain responses and behavior25. The structure-function relationship opens a crucial role for computational models that build on structural data to predict brain dynamics and associated functional attributes. Typically, such models consist of neuronal units whose dynamics are described by differential equations based on biophysics (for example, ionic membrane conductance). These neuronal units are then interconnected via a structural network of synaptic links that allows units to communicate and influence each other’s dynamic state. Neural communication may involve noisy signal transmission and conduction delays, and neural activity is typically tracked by measuring activation level, firing rate or membrane potential across time. Hence, such models deliver time-series data that can be analyzed using methods very similar to those used for focus on mapping the brain the analysis of empirically measured brain dynamics. Connectome data on structural connectivity can supply the coupling matrix for such computational models, and initial results suggest that these models can partly reproduce the spatial pattern and temporal dynamics of empirically observed functional networks26. First applications of such computational network models have already provided considerable theoretical insight into the network basis of functional connectivity in the resting brain27. As such computational models become more realistic and sophisticated, they may also become important tools for explaining empirical observations and predicting new findings. A highly ambitious project to build a working model of a human brain incorporating cellular and subcellular detail has commenced28, and structural connectome data coming from both micro and macro scales will likely be an important ingredient. A parallel effort to construct a ‘virtual brain’ builds on whole-brain connectome data sets to generate complex neural dynamics whose spatiotemporal signatures can be compared against empirical brain recordings29, with the explicit aim to create a computational tool that can be of clinical benefit. Simulations of brain activity in human patients, informed by mapping their individual connectome, may be just over the horizon. The confluence of connectome mapping and computational modeling once again parallels ongoing developments in genomics and systems biology, where computational models have become an integral component of many research projects. The idea that computer models can help understand the workings of a cell has a long history, and the availability of comprehensive data on cell structure, genes and gene products has now put the construction of a working model of a cell within reach30. As is the case with the virtual cell, building a virtual brain will open a wealth of new possibilities for testing and manipulating brain networks at all levels and probing for architectural features that are necessary for efficient and flexible network function. As the emerging field of connectomics continues to mature, computational models will become indispensable tools for guiding empirical studies of brain networks. Conclusion The promise of concerted efforts to map the brain’s structural and functional connections is that a renewed focus on the brain as a com- plex networked system can deliver fundamentally new insights and knowledge. This promise can only be fulfilled with the help of new approaches to data analysis and reduction, driven by network theory and computational modeling. As increasing amounts of brain network data become available, there will be a growing need for neuroscientists to work more closely with colleagues in other disciplines who are tackling complex technological, social and biological systems. The adoption of a network perspective, the development of new ways to analyze and describe brain connectivity, and the deployment of sophisticated integrative models are all important steps toward making sense of brain network data. ACKNOWLEDGMENTS Supported by the J.S. McDonnell Foundation. COMPETING FINANCIAL INTERESTS The author declares no competing financial interests. 1. Sporns, O., Tononi, G. & Kötter, R. PLoS Comput. Biol. 1, 245–251 (2005). 2. Denk, W., Briggman, K.L. & Helmstadter, M. Nat. Rev. Neurosci. 13, 351–358 (2012). 3. Van Essen, D.C. & Ugurbil, K. Neuroimage 62, 1299–1310 (2012). 4. Akil, H., Martone, M.E. & van Essen, D.C. Science 331, 708–712 (2011). 5. Marcus, D.S. et al. Front. Neuroinform. 5, 4 (2011). 6. Hood, L. et al. Science 306, 640–643 (2004). 7. Sporns, O. Discovering the Human Connectome. (MIT Press, 2012). 8. Bourne, J.N. & Harris, K.M. Curr. Opin. Neurobiol. 22, 372–382 (2012). 9. Jbabdi, S. & Johansen-Berg, H. Brain Connect. 1, 169–183 (2011). 10. Friston, K.J. Brain Connect. 1, 13–36 (2011). 11. Smith, S.M. Neuroimage 62, 1257–1266 (2012). 12. Wig, G.S., Schlaggar, B.L. & Petersen, S.E. Ann. NY Acad. Sci. 1224, 126–146 (2011). 13. Fortunato, S. Phys. Rep. 486, 75–174 (2010). 14. Power, J.D. et al. Neuron 72, 665–678 (2011). 15. Guimerà, R. & Amaral, L.A.N. Nature 433, 895– 900 (2005). 16. Kashtan, N. & Alon, U. Proc. Natl. Acad. Sci. USA 102, 13773–13778 (2005). 17. Zuo, X.N. et al. Cereb. Cortex 22, 1862–1875 (2012). 18. Bargmann, C.I. Bioessays 34, 458–465 (2012). 19. Bassett, D.S. et al. Proc. Natl. Acad. Sci. USA 108, 7641–7646 (2011). 20. Jones, D.T. et al. PLoS ONE 7, e39731 (2012). 21. Alivisatos, A.P. et al. Neuron 74, 970–974 (2012). 22. Fox, P.T. et al. Hum. Brain Mapp. 25, 185–198 (2005). 23. Yarkoni, T. et al. Nat. Methods 8, 665–670 (2011). 24. Vazquez, A. et al. Nat. Biotechnol. 21, 697–700 (2003). 25. Saygin, Z.M. et al. Nat. Neurosci. 15, 321–327 (2012). 26. Adachi, Y. et al. Cereb. Cortex 22, 1586–1592 (2012). 27. Deco, G., Jirsa, V.K. & McIntosh, A.R. Nat. Rev. Neurosci. 12, 43–56 (2011). 28. Markram, H. Sci. Am. 306, 50–55 (2012). 29. Jirsa, V.K. et al. Arch. Ital. Biol. 148, 189–205 (2010). 30. Karr, J.R. et al. Cell 150, 389–401 (2012). nature methods | VOL.10 NO.6 | JUNE 2013 | 493 commentary FOCUS ON MAPPING THE BRAIN Why not connectomics? Joshua L Morgan & Jeff W Lichtman npg © 2013 Nature America, Inc. All rights reserved. Opinions diverge on whether mapping the synaptic connectivity of the brain is a good idea. Here we argue that albeit their limitations, such maps will reveal essential characteristics of neural circuits that would otherwise be inaccessible. Neuroscience is in its heyday: large initiatives in Europe1 and the United States2 are getting under way. The annual neuroscience meetings, with their tens of thousands of attendees, feel like small cities. The number of neuroscience research papers published more than doubled over the last two decades, overtaking those from fields such as biochemistry, molecular biology and cell biology3. Given the number of momentous advances we hear about, surely must we not be on the threshold of knowing how the healthy brain works and how to fix it when it does not? Alas, the brain remains a tough nut to crack. No other organ system is associated with as long a list of incurable diseases. Worse still, for many common nervous system illnesses there is not only no cure but no clear idea of what is wrong. Few psychiatric illnesses, learning disorders or even severe pain syndromes such as migraine have pathognomonic signs: no blood tests, radiographic or electrophysiological findings, or even brain biopsies enable diagnoses. This predicament is unlike that for other organ systems, where disease is nearly always associated with tissue pathological and/or biochemical signs. A patient may come for help because of pain in their belly, but the physician discovers the cellular or molecular cause of the pain before a treatment regimen begins. Why is this so different for the brain? Those of us with an interest in a deeper understanding of the structure of the brain, Joshua L. Morgan and Jeff W. Lichtman are in the Department of Molecular and Cellular Biology and The Center for Brain Science, Harvard University, Cambridge, Massachusetts, USA. e-mail: [email protected] or [email protected] 494 | VOL.10 NO.6 | JUNE 2013 | nature methods not surprisingly, see the problem in structural terms. The nervous system is a physical tissue that is quite unlike other organ systems4. For one thing, it contains far more types of cells. The retina alone has more than 50 different kinds of cells5, whereas the liver has closer to five. Second, neurons, by virtue of their complicated geometry, connect (via synapses) to many more cellular partners than the limited associations of immediately adjacent cells in other organs. Third, the neuronal contacts give rise to directional circuits that have no analogs in other tissues. Fourth, the fine structure of these neural circuits is quite diverse and differs from the marked structural redundancy in other organs where the same multicellular motif (for example, the renal nephron) is iterated many times. Fifth, perhaps the most intriguing difference between neural tissue and other organs is that the cellular structure of neural tissue is a product of both genetic instruction and experience. Thus, the structure of each of our nervous systems is personalized by our own set of experiences. The special structural features of the nervous system are likely the reason why it is more difficult to understand than other organs are. We contend, however, that a more complete rendering of neural circuit synaptic connectivity (that is, connectomics6) would go a long way to solving this problem. With this understanding, diseases that manifest as abnormalities of behavior, thought or learning, or as pain might become as clearly linked to underlying pathological structure, as is the case for diseases in other organ systems. Knowing what is wrong is a good first step to finding a solution. For some proponents of connectomics, the potential clinical payoffs provide suf- ficient justification for such studies. But not everyone agrees. First, connectomics can require an industrialized effort that is akin to initiatives that allowed genomics to flourish. In light of present severe budget limitations, connectomics might be ill-advised. Moreover, some have argued that pursuing connectomics would be a waste of money, even if it were free. They have argued that anatomical maps fundamentally do not reveal how the brain works. Below we address some of these arguments. Top ten arguments against connectomics Number ten: circuit structure is different from circuit function. One argument put forward is that the nervous system’s macroscopic functions (that is, behaviors) are derived directly from the functional (that is, electrical) properties of the neurons rather than the anatomical connections between neurons. Hence it is the relationship between the firing patterns of action potentials of neurons and function that is the key to bridge the gap between the cells of the nervous system and behavior. The recent proposal of an initiative to get a ‘brain activity map’ reflects this view, as do many initiatives for Ca2+ imaging from ever larger numbers of nerve cells. Although no one doubts that neural connections underlie signaling between nerve cells, the focus on neuronal function (action potentials) is based in part on the belief that the structural wiring details per se are insufficient to derive the firing patterns. Part of the problem is that the spiking properties of a neuron come not only from the electrical signals they receive from their presynaptic partners but also from the distribution and characteristics of their FOCUS ON MAPPING THE BRAIN npg © 2013 Nature America, Inc. All rights reserved. intrinsic voltage-gated channels. Although in principle connectomics can reveal all the presynaptic partners, distributions of ion channels would not be easily revealed by a connectional map. Furthermore, even if one knew the molecular identities and precise location of all the voltage-gated channels in every neuron (determined by very sophisticated immunological labeling, for example), this information would still not be sufficient to provide the activity patterns underlying behavior. The reason is that behavior is often driven by the particular trains of action potentials set up by sensory inputs. This sensory experience is extrinsic to the nervous system and hence inaccessible from just looking at a connectional map. Response. Brains can encode experiences and learned skills in a form that persists for decades or longer. The physical instantiation of such stable traces of activity is not known, but it seems likely to us that they are embodied in the same way intrinsic behaviors (such as reflexes) are: that is, in the specific pattern of connections between nerve cells. In this view, experience alters connections between nerve cells to record a memory for later recall. Both the sensory experience that lays down a memory and its later recall are indeed trains of action potentials, but in-between, and persisting for long periods, is a stable physical structural entity that holds that memory. In this sense, a map of all the things the brain has put to memory is found in the structure—the connectional map. An ‘activity map’ of the brain that only shows trains of action potentials would certainly be an incomplete map, as most behaviors and memories will not be visible in any finite recording session. Decoding the way experience via electrical activity becomes stably embedded in physical neuronal networks is the unmet challenge that connectomics attempts to solve. In our view, more challenging perhaps will be ferreting out the relationships between neural circuits and behaviors that are intrinsic and unlearned. In most animals, the behavioral repertoire is mostly genetically encoded. Such inherited circuits may have arbitrary features that came about accidentally at some instant in an animal’s evolution. We therefore think that genetically encoded behaviors could be instantiated in quite diverse ways compared to a more limited number of rules governing experience-dependent formation of circuits. From this perspective, the structural under- pinnings of a behavior of Caenorhabditis elegans may actually be more difficult to decode than the circuit underlying a learned behavior in a mammal. Number nine: signals without synapses and synapses without signals. There is abundant evidence that neurotransmitter released at a synapse is not always restricted to the ‘intended’ postsynaptic target. Spillover of glutamate from excitatory synapses has been shown to affect nearby postsynaptic sites7. It is also well known that neuronal activity can affect the behavior of nearby glial cells, which can then convey this activity to other glial cells through junctions between glia. In all brains, there are also many signals that pass between brain cells and other organ systems via hormones. Steroids originating in the adrenal cortex have effects on brain function as do those from sex organs. Growth hormone, thyroid hormone, insulin, leptin and many others also affect brain function. Furthermore, some forms of neurotransmitter release do not rely on classical synapses8. Lastly, neurons often release peptides that act over large areas, using volume transmission as opposed to restricted transmission at adjacent synapses9. All of the aforementioned examples show that brain function is profoundly affected by chemical cues in ways that would be difficult or impossible to infer from anatomical wiring diagrams. Conversely, it is generally accepted that a substantial fraction of excitatory synapses can be structurally present but functionally silent10. These synapses can be switched on via a Hebbian learning step. Obviously in the absence of knowing which synapses are silent, a connectome provides a distorted picture of the functionally useful connections in the brain. Response. It is true that that a map of synaptic connectivity is not identical to a map of the signaling pathways of the brain. In terms of signals without synapses, there is no denying the importance of the brain’s chemical milieu on behavior. The ability of pharmacological agents to rapidly induce sleep, tranquility, excitement, hallucinations and so on means that the behavioral state can be dramatically altered probably without any modification to the connectome. Therefore, we should be careful not to confuse the goals of connectomics with the aims of maps of the activity patterns commentary in a nervous system. The connectome certainly would not be a map of the behavioral state at the moment the brain is preserved. We think the connectome would be much more than that, as it could provide the entire behavioral capability of the brain. But extracting such information may require detailing an unprecedented amount of anatomical data. This anatomically stringent view of connectomics contrasts with interesting recent ideas about determining the connections between brain cells without information about anatomy or physiology11. A connectional matrix minus the anatomy information, of course, can reveal neither the sites of spillover nor the proximities necessary for the spread of peptides, whereas an anatomical connectome could reveal less direct paths for neurotransmitter and neuropeptide signaling. However, the hormonal milieu via blood flow or cerebrospinal fluid would still remain invisible unless the effects of diffusible factors on behavior had a structural correlate12. There are also anatomical approaches to deal with the problem of synapses without signals. Some of the earliest evidence for a class of silent synapses came from serial electron microscopy reconstructions13, and we imagine that such information might become available for glutaminergic synapses as well. If silent synapses are structurally different from transmitting ones by virtue of the neurotransmitter receptor, then in principle that difference can be revealed. Immunolabeling for AMPA-type glutamate receptors should solve this problem. Number eight: ‘junk’ synapses. Given the trillions of synapses in a human brain, most individual synapses are functionally negligible. The brain is probably organized such that the precise number and identity of synapses are unimportant, and what really matters is the general likelihood of connectivity. Thus, some fraction of the synaptic connections could rightly be called ‘junk’: they have no functional role but have so little cost that such scraps persist. If junk synapses are common, then going to the trouble of itemizing all the synaptic connections is both a waste of time and a distraction from more important functional questions. Response. Our view is that it is premature to assign an influence to individual synapses in wiring diagrams when so few complete wiring diagrams have been described. The lesson from genomics is that many noncodnature methods | VOL.10 NO.6 | JUNE 2013 | 495 commentary npg © 2013 Nature America, Inc. All rights reserved. ing sequences that initially were thought to be unimportant and sometimes called junk turned out to have functional importance. Moreover, we are not convinced that the nervous system tolerates large numbers of synapses that are not serving a useful function. In fact, the synapses observed in the adult may reflect the small percentage of synapses that survive an extended period of synapse elimination during development14. Number seven: same structure, many functions. Connectomics might be tractable if each behavior was partitioned to a different circuit element (that is, one circuit, one behavior). However, good evidence shows that multiple sensations or behaviors use the same neurons in different ways15. Neurons can rapidly switch their functional roles in response to chemical signals such as peptides, other neuromodulators or activity levels. Because of this flexibility, it is really not possible to assign a neuron to a function without knowing the behavioral state—something that is invisible in connectomics. Response. Without an understanding of the physiological responses of neurons to various chemical messengers such as peptides, we agree it will be difficult to define the roles of neurons that switch their function depending on the chemical milieu. The extent to which this kind of switching is a general feature of nervous systems is not yet known. But whatever the case, it is important to emphasize one limitation of connectomics: it is not a replacement for insights gained by physiological or pharmacological studies. Rather, connectomics may associate specific physiological phenomena with specific neural-circuit motifs so that the next time that motif is observed in the same tissue, it will signify a physiological process without the requirement of repeating the physiological analysis each time. Because in many nervous systems the same neuron types are duplicated many thousands or millions of times, it seems likely that the same motifs will be repeated multiple times. Thus, a little physiology may go a long way. Number six: same function, many structures What if there is great variability in the circuits that give rise to a single behavioral output? Without a stereotyped pattern of neural circuits that underlie a particu496 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN a 33% connected b Figure 1 | Potential results from two approaches to studying circuit connectivity. (a) Probing pairs of neurons in multiple subjects determines the probability that neighboring neurons are connected. (b) Connectome of the same tissue reveals network motifs. lar function, it will be impossible to relate structure to function. Response. In mammals, at least, variability seems to be the rule. Even the pattern of nerve-muscle connections in the same muscle is quite different from one instantiation to the next16. However, variability in connections does not necessarily mean that the result is incomprehensible. In muscle, for example, there is a skewed distribution of the size of motor units that is the same in each muscle even though the exact location and branching pattern of each axon is unique. Thus, each instantiation appears to be a variation on the same common theme, just as every chess match is different, but they all obey the same rules. One of the reasons we believe connectomics is necessary, is that it is the only way to derive network principles despite variability in the connectivity of individual neurons. Indeed, deriving network principles without connectomics may be nearly impossible in some cases. Without connectomics, neural circuit diagrams can only be assembled by identifying connected pairs of cells in many different subjects. This approach, although accurate as far as it goes, cannot reveal wiring rules as simple as ‘cell A only connects to cell C when cell B connects to cell C’. Such conditional rules have been revealed from fortuitous circuits that lend themselves to easy analysis. For example, triad synapses in the lateral geniculate nucleus allow the connectivity of cell A, B and C to all be seen in a single electron microscopy section as all the synapses are adjacent17. If more complicated patterns of conditional connectivity exist (as well they might18,19), then there is no way to find these without reconstructing the entire circuit and including all the cellular elements (Fig. 1). Number five: statistical synapses should suffice. Statistical mechanics was a great advance in physics: when many similar events occur at the same time (such as the collisions of molecules in a gas), the behavior of the system can be predicted with accuracy without ascertaining the behavior of each and every particle’s trajectory. Given the trillions of synapses in a brain, is there really any alternative except to take a statistical approach to synaptic connectivity? Members of the European Union’s initiative on Future and Emerging Technologies have decided, for example, to fund the large ‘Human Brain Project’. In this case, the connectional associations will be determined by characterizing “...the morphologies of different types of neuron present in different regions of the human brain. Combined with modeling, the results would enable the project to predict a large proportion of the short-range connectivity between neurons, without measuring the connectivity experimentally”1. That is, they will create the wiring diagram without having to go to the trouble of obtaining an anatomic connectome. Response. Because of the heterogeneous nature of brain tissue, statistical approaches to studying neural-circuit wiring are, of course, much more difficult than those for a homogeneous gas. For a statistical model to capture how brains actually generate behavior, ideally it would need to include all the neuronal connectivity motifs. Unfortunately, the number of ways the dozens of different cell types are interconnected, and especially how these connections are contingent on the full cohort of synaptic partners of each cell, is presently unknown. But the advocates of statistical approaches to studying connectivity are not unduly concerned because they believe that models that combine a limited amount of connectivity data with the right set of FOCUS ON MAPPING THE BRAIN npg © 2013 Nature America, Inc. All rights reserved. learning rules will produce fully functional neural circuits. We, however, think the only way to know whether the results of this strategy are actually consistent with a biological brain, is to compare the predicted wiring to an actual connectome. This is why we submit that ‘measuring the connectivity experimentally’ is a good idea: it provides all the circuit motifs; with analysis, it may also provide the physical instantiation of the learning rules (see below) and is a direct path to building a working model of the brain. Number four: the mind is no match for the complexity of the brain. Might it be the case that ‘connectomisists’ have bit off more than they can chew? Might it be that the brain’s structural complexity far outstrips the complexity of the thoughts that emanate from even the best and the brightest brains? Although some may hope that there are organizing principles and regularities that will permit substantial compression of the connectional information into a simpler framework, is there really any biological reason such regularities are inevitable? Could the most succinct description that embodies all that a brain does be the brain itself? If so, then there is little point in mapping the connections in great detail because in the best case, one would end up with a description that would be as complicated and intractable as the brain itself. Response. The proximate goal of connectomics is to generate detailed renderings of the connections between neurons in large volumes of the brain. The purposes that such connectional maps could be put to are numerous. Comparisons between healthy and diseased brains might point to the physical underpinnings of psychiatric diseases. Comparisons between young and old brains might provide an understanding of what kinds of network changes are associated with development and aging. Comparisons between human and other primate brains may provide insight into what underlies intelligence. So even if ‘understanding’ the brain is not likely to be an early triumph of connectomics, there are potentially many fundamental things that this kind of data will reveal. To put it perhaps a bit too bluntly, we think that understanding big data, may be overrated as a goal. Connectomes can immediately provide insights even if understanding lags. Number three: merely descriptive neuroanatomy, just more expensive. The only hope for curing diseases and uncovering the ways brains work is to get insights into underlying mechanisms. Mechanistic insights come from experiments that test hypotheses by manipulating variables. There is of course a time and a place for descriptive studies in a field, and neuroscience owes Cajal a great debt for his landmark description of the cellular underpinnings of neural tissue, but the time for mere description is long past. Indeed, connectomics sounds modern but is really just a throwback: it is neuroanatomy gussied up with a fancy new word and ultra-expensive machines. Connectomics harkens to a time when description was all we could do. We can do more now. Response. The Hubble telescope, archaeology, the human genome and much of structural biology are also merely descriptive, but few would argue that they have not provided fundamental insights. The hope of descriptive approaches is that they provide specific data that lead us to general hypotheses (inductive reasoning). They can reveal associations and frameworks that were not readily apparent before. In a variety of fields, big-data initiatives are being used to challenge the sacrosanct ‘hypothesis first’ world view of scientific investigation. Number two: not much was learned from the connectomes we have. The 10-year effort to generate the connectome of C. elegans20 is widely cited, but it has had less utility than had been imagined originally. No one can claim that the relationship between structure and function has been settled in this very small nervous system. In some respects, connectomes distract from more mechanistic analyses because they reveal many more synaptic interconnections between neurons than would seem necessary. Given the struggle to make sense of the mind of a worm with about 300 neurons, it is very unlikely that we will get anywhere trying to fathom a mammalian nervous system that could be roughly a billion times bigger. Speaking of big brains, there is already a big Human Connectome project21, so why bother with maps of mice and flies? Response. At the very least the worm connectome has been helpful in constraining circuit analysis by providing for each neu- commentary ron the list of upstream and downstream neurons. This is no doubt beneficial, having been the starting point for understanding the neural pathways for touch-induced movements22, for example. But the worm connectome may not be as useful as initially imagined because of its surprising complexity. The high level of interconnectivity among the 300 neurons—a revelation in itself—does not lend itself to easy analysis. The interconnectedness of C. elegans neurons potentially provides for a much more adaptable and diverse behavioral repertoire than one with a simpler wiring diagram but at the expense of easy comprehension by humans. A particular technical limitation of the worm’s connectome is that its inhibitory and excitatory connections cannot be differentiated in electron microscopy images (unlike the situation for mammals), reducing the overall power of the wiring diagram to reveal new concepts. In addition, there are particular challenges in interpreting a map of connectivity in a highly differentiated nervous system where each neuron has many subcellular compartments, where all the signaling is due to local potentials, and evolution has specialized nearly every cell into a unique type. This specification contrasts to the millions of cells of a single class found in many parts of the (perhaps less evolved) vertebrate nervous system. Thus, the worm’s nervous system may use a fundamentally different strategy than the one predominating in much larger nervous systems in which cell types reflect populations of neurons whose connectivity is organized by experience. As to the Human Connectome project, this is an important and ambitious multiinstitution study to gain information about the organization of the brains of many individuals including many twins. One of its goals is to map the pathways that project between various brain regions (‘projectomics’). However, none of its many goals relate to describing the synaptic connectivity of the brain. Number one. “If you want to Google on connectome and look [it] up, you can see some gorgeous pictures that are being made of the wiring diagram of the human brain, showing you how the wires move from front to back and up and down, side to side. But again, it’s a static picture. It’d be like, you know, taking your laptop and prying the top off and staring at the parts inside, you’d nature methods | VOL.10 NO.6 | JUNE 2013 | 497 commentary npg © 2013 Nature America, Inc. All rights reserved. be able to say, yeah, this is connected to that, but you wouldn’t know how it worked” (Francis Collins, NPR Science Friday; 5 April 2013). Response. Maybe picking an argument in this case is not so wise, but prying the lid off a computer to see how everything is connected does not seem like such a bad idea, at least as a first step. Without knowing the parts list and what is connected to what, how could one ever really know how it works? In our view, the ideal brain-imaging technology would provide both a complete map of synaptic connectivity as well as a complete map of the activity of all neurons and synapses in real time during normal behaviors. Even better, would be to do this in a human being who can report on their thoughts while behaving. Unfortunately, we are a long way from such technologies; so what do we do in the meantime? We think the computer analogy gives us a lead. The computer can be turned off and on without losing much data because the instructions that make it work are embedded in its ‘static’ physicality. A deep understanding of how information is stably stored in the structure of hard discs, the input and output wires of each chip, the physical structures that explain the working of those chips and so on would be enormously helpful in making sense of a computer. Might the same be said of nervous systems? We think that the static connectivity of the brain has embedded within it much of the ‘instruction book’ that guides the billions of impulses through networks of neurons to ultimately generate outward behavior. We see a connection between the relationship between the connectome and the brain’s functional properties to not only computers but also to the way the static human genome encodes much of how an organism works. Structural defects in the genome that give rise to a range of diseases are perhaps a good analogy for defects in wiring that give rise to diseases of brain function. In each case, tracing the causal linkages between the static structure (be it genotype or connectotype) and the ultimate phenotype (be it cancer or schizophrenia) is difficult. But this difficulty has not deterred cancer biologists from seeking the ultimate causes (the physical genes) that affect the likelihood of cancer, and we think should not deter neuroscientists from 498 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN a b Figure 2 | Two types of wiring diagrams. (a) Schematic of mammalian retina drawn by Santiago Ramón y Cajal in 1901. Image courtesy of J. De Carlos. (b) Dense connectivity of the Twitter network of David Rodrigues as rendered by D. Rodrigues (http://www.flickr.com/photos/ dr/2048034334/). seeking the physical wiring abnormalities that underlie brain disorders, even if this requires high resolution connectomics. In both situations, the structural data is overwhelming but nonetheless holds fundamental truths. To quote Francis Collins referring to the human genome project: “When you have for the first time in front of you this 3.1-billion-letter instruction book that conveys all kinds of information and all kinds of mystery about humankind, you can’t survey that going through page after page without a sense of awe. I can’t help but look at those pages and have a vague sense that this is giving me a glimpse of God’s mind”23. We likewise feel that the static maps of the brain may engender awe and it might not be so unrealistic to hope that in staring into such a map we might get a glimpse of the human mind. Why we still want to do connectomics The large (but we are sure still incomplete) list of arguments above might seem discouraging enough to put a damper on all but the most irrationally enthusiastic proponents of connectomics. In addition to the reasons we provided in our rebuttals, we remain committed to pursuing connectomics because of the absence of one fundamental kind of information about nervous systems that we do not think is attainable in any other way. What Cajal did not do for neuroscience. Thanks to both his genius and the extraordinary power of the sparse labeling of the Golgi stain, Cajal developed a view about neurons that has had a dominating effect on the way in which we imagine neural circuits to work. Schematic textbook diagrams of retinal circuits, cortical circuits, cerebellar circuits, among many others, provide stu- dents with a sense of which cell types connect with one another. These diagrams also serve as a model of how information flows through a circuit of cells because many of these diagrams include arrows that harken back to the arrows found in the original Cajal diagrams (Fig. 2a). The sense one might take from these kinds of diagrams is that this kind of structural class-wise connectivity information explains the way a neural circuit works. One may quibble over the extent to which the details of these connectional maps have been worked out, but the implication is that knowing this kind of neuronal-class connectivity is sufficient to get some understanding about how neurons process information. In certain cases, this view is probably correct. In an animal in which the majority of neurons are each genetically unique and each has a stereotyped connectivity, connections of individual neurons are the same thing as a class-wise wiring diagram (because to a first approximation each cell is its own class). Thus, one can imagine a detailed map describing exactly which cells are connected, and that map would be exactly the same as what actually exists in the animal. But in other cases these diagrams are quite different from the actual connectional array. They lack both quantitative information and ignore the different wiring patterns of the cells of the same class. Thus, a retinal schematic diagram shows what classes of cells are interconnected but does not show critical details such as how many different amacrine cells converge on each retinal ganglion cell or which particular amacrine cells are making those connections. A cortical schematic diagram does not show whether two interconnected npg © 2013 Nature America, Inc. All rights reserved. FOCUS ON MAPPING THE BRAIN pyramidal cells have a common third input. A cerebellar schematic diagram does not show whether the cohort of parallel fibers connected to a given Purkinje cell predicts which parallel fibers will connect with adjacent Purkinje cells. This kind of information is, however, being addressed by new generation of connectomic circuit maps that reveal connectivity between dozens or hundreds of cells in a single piece of tissue17,24–26. These new efforts and several earlier attempts 20,27, are beginning to shift the paradigm from class-wise connectivity questions to questions about the particular pattern of connections in a cell class. Variability in the connections of cells of the same class, is the essential feature that is lacking in classical wiring diagrams but that exists in actual neural circuits. Because the vertebrate brain is mostly composed of many copies of each of many cell types (unique neurons such as the Mauthner cell in teleosts are quite exceptional), this aspect of neural circuit organization cannot be ignored. In sum, Cajal’s use of sparse labeling did not provide a means for coming to grips with the way redundant populations are used in circuits. Connectomics, however, does. Engramics? In 1904, Richard Semon, a German evolutionary biologist, coined the term ‘engram’ as the physical memory trace that is somehow embedded in an organism after an experience. He was motivated to think about the physicality of memory in an attempt to formulate a means of inheritance of acquired characteristics 28. Although this Lamarckian notion is refuted for gene-based inheritance, there is little doubt that humans do acquire information (by learning) during their lives that affects their behaviors and then pass this information on to their children (among others) to alter their behaviors in nongenetic ways. A longstanding goal of neuroscience is to determine the physical basis of an engram. An interesting feature of the highly redundant populations of neurons in mammalian nervous systems is that they undergo dramatic changes in their connectivity in early postnatal life. Because these alterations are activity-dependent and transform diffusely connected networks into many distinct subnetworks, this process could be the physical underpinnings of memory14. Regardless of the mechanism, we wonder whether eventually a subfield of neuroscience will exist that is devoted to understanding the encoding and decoding of experience-based changes in neural networks. Engramics would certainly require a hefty amount of connectomics along with sophisticated analysis and ultimately simulations to test the idea that a particular neural network encodes a particular memory. What is ahead for connectomics? Assuming that connectomics has a future, several mostly technical obstacles will need attention. Statistical and analytic tools. In many ways, maps of complete connectomes might look less like the canonical wiring diagrams of Cajal and more like modern renderings of social networks (Fig. 2b). How to compare and analyze these connectional graphs of the brain is a new challenge that will require new mathematical techniques for analyzing graphs, new statistical tests to compare one circuit with another and finally a cadre of neuroscientists who will find mining big data appealing. Ultimately, however, these tools will only be useful if connectomes can be produced easily. Faster data collection. There are two reasons to collect connectomics data faster. First, the production of connectomes must eventually be easy enough so that multiple experimental conditions can be compared and findings can be replicated. Second, the variety of questions that can be asked increases as volumes enlarge and multiple samples can be processed. If a complete map of every synapse in a human brain is wanted, a serious obstacle is the fact that at current speeds it would take 10 million years to complete. How fast can we go? The best technology currently available for producing connectomics data sets is high-throughput electron microscopy. Large-scale electron microscopy data sets are now being produced that are about 300 cubic micrometers. For many circuits, this size is large enough to encompass multiple complete axonal and dendritic arbors. There are a number of ways that this technology could be modified to produce a 10–50 fold increase in data-acquisition rates in the next few years. However, the commentary current bottleneck for obtaining connectomics data is not image acquisition but image segmentation29. Many connectomics data sets rely on a human painting voxels on a computer screen. One potential speedup is crowdsourcing (http://eyewire.org/). Alternatively learning algorithms can use human segmentations to guide computer efforts30. At the moment the results still require substantial human editing. All of these directions are being pursued, with competitions between groups stimulating new approaches to this daunting problem (http://brainiac.mit.edu/ SNEMI3D/). Ironically, much of this effort in image analysis is to make machines do what human brains do easily—something that we might better understand once connectomics analysis of the visual system is complete. Proteome meets the connectome. The structural mapping of connections will only be useful if the cell type of each of the neurons involved in a circuit is known. Serial reconstruction of a neuron’s shape and location will in some cases be sufficient for the type of neuron to be identified. But it is likely that there is heterogeneity among neurons that look alike. Fortunately the past few decades have seen an explosion in knowledge of the molecular classes of nerve cells31. What remains to be done is finding ways to insert that kind of information into mapping studies. Correlative fluorescence immunostaining with electron microscopy is one approach32. Other techniques are certainly on the horizon. Conclusion Despite the many arguments against undertaking a connectomics analysis, we think it must be done for three reasons. First, the conditional patterns of synaptic connectivity generated during development and by experience are inaccessible to techniques that sample from only a few cells at a time. Second, neuroscientists cannot claim to understand brains as long as the network level of brain organization is uncharted; without this detailed information, neuronal physiology is connected to systems physiology by a black box. Third, there is a high likelihood that such exploration will reveal unexpected properties by shining a light (or electrons) on this most mysterious tissue. nature methods | VOL.10 NO.6 | JUNE 2013 | 499 commentary ACKNOWLEDGMENTS Our work was supported by a Conte Center grant (US National Institute of Mental Health), the US National Institutes of Health, the Gatsby Charitable Trust and Center for Brain Science Harvard University. We thank D. Rodrigues for use of his rendering of his Twitter network and J. De Carlos for Cajal’s drawing of the retina. COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests. npg © 2013 Nature America, Inc. All rights reserved. 1. Markram, H. et al. A Report to the European Commission. <http://www.humanbrainproject. eu/files/HBP_flagship.pdf> (April 2012). 2. Collins, F. & Prabhakar, A. The White House Blog <http://www.whitehouse.gov/ blog/2013/04/02/brain-initiative-challengesresearchers-unlock-mysteries-human-mind> (2 April 2013). 3. Pautasso, M. Sustainability 4, 3234–3247 (2012). 4. Lichtman, J.W. & Denk, W. Science 334, 618– 623 (2011). 5. Masland, R.H. Curr. Opin. Neurobiol. 11, 431– 436 (2001). 500 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN 6. Lichtman, J.W. & Sanes, J.R. Curr. Opin. Neurobiol. 18, 346–353 (2008). 7. Diamond, J.S. Nat. Neurosci. 5, 291–292 (2002). 8. Oláh, S. et al. Nature 461, 1278–1281 (2009). 9. Bargmann, C.I. Bioessays 34, 458–465 (2012). 10. Kerchner, G. & Nicoll, R. Nat. Rev. Neurosci. 9, 813–825 (2008). 11. Zador, A.M. et al. PLoS Biol. 10, e1001411 (2012). 12. Butcher, A.J. et al. J. Biol. Chem. 286, 11506– 11518 (2011). 13. Jahromi, S.S. & Atwood, H.L. J. Cell Biol. 63, 599–613 (1974). 14. Lichtman, J.W. & Colman, H. Neuron 25, 269– 278 (2000). 15. Weimann, J.M. & Marder, E. Curr. Biol. 4, 896– 902 (1994). 16. Lu, J., Tapia, J.C., White, O.L. & Lichtman, J.W. PLoS Biol. 7, 13 (2009). 17. Sherman, S.M. & Guillery, R.W. J. Neurophysiol. 76, 1367–1395 (1996). 18. Seung, H.S. Neuron 62, 17–29 (2009). 19. Song, S., Sjöström, P.J., Reigl, M., Nelson, S. & Chklovskii, D.B. PLoS Biol. 3, e68 (2005). 20. White, J.G., Southgate, E., Thomson, J.N. & Brenner, S. Phil. Trans. R. Soc. Lond. B Biol. Sci. 314, 1–340 (1986). 21. Sporns, O., Tononi, G. & Kötter, R. PLoS Comput. Biol. 1, e42 (2005). 22. Chalfie, M., Sulston, J.E., Thomson, J.N. & White, G. J. Neurosci. 5, 956–964 (1985). 23. Swinford, S. The Sunday Times of London (11 June 2006). 24. Anderson, J.R. et al. Mol. Vis. 17, 355–379 (2011). 25. Bock, D.D. et al. Nature 471, 177–182 (2011). 26. Briggman, K.L., Helmstaedter, M. & Denk, W. Nature 471, 183–188 (2011). 27. Freed, M.A. & Sterling, P. J. Neurosci. 8, 2303– 2320 (1988). 28. Semon, R. W. & Simon, L. The Mneme. (G. Allen & Unwin Ltd., 1921). 29. Helmstaedter, M. Nat. Methods 10, 501–507 (2013). 30. Kaynig, V. & Vazquez-Reina, A. IEEE Trans. Med. Imaging 1, 1–7 (2012). 31. Lein, E.S. et al. Nature 445, 168–176 (2007). 32. Micheva, K.D., Busse, B., Weiler, N.C., O’Rourke, N. & Smith, S.J. Neuron 68, 639–653 (2010). focus on mapping the brain perspective Cellular-resolution connectomics: challenges of dense neural circuit reconstruction npg © 2013 Nature America, Inc. All rights reserved. Moritz Helmstaedter Neuronal networks are high-dimensional graphs that are packed into three-dimensional nervous tissue at extremely high density. Comprehensively mapping these networks is therefore a major challenge. Although recent developments in volume electron microscopy imaging have made data acquisition feasible for circuits comprising a few hundreds to a few thousands of neurons, data analysis is massively lagging behind. The aim of this perspective is to summarize and quantify the challenges for data analysis in cellular-resolution connectomics and describe current solutions involving online crowdsourcing and machine-learning approaches. Brains are unique not in the number of cells they comprise (about 85 billion neurons in the case of the human brain1,2) but in the extent of direct and specific communication between their cells via synaptic connections (each neuron has on the order of 1,000 synaptically coupled partner neurons). Mapping the resulting complex connectivity graph is the goal of connectomics. At the coarse level, inter-areal projections are tracked either noninvasively using variants of diffusion imaging in humans3 (see Review4 in this issue) or invasively using tracer injections combined with high-throughput imaging in mice (Mouse Brain Architecture Project5 (http://brainarchitecture.org/) and Allen Brain Connectivity Atlas (http://connectivity. brain-map.org/) among other initiatives; see Review6 in this issue). However, the mapping coverage achieved using these approaches is still not sufficient to resolve the complete set of neuronal networks contained in the tissue. One voxel of magnetic resonance imaging data with typical one-cubic-millimeter resolution, for example, contains millions of neuronal cell bodies and several kilometers of neuronal wires. At the other end of the resolution spectrum, the aim of connectomics is to image and analyze neuronal circuits densely in sufficiently large volumes but at single-cell and single-neurite resolution. Neuroscience is very data-poor in this respect, contrary to the assumptions made by contemporary simulation initiatives (The Human Brain Project7; http://www.human brainproject.eu/), and it is not known whether there are cases in which the measurement of cellular-resolution connectivity graphs can in fact be replaced by simplified low-order statistical assumptions about neuronal wiring. To date, only a few neuronal circuits have been analyzed comprehensively—with the connectivity map of the entire nervous system of Caenorhabditis elegans, comprising 302 neurons8,9, being the most notable and largest for decades. Neuronal-circuit reconstruction is so difficult because of the small size of neuronal processes and the extremely high packing density of the neuropil (Fig. 1a; here the term neuropil is used to describe nervous tissue containing densely packed axons and dendrites, even if interspersed with cell bodies, glia cells and blood vessels). For more than a century, the analysis of neuronal circuitry was therefore focused on using very sparse labeling techniques that stained only every 10,000th to 100,000th neuron in a given volume (Fig. 1a). Only recent advances in high-throughput electron microscopy have made possible the imaging of substantial volumes of neuronal tissue at high resolution, allowing the reconstruction of larger circuits at a much faster pace. The main challenge, however, is the analysis of imaging data, which is currently the limiting step by several orders of magnitude. Overcoming this gap in data analysis is therefore the main methodological focus of cellular-resolution connectomics. Resolution requirements The mapping of dense circuits requires the reconstruction of a large fraction of the neuronal wires in a given volume of nerve tissue (Fig. 1a). In most cases, Structure of Neocortical Circuits Group, Max Planck Institute of Neurobiology, Munich-Martinsried, Germany. Correspondence should be addressed to M.H. ([email protected]). Received 11 February; accepted 15 April; published online 30 may 2013; doi:10.1038/nmeth.2476 nature methods | VOL.10 NO.6 | june 2013 | 501 perspective focus on mapping the brain neuropil is largely isotropic: neuronal processes can locally turn in any direction, even in cases where there is a preferred direction such as the light axis in the retina or the radial developmental axis in the neocortex. This means that the lowest-resolution dimension of the imaging techniques used should account for the smallest neurite diameter in the chosen tissue volume. Examples of the smallest neurites in several parts of the nervous system (Fig. 1b) are dendritic spine necks (40–50 nanometer (nm) diameter, for example, in mouse hippocampus10 and neocortex), the thinnest parts of mouse cortical axons (~50 nm; K.M. Boergens and M.H.; unpublished data), amacrine cell dendrites in the mouse retina (~50 nm (ref. 11; K.L. Briggman, personal communication) and dendrites in the fly mushroom body (~30 nm (ref. 12)). The minimal required imaging resolution is then roughly half the smallest neurite diameter, if one requires each neurite diameter to be represented by at least two voxels in the three-dimensional (3D) image data set. It may be possible in some cases to relieve this resolution requirement slightly based on prior knowledge about the shape of neurites and the continuity of intracellular organelles. Thus, the minimal required imaging resolution is ~20–30 nm for most circuits but can be as small as 10–15 nm in certain model systems. a A B C D E F 10 –5 b Mushroom body dendrites 30 10 –4 –3 40 –2 10 10 Fraction of labeled neurons Dendritic spine neck 10 Amacrine cell dendrite 50 60 Smallest neurite diameter (nm) –1 1 Pyramidal cell axon 70 80 Volume requirements: minimal circuit dimension When mapping neuronal circuits (Fig. 2a), it is important to detect synaptic contacts between neurons, but it is in many cases even more important to be able to exclude synaptic connectivity between neurons to determine the structure of a wiring diagram (measuring the ‘zeros’ of the connectivity matrix). To exclude the existence of a synapse between two neurons, one has to image at least one of these two neurons in entirety. More precisely, in cases where neurons have an input part (for example, pyramidal cell dendrites, which are assumed to be only postsynaptic for chemical synapses) and an output part (for example, the axon of a local interneuron), at least one of the two neuronal arbors have to be fully contained in the imaged volume. This notion implies that for each circuit to be mapped, one can define a minimal circuit volume that fulfills the criterion of sufficient completeness for a sufficient number of the relevant neurons (Fig. 2b–e). Then, the imaging technique should be capable of imaging volumes such that the smallest imaged dimension is at least as large as the minimal required circuit dimension. ~1 mm ~300 µm ~500 µm 502 | VOL.10 NO.6 | june 2013 | nature methods ~300 µm ~80 µm a Connectome b Mouse retina c Mouse S1 cortex e Olfactory bulb Figure 2 | Minimal circuit dimensions. inner plexiform layer layer 4 glomerulus (a) A ‘connectome’, the connectivity matrix Postsynaptic ~40 µm ‘1’ between a set of presynaptic and postsynatptic neurons, with 15% positive entries (‘1’, blue dots) representing the existence of a synaptic connection between two neurons and 85% negative entries (‘0’, white) representing the absence of a synaptic connection. Detection of a synaptic connection is a local decision Mouse cortex layer 2/3 d ‘0’ (on the order of a few micrometers; top right). To exclude synaptic connectivity between two neurons, at least one of them has to be imaged in its entirety, yielding the notion of a minimal circuit volume (dashed box) >1–2 mm and the minimal circuit dimension (the smallest of the volumes’ dimensions). (b–e) Approximate minimal circuit dimensions for several example circuits, based on the requirement to measure the existence and the absence of synaptic connections and the spatial extent of the relevant neurons in the circuit. Not drawn to scale. Dashed boxes indicate the respective minimal circuit volumes. Presynaptic npg © 2013 Nature America, Inc. All rights reserved. Figure 1 | Density of neuronal circuits and minimal resolution requirements. (a) Fraction of labeled neurons using light microscopy (left) and electron microscopy (right). In light microcopy images single neurons can be detected only because of extremely sparse staining. Images show (left to right): two synaptically coupled neurons from layers 4 and 2/3 in rat somatosensory cortex (modified from ref. 48), sketch of cortical neurons (from ref. 49, image from Wikimedia), population of fluorescently labeled neurons in neocortex that were presynaptic to one postsynaptic pyramidal neuron (modified from ref. 50), reconstruction of mouse retina inner plexiform layer (modified from ref. 36) and depiction of C. elegans neurons (from wormbase.org and openworm.org, courtesy of C. Grove and S. Larson). (b) Minimal resolution requirements for cellular connectomics. The thinnest structures in various neuronal circuits are indicated, from left to right: mushroom body dendrite (modified from ref. 12) dendrite of layer 4 neuron in mouse neocortex (image by B. Cowgill), starburst amacrine cell dendrite, modified from ref. 11, GABAergic axon in mouse neocortex (image by K. Boergens and N. Marahori). Not drawn to scale. Red arrows indicate examples of the respective neurites. focus on mapping the brain a perspective e Manual ultrathin serial sectioning TEM(CA) Human cortex b Automated serial sectioning, tape collection c SEM d Minimal circuit dimension 1 cm M.ret. w.f. Z.f. larv. w.b. M.o.b. glom. D.m. m.b. 1 mm 100 µm Mouse cortex S1 layer 2/3 * ssTEM(CA) C. elegans w.b. SB EM M. retina s.f. 10 µm D.m. medulla FIB-SEM npg © 2013 Nature America, Inc. All rights reserved. Diamond knife, serial block-face SEM FIB, serial block-face SEM 0 20 ATUMssSEM 40 60 Minimal required resolution (nm) 80 Figure 3 | Volume electron microscopy techniques for cellular connectomics and their spatial resolution and scope. (a–d) Sketches of the four most widely used methods for dense-circuit reconstruction: conventional manual ultrathin sectioning of neuropil (a) followed by TEM or TEMCA imaging16 (a), ATUMSEM17 (b), SBEM18 (c) and FIB-SEM19 (d). In a,b, tissue is first sectioned and then (potentially later) transferred into the electron microscope for imaging. In c,d, the tissue block is abraded while imaging inside the electron microscope. (e) Approximate minimal resolution and smallest spatial dimension typically attainable with the imaging techniques in a–d, based on published results (gray shading); dashed lines indicate likely future extensions. Values also depend on the quality of staining and neurons of interest in a circuit. Approximate minimal resolution and minimal circuit dimension required to image indicated circuits. C. elegans w.b., C. elegans whole-brain reconstruction8,9,35; solid line indicates longest series from one worm and dashed line, the combined series length from three worms9. D.m. m.b., Drosophila melanogaster mushroom body; minimal required resolution based on estimate of smallest dendrites (30 nm diameter; Fig. 1b); D.m. medulla, D. melanogaster medulla, 1 cartridge (diameter of ~6 µm) with smallest processes less than 15 nm diameter. Human cortex, minimal circuit volume containing entire L5 pyramidal neuron dendrites and their local axons. Mouse cortex S1 layer 2/3, minimal circuit volume (Fig. 2d). *, mouse cortex S1 layer 4 minimal circuit volume (Fig. 2c). M.o.b.glom., mouse olfactory bulb, 1 glomerulus, only intraglomerular circuitry (Fig. 2e). M. retina s.f., mouse retina, small field (Fig. 2b). M.ret. w.f., mouse retina, wide field (including the largest amacrine and ganglion cells). Z.f. larv. w.b.: zebrafish larva whole brain. For example, the minimal circuit volume for bipolar-toganglion cell connectivity in the mouse retina is dictated by the size of the axon of bipolar cells, which is contained in a volume of ~20 µm × 20 µm × 40 µm for most bipolar cells (20 µm in the plane of the retina, and 40 µm along the light axis; Fig. 2b). If one wants to map the circuits of roughly a dozen bipolar cells per bipolar cell type, the resulting minimal circuit volume is ~80 µm × 80 µm × 40 µm (Fig. 2b); thus the minimal circuit dimension is 40 µm. Other examples are the mouse olfactory bulb, where the entire circuit within one glomerulus, excluding mitral cells, is contained in a volume of ~300 µm × 300 µm × 300 µm (Fig. 2e). If one aims to include mitral cells into the circuit analysis, the minimal volume increases to ~1 mm × 500 µm × 500 µm (Fig. 2e; A. Schaefer; personal communication). In mouse neocortex, the minimal circuit volume is thought to be smallest in layer 4 of primary somatosensory cortex (~300 µm × 300 µm × 300 µm, Fig. 2c), it is ~1 mm × 1 mm × 500 µm in layer 2/3, ~3 mm × 3 mm × 1 mm in layer 5 and at least as large in other cortical areas where neurons have widespread axonal projections (Fig. 2d). Together, the required minimal imaging resolution, as dictated by the smallest neurite diameters, and the minimal circuit dimension, as dictated by the types and spatial extent of the relevant neurons, differ widely between different circuit model systems. To map these circuits, it is therefore crucial to choose the appropriate volume imaging method (Fig. 3). Volume electron microscopy techniques All current imaging techniques used for cellular-resolution connectomics rely on sequential two-dimensional (2D) imaging of tissue combined with slicing, abrading or evaporating the tissue. Then images are combined into a 3D image volume. Only in this sense are today’s imaging methods volume imaging methods. The key techniques are serial-section transmission electron microscopy (ssTEM9,13–15) in some cases combined with fast camera arrays (TEMCA16, Fig. 3a), automated serial-section tape-collection scanning electron microscopy (ATUM-SEM17; Fig. 3b), serial block-face SEM (SBEM11,18, Fig. 3c) and focused ion beam SEM (FIB-SEM19, Fig. 3d); these microscopy techniques are reviewed in ref. 20. Briefly, these methods differ in the sequence and quality of cutting and imaging, and provide different minimal imaging resolutions and dimensions. Essentially, the resolution is currently lowest, and the acquirable dimension is smallest, along the cutting axis. In ssTEM (Fig. 3a), tissue is first cut into ultrathin sections, which are then imaged using TEM or speed-enhanced TEMCA. As sections are cut manually and are physically maintained for the imaging step, these techniques have a minimal resolution of ~40 nm. They provide large imaging areas in the plane of imaging but yield typically small dimensions along the cutting axis because manually cutting and maintaining more than a few thousand sections is hardly possible. ATUM-SEM (Fig. 3b) overcomes the cutting-thickness limitation of ssTEM by automated slicing and automated tape collection of the sections, yielding a minimal resolution of up to 30 nm and providing long series of collected sections. These are imaged using SEM as the collection tape is electron-dense. The two block-face techniques (Fig. 3c,d) interlace and invert the sequence of imaging and cutting: the entire tissue block is transferred into the electron microscope chamber, and nature methods | VOL.10 NO.6 | june 2013 | 503 perspective focus on mapping the brain a b d Contouring 1 Difficult configuration Synapse 2 .... 1 c 6 Skeleton tracing e Missed branchpoint 3 Axon npg © 2013 Nature America, Inc. All rights reserved. 4 5 Dendrite 6 Figure 4 | Manual and automated reconstruction challenges in electron microscopy–based connectomics. (a) Volume reconstruction of a synaptic contact between an excitatory axon and dendritic spine in mouse somatosensory cortex imaged using SBEM (top right; K.M. Boergens and M.H., unpublished data). Images show electron microcopy slices at various depths; distance between slices 1 and 6 is 500 nm. (b,c) Manual data annotation is performed mostly either by contouring of neurite cross-sections (b) or center-line reconstruction (skeleton tracing, c). Contouring is ~50 times faster but provides no direct volume reconstruction. Modified from ref. 51. (d) Sketch of three possible neurite configurations for a configuration that automated algorithms typically cannot resolve but human annotators can (panels show three possible configurations). Dashed lines, imaging plane in which analysis is most difficult. Note that one neurite runs parallel to imaging plane (blue and green). Such a configuration is only resolvable if minimal resolution is greater than the smallest neurite diameter. Analysis may be facilitated when including priors about neurite diameters (pink neurite). (e) Branchpoint that can be missed by human annotators (unforced attention-related error25) corresponding to the dashed line in scheme on the left (from ref. 25; blue and white, skeletons traced independently by two expert annotators). images are taken from the surface of the block using SEM. Then, the top of the tissue block is abraded either using a diamondknife microtome installed into the electron microscope chamber (SBEM, Fig. 3c) or using a focused ion beam attached to the electron microscope (FIB-SEM, Fig. 3d). Diamond knife–based cutting currently limits the abrasion thickness to 25 nm but allows for long imaging sequences (smallest dimension is currently 300 µm (ref. 11), K.M. Boergens and M.H.; unpublished data), whereas FIB-SEM is unique in providing a minimal resolution of 4 nm, but so far limited in its smallest dimension (~40 µm (ref. 19)). The relationship between the available electron microscopy imaging techniques and the resolution and size requirements of several example circuits is summarized in Figure 3e. In a nutshell, ssTEM is most amenable for circuits that can tolerate a minimal required resolution of more than 40 nm and a minimal circuit dimension of less than a few hundred micrometers, whereas FIBSEM is uniquely suited for circuits that require very high imaging resolution, such as most circuits in the fly nervous system, but can tolerate smaller volumes. ATUM-SEM and SBEM are currently best suited for dense-circuit reconstruction in systems with ~25–30 nm minimal resolution and several hundred micrometers minimal circuit dimension. The aims of ongoing method developments are to increase the spatial scope along the cutting dimension, increase imaging speed in the plane of imaging and potentially decrease the cutting or abrasion thickness (Fig. 3e). Circuit reconstruction All high-resolution imaging techniques generate large volumes of images, which totaled several hundred gigabytes per data set for published projects11,16,21, are currently around a dozen terabytes per data set and will likely be several petabytes per data set in 504 | VOL.10 NO.6 | june 2013 | nature methods the foreseeable future. To reconstruct neuronal circuits from the imaging data, synapses have to be identified, and the pre- and postsynaptic neurites contributing to each synapse have to be followed back to their respective cell bodies to attribute the synapse to the correct neuron. As soon as the imaging resolution and staining quality are sufficient for detecting postsynaptic densities and synaptic vesicles (which are ~40 nm in diameter and thus at the same scale as most minimal neurite diameters, see above), the detection of synapses becomes feasible. Synapse detection is more difficult in systems where one presynaptic site can have multiple postsynaptic partners (such as in the mammalian retina and the fly optical system), presynaptic vesicle pools are small and/or postsynaptic specializations are difficult to identify (for example, symmetrical synapses of inhibitory axons in the neocortex). In all cases, volume imaging substantially improves the detection of synapses compared to the identification of synapses in two-dimensional images. For example, the 3D imaging of a synapse in mouse neocortex (Fig. 4a, at 25 nm slice thickness using SBEM, K.M. Boergens and M.H., unpublished data) substantially improves the reliability of synapse detection: judging the synapse from just one of the 2D images may be questionable, but the sequence of several 2D images of the same synaptic location relieves uncertainty of detection. Synapse detection is additionally improved by high-resolution 3D imaging as obtained using FIB-SEM22 and can in these cases be automated22,23. Automated synapse-detection methods for SBEM and ssTEM data are likely to become available soon. The main challenge for cellular connectomics, however, is not the detection of synapses but the reconstruction of neuronal wires (Fig. 4b–e). This immense difficulty has two main origins: first, neurites vary greatly in diameter and in local entanglement, generating a substantial frequency of locations at which the path of focus on mapping the brain npg © 2013 Nature America, Inc. All rights reserved. a Data acquisition Reconstruction Gaming 109 106 Completed Analysis gap Completed *** SBEM ssTEM * * * * ** 3 10 ssTEMCA SBEM Time (hours) Figure 5 | Imaging and analysis times illustrating the analysis gap in cellular connectomics. (a) Data acquisition times and respective reconstruction durations for completed projects and estimates for ongoing or planned projects. Blue and red shading indicates experiment and reconstruction, respectively. Data-acquisition projections assume ~5 megahertz imaging speed and SBEM experiment (*); human cortex minimum circuit volume–acquisition projection assumes an additional ~200-fold increase in imaging speed as potentially attainable with multibeam electron microscopes (**). A recent C. elegans data analysis had a tenfold increase in the speed of annotation26 (dashed line, ***). Gaming estimates are provided for successful science games in total annotation hours and an online casual game (Angry Birds) in annotation hours per year. FoldIt estimate is based on the number of players; Galaxy Zoo estimate is based on the reported number of classifications and an estimate of on average 30 seconds per classification. (b,c) Screen shots of ongoing online gaming initiatives in cellular connectomics: Eyewire (b; image courtesy of H.S. Seung) and Brainflight (c). perspective ) y 7) 9) 4) 6) 4 s /3 in it 3) 9) 4) 6) us 4 /3 in nse ar 43 om 6, f. . 1 3 x L ulu L2 ra rcu e ,3 f. 1 3 ul x L L2 ra ye er rte ex le b de (ref. ron s. (re (ref (ref. rte mer ex le b ci um ,24 (re (ref. (ref. f x er t e p s om co ort ho (re tina ex se 1 co glo cort ho um vol 6,7 rse se se g l t r g 1 c w . in a rds ns re ort en S lb 1 a w nim co fs spa par en ld o lb S S1 va fo Zo Bi ga e c d e u S rv i r (re ina x s a d y bu an y le ous V1 ina ous ry b se la ex m la in axy e n m e t e i r e u h t t h s t gr u t t al o s n Re or re t n i o H C. M use e re Macto Mo afis cor r c e A F c e p G d f o lfa br n it s ol M ous V1 ous O ld an Ze uma se M M Fo eg l H ou e M C. neurites can only be revealed after intense inspection (Fig. 4d). Second, and most notably, errors of neurite continuity have a highly correlated effect such that any neurite break along the path from a synapse to its soma creates an error in the assignment of the synapse to a neuron. As these entangled paths can be as long as several millimeters in many circuits (note that this is the neuronal path length between synapse and soma, not the Euclidean distance), neurite reconstruction has to be extremely reliable to provide tolerable error rates in the connectivity matrix. In spite of these difficulties, human annotators can reconstruct electron microscopy data by either contouring the neurites in sequential image planes (Fig. 4b (refs. 21,24)) or by reconstructing only the neurite center line in three dimensions (Fig. 4c (refs. 25,26). Most notably, human annotators can resolve difficult locations with high accuracy (Fig. 4d). However, such annotation is very slow, and annotators make many unforced attentionrelated mistakes25 (Fig. 4e). In an attempt to speed up reconstruction, a main focus in connectomic data analysis was to devise automated reconstruction algorithms15,22,27–33. Unfortunately, however, these algorithms so far do not provide sufficient reconstruction accuracy to replace human annotators. Currently, automated reconstructions generate erroneous neurite breaks or erroneous neurite mergers every few tens of micrometers of neuronal path length, which results in the loss or misassignment of most of the synapses of a given neuron. Automated algorithms can, however, substantially support data annotation by human annotators, as discussed below. Why do automated algorithms fail at resolving what human annotators can resolve? Figure 4d is an illustration of a local neurite configuration that could be interpreted in multiple ways by only subtle local changes of the membranes separating the respective neurites. Presently available automated algorithms are notoriously bad at resolving such locations, especially when the concerned neurites are very thin. It can be suspected that such neurite configurations can only be resolved when explicitly comparing the different possible segmentation solutions (Fig. 4d), which is what the human visual system may compute when inspecting such a location in the data. Recent developments in algorithms aim to account for such segmentation-based data analysis31,34. Analysis of electron microscopy data is thus meeting a substantial challenge: humans can solve difficult locations but are slow b c and make attention-based errors, whereas machines are efficient at solving easy locations but fail when neurites become small and their packing is dense. The key approach to resolving the reconstruction problem in electron microscopy–based connectomics has therefore been to combine massive human data-annotation efforts with automated image analysis. In essence, human annotators provide the long-range connectivity information and solve the most difficult annotation problems, and automated algorithms provide the local volume reconstruction at less difficult locations. This analysis approach has been implemented in two variants: either automated algorithms are first used to presegment the imaging data, followed by manual inspection and correction (iterated proofreading15,21,35, usually involving well-trained full-time annotators), or manual annotation is parallelized, followed by automated consensus computation, and by combination with local volume segmentation (forward-only annotation, usually using lightly trained part-time annotators such as undergraduate students25). The analysis gap in cellular connectomics What are the resulting experiment and analysis times for cellularresolution connectomics today? In Figure 5a, durations for the accomplished circuit imaging and reconstruction experiments are reported as well as estimates for ongoing and envisioned projects. Data acquisition took around several hundred to a thousand hours for the reconstruction of the C. elegans connectome9 (ssTEM), the direction-selectivity circuit analysis in the mouse retina11 (SBEM), a local circuit analysis experiment in V1 mouse cortex16 (ssTEMCA) and the dense local reconstruction of the inner plexiform layer in mouse retina36 (SBEM). This time was spread over several months for the ssTEM methods and compressed into 6–8 weeks in the SBEM experiments. Imaging and acquisition speed has since substantially increased, bringing much larger volumes within the reach of a several-month continuous experiment: for example, one glomerulus in mouse nature methods | VOL.10 NO.6 | june 2013 | 505 npg © 2013 Nature America, Inc. All rights reserved. perspective olfactory bulb (~300 µm × 300 µm × 300 µm), the minimal circuit volume in layer 4 of mouse S1 cortex (~300 µm × 300 µm × 300 µm) or the entire brain of the larva-stage zebra fish (~400 µm × 500 µm × 500 µm). Highly parallel SEM may promise an additional speed-up of more than two orders of magnitude, which would even enable the imaging of a relevant volume of human cortex (~2 cm × 2 cm × 1 cm). Improvements in imaging and acquisition rates appear to have moved many circuits of interest into the realm of experimental feasibility, which is ~1,000– 2,000 hours of imaging time (Fig. 5a; continuous experiments of up to a year (~8,000 hours) may become feasible with improvements in automation). However, if one considers the required reconstruction time, cellular-resolution connectomics imposes daunting challenges. The two recently published mouse connectomics studies made use of very sparse circuit reconstructions: only around 30 neurons were reconstructed for the direction-selectivity circuit analysis11 and the V1 cortex study16, already amounting to ~1,000 hours of analysis time. For dense-circuit reconstruction, the C. elegans reconstruction took around 12 years of part-time work by one scientist9,26, also amounting to around a few thousand hours of work. Including the additional studies8,35, the C. elegans connectome may have consumed around 10,000–20,000 hours spread across three decades. The most recent mammalian dense-circuit reconstruction in mouse retina36 took ~30,000 hours, by hundreds of parttime undergraduate annotators. Thus, all cellular-resolution connectomics studies to date have involved thousands to tens-of-thousands of hours of reconstruction. Although imaging speed has substantially increased, reconstruction speed is massively lagging behind. For any of the proposed dense-circuit reconstructions (mouse neocortex, olfactory bulb, fish brain and human neocortex; Fig. 5a), analysistime estimates are at least one if not several orders of magnitude larger than what has been accomplished to date (Fig. 5a): requirements for these envisioned projects are around several hundred thousand hours of manual labor per project. These enormous numbers constitute what can be called the analysis gap in cellular connectomics: although imaging larger circuits is becoming more and more feasible, reconstructing them is not. Addressing the analysis problem How is the field approaching this massive methodological challenge? There are two lines of solutions. First, improvements in automated algorithms promise to increase the efficiency of reconstruction by human annotators. Second, massive online crowdsourcing initiatives are being built to ‘recruit’ the required work hours over the internet. The current accuracy of automated classifiers is at least three orders of magnitude less than what is needed for reconstruction of entire neurons. It is therefore unlikely to expect that automated algorithms will soon take over image analysis for connectomics completely. Improvements in algorithms have saturated at the level of voxel classification (‘voxel classifiers’27,28,32,37–39), and the current focus is on segmentation-based classifiers (‘supervoxel classifiers’31,34,40). Rather than waiting for such algorithms to reach the required accuracy, automated methods have been made useful in reducing the analysis challenge when they were well integrated with human annotation15,25,41–44. For example, 506 | VOL.10 NO.6 | june 2013 | nature methods focus on mapping the brain the estimates of circuit reconstruction reported in Figure 5 are based on manual-algorithmic integration that already yielded a 10–50-fold increase in reconstruction efficiency. Thus, on one hand, the key prospect for the improvement of automated analysis remains in lowering error rates (that is, generating longer correctly reconstructed neurite segments), but on the other hand, the focus is on better integration with human annotation. In an interesting twist, all automated methods rely heavily on training data. It is typically very expensive, however, to generate training data, and this in many cases limits algorithm improvements. Therefore, the attempts to recruit annotators to perform large amounts of human annotation, as described next, also may provide a critical boost to improve automated reconstruction methods. Online crowd-sourcing has been successful in a few scientific fields: most notably the protein-folding game FoldIt45 and the galaxyclassification initiative Galaxy Zoo46. Although these approaches proved very fruitful in their respective settings, the time requirements in cellular connectomics are still large by comparison. Galaxy Zoo has (over several years) probably recruited the work time required for the reconstruction of one dense cortical circuit (Fig. 5a). The comparison to casual gaming, however, seems promising: even a single successful game is reported to be played millions of hours every day, amounting to as many hours a year as would be needed to reconstruct a relevant piece of human cortex (~1 billion hours; Fig. 5a). This setting of successful citizen-science games and the order of magnitude of hours spent online playing games has fueled hopes of recruiting the required analysis via the internet. As a first prerequisite, data annotation had to move online, overcoming the challenges of streaming the required data to clients’ browsers. The Open Connectome project using the collaborative annotation toolkit for massive amounts of image data (CATMAID) software47 focused on ssTEM and ssSEM data analysis, the Eyewire project for SBEM data (http://www.eyewire.org/; Fig. 5b) and the Brainflight project (http://www.brainflight.org/; Fig. 5c) are notable examples of online annotation, with Eyewire and Brainflight using game-like features to enhance engagement. It is an exciting endeavor today to explore whether citizen science can help to overcome the massive analysis gap we are facing in cellular-resolution connectomics. Only if hundreds of thousands of hours can be recruited will dense-circuit reconstruction be feasible. It is also possible that improved automation will provide the crucial gain in the efficiency of analysis, but most likely it is going to be a combination of the two that will make dense-circuit reconstruction more and more realistic. Notably, it can be expected that once such large data sets are being reconstructed, the large amounts of training data needed to create more powerful analysis algorithms will be available. Then, finally, can connectomics analysis become a routine method with applications in many neuroscience laboratories, and cellular-resolution connectomics may further contribute to unraveling the computations neuronal circuits perform. Acknowledgments I am grateful to the members of my laboratory for many fruitful discussions, specifically to K.M. Boergens, Y. Buckley, F. Isensee, N. Marahori, A. Mohn and H. Wissler for help with generating figures, and E. Dow for discussions concerning game development. I thank M. Berning, K.M. Boergens, E. Dow and A. Schaefer for helpful comments on the manuscript. COMPETING FINANCIAL INTERESTS The author declares no competing financial interests. focus on mapping the brain npg © 2013 Nature America, Inc. All rights reserved. Reprints and permissions information is available online at http://www.nature. com/reprints/index.html. 1. Herculano-Houzel, S. The human brain in numbers: a linearly scaled-up primate brain. Front. Human Neurosci. 3, 31 (2009). 2. Azevedo, F.A. et al. Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. J. Comp. Neurol. 513, 532–541 (2009). 3. Wedeen, V.J. et al. The geometric structure of the brain fiber pathways. Science 335, 1628–1634 (2012). 4. Craddock, R.C. Imaging human connectomes at the macroscale. Nat. Methods 10, 524–539 (2013). 5. Bohland, J.W. et al. A proposal for a coordinated effort for the determination of brainwide neuroanatomical connectivity in model organisms at a mesoscopic scale. PLOS Comput. Biol. 5, e1000334 (2009). 6. Osten, P. & Margrie, T.W. Mapping brain circuitry with a light microscope. Nat. Methods 10, 515–523 (2013). 7. Markram, H. The blue brain project. Nat. Rev. Neurosci. 7, 153–160 (2006). 8. Varshney, L.R., Chen, B.L., Paniagua, E., Hall, D.H. & Chklovskii, D.B. Structural properties of the Caenorhabditis elegans neuronal network. PLOS Comput. Biol. 7, e1001066 (2011). 9. White, J.G., Southgate, E., Thomson, J.N. & Brenner, S. The structure of the nervous system of the nematode Caenorhabditis elegans. Philos. Trans. R. Soc. Lond. B Biol. Sci. 314, 1–340 (1986). 10. Chicurel, M.E. & Harris, K.M. Three-dimensional analysis of the structure and composition of CA3 branched dendritic spines and their synaptic relationships with mossy fiber boutons in the rat hippocampus. J. Comp. Neurol. 325, 169–182 (1992). 11. Briggman, K.L., Helmstaedter, M. & Denk, W. Wiring specificity in the direction-selectivity circuit of the retina. Nature 471, 183–188 (2011). 12. Butcher, N.J., Friedrich, A.B., Lu, Z., Tanimoto, H. & Meinertzhagen, I.A. Different classes of input and output neurons reveal new features in microglomeruli of the adult Drosophila mushroom body calyx. J. Comp. Neurol. 520, 2185–2201 (2012). 13. Fahrenbach, W.H. Anatomical circuitry of lateral inhibition in the eye of the horseshoe crab, Limulus polyphemus. Proc. R. Soc. Lond. B Biol. Sci. 225, 219–249 (1985). 14. Harris, K.M. et al. Uniform serial sectioning for transmission electron microscopy. J. Neurosci. 26, 12101–12103 (2006). 15. Mishchenko, Y. et al. Ultrastructural analysis of hippocampal neuropil from the connectomics perspective. Neuron 67, 1009–1020 (2010). 16. Bock, D.D. et al. Network anatomy and in vivo physiology of visual cortical neurons. Nature 471, 177–182 (2011). 17. Hayworth, K.J., Kasthuri, N., Schalek, R. & Lichtman, J.W. Automating the collection of ultrathin serial sections for large volume TEM reconstructions. Microsc. Microanal. 12 (suppl. 2), 86–87 (2006). 18. Denk, W. & Horstmann, H. Serial block-face scanning electron microscopy to reconstruct three-dimensional tissue nanostructure. PLoS Biol. 2, e329 (2004). 19. Knott, G., Marchman, H., Wall, D. & Lich, B. Serial section scanning electron microscopy of adult brain tissue using focused ion beam milling. J. Neurosci. 28, 2959–2964 (2008). 20. Briggman, K.L. & Bock, D.D. Volume electron microscopy for neuronal circuit reconstruction. Curr. Opin. Neurobiol. 22, 154–161 (2012). 21. Cardona, A. et al. An integrated micro- and macroarchitectural analysis of the Drosophila brain by computer-assisted serial section electron microscopy. PLoS Biol. 8, e1000502 (2010). 22. Kreshuk, A. et al. Automated detection and segmentation of synaptic contacts in nearly isotropic serial electron microscopy images. PLoS ONE 6, e24899 (2011). 23. Becker, C., Ali, K., Knott, G. & Fua, P. Learning context cues for synapse segmentation in EM volumes. Med. Image Comput. Comput. Assist. Interv. 15, 585–592 (2012). 24. Fiala, J.C. Reconstruct: a free editor for serial section microscopy. J. Microsc. 218, 52–61 (2005). 25. Helmstaedter, M., Briggman, K.L. & Denk, W. High-accuracy neurite reconstruction for high-throughput neuroanatomy. Nat. Neurosci. 14, 1081–1088 (2011). Algorithm-based apparatus for combining the reconstructions of many slightly trained annotators into a consensus reconstruction, a prerequisite for large-scale crowd-sourcing. 26. Xu, M. et al. Computer assisted assembly of connectomes from electron micrographs: application to Caenorhabditis elegans. PLoS ONE 8, e54050 (2013). perspective 27. Jain, V. et al. Supervised learning of image restoration with convolutional networks. IEEE 11th International Conference on Computer Vision 1–8 (2007). 28. Jain, V., Seung, H.S. & Turaga, S.C. Machines that learn to segment images: a crucial technology for connectomics. Curr. Opin. Neurobiol. 20, 653–666 (2010). 29. Turaga, S.C. et al. Convolutional networks can learn to generate affinity graphs for image segmentation. Neural Comput. 22, 511–538 (2010). 30. Chklovskii, D.B., Vitaladevuni, S. & Scheffer, L.K. Semi-automated reconstruction of neural circuits using electron microscopy. Curr. Opin. Neurobiol. 20, 667–675 (2010). 31. Andres, B. et al. 3D segmentation of SBFSEM images of neuropil by a graphical model over supervoxel boundaries. Med. Image Anal. 16, 796–805 (2012). 32. Andres, B., Köthe, U., Helmstaedter, M., Denk, W. & Hamprecht, F. Segmentation of SBFSEM volume data of neural tissue by hierarchical classification. in Pattern Recognition, Proc. DAGM 2008, Lecture Notes in Computer Science (ed., G. Rigoll) 5096, 142–152 (Springer, 2008). 33. Funke, J., Andres, B., Hamprecht, F.A., Cardona, A. & Cook, M. Efficient automatic 3D-reconstruction of branching neurons from EM data. Proc. IEEE Conference on Computer Vision and Pattern Recognition 1004–1011 (2012). 34. Jain, V. et al. Learning to agglomerate superpixel hierarchies. in Advances in Neural Information Processing Systems (eds., J. Shawe-Taylor, R.S. Zemel, P. Bartlett, F.C.N. Pereira and K.Q. Weinberger) 24, 648–656 (2011). 35. Jarrell, T.A. et al. The connectome of a decision-making neural network. Science 337, 437–444 (2012). 36. Helmstaedter, M., Briggman, K.L., Turaga, S., Jain, V., Seung, H.S. & Denk, W. Connectomic reconstruction of the inner plexiform layer in the mouse retina. Nature (in the press). 37. Turaga, S.C. et al. Convolutional networks can learn to generate affinity graphs for image segmentation. Neural Comput. 22, 511–538 (2010). 38. Mishchenko, Y. Automation of 3D reconstruction of neural tissue from large volume of conventional serial section transmission electron micrographs. J. Neurosci. Methods 176, 276–289 (2009). 39. Seyedhosseini, M. et al. Detection of neuron membranes in electron microscopy images using multi-scale context and radon-like features. Med. Image Comput. Comput. Assist. Interv. 14, 670–677 (2011). 40. Vazquez-Reina, A. et al. Segmentation fusion for connectomics. IEEE International Conference on Computer Vision 177–184 (2011). 41. Sommer, C., Straehle, C., Koethe, U. & Hamprecht, F.A. ilastik: interactive learning and segmentation toolkit. Proc. IEEE International Symposium on Biomedical Imaging 8, 230–233 (2011). 42. Straehle, C.N., Kothe, U., Knott, G. & Hamprecht, F.A. Carving: scalable interactive segmentation of neural volume electron microscopy images. Med. Image Comput. Comput. Assist. Interv. 14, 653–660 (2011). 43. Jeong, W. et al. Ssecrett and NeuroTrace: interactive visualization and analysis tools for large-scale neuroscience data sets. IEEE Comput. Graph. Appl. 30, 58–70 (2010). 44. Roberts, M. et al. Neural process reconstruction from sparse user scribbles. Med. Image Comput. Comput. Assist. Interv. 14, 621–628 (2011). 45. Cooper, S. et al. Predicting protein structures with a multiplayer online game. Nature 466, 756–760 (2010). Example of a computer game used for scientific discovery. 46. Land, K. et al. Galaxy Zoo: the large-scale spin statistics of spiral galaxies in the Sloan Digital Sky Survey. Mon. Not. R. Astron. Soc. 388, 1686–1692 (2008). 47. Saalfeld, S., Cardona, A., Hartenstein, V. & Tomancak, P. CATMAID: collaborative annotation toolkit for massive amounts of image data. Bioinformatics 25, 1984–1986 (2009). Online annotation platform for electron microscopy image data presented in successive 2D image planes. 48. Feldmeyer, D., Lubke, J., Silver, R.A. & Sakmann, B. Synaptic connections between layer 4 spiny neurone-layer 2/3 pyramidal cell pairs in juvenile rat barrel cortex: physiology and anatomy of interlaminar signalling within a cortical column. J. Physiol. (Lond.) 538, 803–822 (2002). 49. Ramón y Cajal, S. Textura del sistema nervioso del hombre y de los vertebrados (Imprenta N. Moya, 1904). 50. Rancz, E.A. et al. Transfection via whole-cell recording in vivo: bridging single-cell physiology, genetics and connectomics. Nat. Neurosci. 14, 527–532 (2011). 51. Helmstaedter, M. & Mitra, P.P. Computational methods and challenges for large-scale circuit mapping. Curr. Opin. Neurobiol. 22, 162–169 (2012). nature methods | VOL.10 NO.6 | june 2013 | 507 perspective focus on mapping the brain CLARITY for mapping the nervous system npg © 2013 Nature America, Inc. All rights reserved. Kwanghun Chung1,2 & Karl Deisseroth1–4 With potential relevance for brain-mapping work, hydrogel-based structures can now be built from within biological tissue to allow subsequent removal of lipids without mechanical disassembly of the tissue. This process creates a tissue-hydrogel hybrid that is physically stable, that preserves fine structure, proteins and nucleic acids, and that is permeable to both visible-spectrum photons and exogenous macromolecules. Here we highlight relevant challenges and opportunities of this approach, especially with regard to integration with complementary methodologies for brainmapping studies. Mammalian brains are staggeringly complex in terms of both scale and diversity; many billions of neurons are present, among them likely at least hundreds of genetically distinct cell types, with each type of cell represented by many distinct projection patterns. CLARITY1 is a newly developed technology that can be used to transform intact biological tissue into a hybrid form in which tissue components are removed and replaced with exogenous elements for increased accessibility and functionality. CLARITY has the potential to facilitate rapid extraction of anatomicalprojection information important for many fields of neuroscience research2,3; such information can be collected along with molecular-phenotype information at the resolution of single cells. Alone or in combination with other methods4,5, such an approach could contribute to the study of function and dysfunction in this complex system. In general, obtaining system-wide detailed information from neural tissue is a formidable challenge (to say nothing of subsequent data curation and analysis). In the mammalian central nervous system, seamlessly intertwined neural processes leave little extracellular space, creating barriers to macromolecule diffusion for in situ hybridization, antibody staining or other forms of molecular phenotyping deeper than the first few cellular layers of intact tissue6. The high density of irregularly arranged lipid interfaces that characterizes this tissue likewise creates an effective scattering barrier to photon penetration for optical interrogation of mammalian brains7, unlike the Caenorhabditis elegans (worm) or larval Danio rerio (zebrafish) nervous systems, which are more accessible owing in part to smaller size and less myelination. Singlephoton microscopy can provide optical transmission of information from only about 50 micrometers below the mammalian brain surface, and even welloptimized two-photon microscopy cannot be used to image deeper than about 800 micrometers, far short of enabling visualization of full projection patterns and global arrangement of cell populations in the intact brain8. Over the past few decades, a great deal of technological innovation has been stimulated by these challenges8–21. First, newer automated methods for mechanical sectioning of tissue have overcome some of the drawbacks of traditional sectioning methods that were laborious, expensive and damaged the tissue. Serial block-face mechanical 9–13 or opticalablative14 methods, in combination with imaging readouts such as two-photon tomography14,15, electron microscopy16 or array tomography17, have been used to map macroscopic to nanoscopic brain structures (see Review18 in this issue). In some cases, molecular labeling is built into these processes, and ongoing work includes approaches for addressing additional challenges such as generation of contrast in tissue before sectioning19,20. Tools for automated analysis, efficient reconstruction and error-free alignment also continue to be developed21 as well as for registration with activity information, and indeed detailed wiring information linked to activity has been obtained from well-defined volumes22,23. Second, optical clearing methods have been developed that involve immersion of the specimen in medium that matches the refractive index of the tissue, thereby reducing light scattering and extending 1Department of Bioengineering, Stanford University, Stanford, California, USA. 2CNC Program, Stanford University, Stanford, California, USA. 3Department of Psychiatry and Behavioral Sciences, Stanford University, Stanford, California, USA. 4Howard Hughes Medical Institute, Stanford University, Stanford, California, USA. Correspondence should be addressed to K.C. ([email protected]) and K.D. ([email protected]). Received 18 March; accepted 22 April; published online 30 may 2013; doi:10.1038/nmeth.2481 508 | VOL.10 NO.6 | JUNE 2013 | nature methods focus on mapping the brain perspective npg © 2013 Nature America, Inc. All rights reserved. Figure 1 | Imaging of nervous system projections in the intact mouse brain with CLARITY. Adult Thy1-EYFP H line mice (4 months old) were perfused transcardially followed by hybridization of biomolecule-bound monomers into a hydrogel mesh and lipid removal as described1. The clarified mouse brain was imaged from the dorsal region to the ventral region (3.4 mm to the midpoint) using a 10× water-immersion objective. Image adapted from ref. 1. the range of optical imaging24–26. For example, benzyl alcohol−benzyl benzoate is an organic solvent that effectively renders biological specimens transparent but reduces the stability of fluorescent protein signals24,25; in Scale, another clearing method, an aqueous solution is used to preserve fluorescence signals, but the rate and the extent of clearing remain limiting26. Notably all current tissue1 mm clearing methods leave the densely packed lipid bilayers intact and therefore still face challenges with regard to penetration by visible-spectrum light and molecules, making these methods largely incompatible with whole-tissue molecular phenotyping. The CLARITY approach1 helps address ongoing challenges by enabling molecular and optical interrogation of large assembled biological systems, such as the entire adult mouse brain. Light-microscopy (Fig. 1) and biochemical-phenotyping (Fig. 2) techniques can be used to rapidly access the entire intact clarified mouse brain with fine structural resolution and molecular detail (to the level of spines, synapses, proteins, single-amino-acid neurotransmitters and nucleic acids) while in the same preparation maintaining global structural information including brainwide macroscopic connectivity. To clarify tissue (Fig. 3), lipid bilayers are replaced with a more rigid and porous hydrogel-based a 3D rendering b DR infrastructure (in a process conceptually akin to petrification or fossilization, except that not only structure but also native biomolecules such as proteins and nucleic acids are preserved). This outcome is achieved by first infusing small organic hydrogelmonomer molecules into the intact brain along with crosslinkers and thermally triggered polymerization initiators; subsequent temperature elevation triggers formation, from within the brain, of a hydrogel meshwork covalently linked to native proteins, small molecules and nucleic acids but not to lipids, which lack the necessary reactive groups (Fig. 3a). Subsequent wholebrain electrophoresis in the presence of ionic detergents actively removes the lipids (Fig. 3b), resulting in a transparent brainhydrogel hybrid that both preserves, and makes accessible, structural and molecular information for visualization and analysis. DR RR z = 2,460 µm SNR SNR c VTA VTA PO z = 1,005 µm d CPu P PO CPu V D A z = 420 µm Figure 2 | Intact mouse brain molecular phenotyping and imaging with CLARITY. (a) Three-dimensional (3D) visualization of immunohistology data, showing tyrosine hydroxylase (TH)-positive neurons and fibers in the mouse brain. The intact clarified brain was stained for 6 weeks as described1, with primary antibody for 2 weeks, followed by a 1-week wash, then stained with secondary antibody for 2 weeks followed by a 1-week wash and imaged 2,500 µm from ventral side using the 10× water-immersion objective. D, V, A and P indicate dorsal, ventral, anterior and posterior, respectively. (b–d) Optical sections at different depths, corresponding respectively to the upper, middle and lower dashed box regions in a. Note that TH-positive neurons are well-labeled and clearly visible even at a depth of 2,460 µm in the intact brain. CPu, caudate putamen; PO, preoptic nucleus; VTA, ventral tegmental area; SNR, substantia nigra; RR, retrorubral nucleus; DR, dorsal raphe. Scale bars, 700 µm (a) and 100 µm (b–d). Image adapted from ref. 1. nature methods | VOL.10 NO.6 | JUNE 2013 | 509 perspective npg © 2013 Nature America, Inc. All rights reserved. Figure 3 | CLARITY technology and instrumentation. (a) Tissue is crosslinked with formaldehyde in the presence of infused hydrogel monomers. Thermally triggered polymerization then results in a hydrogel-tissue hybrid which physically supports tissue structure and chemically incorporates native biomolecules into the hydrogel mesh. (b) In electrophoretic tissue clearing (ETC), an electric field is applied across the hybrid immersed in an ionic detergent solution to actively transport ionic micelles into the hybrid and extract membrane lipids out of the tissue, leaving structures and cross-linked biomolecules in place and available for imaging and molecular phenotyping1. The ETC setup consists of the custom ETC chamber, a temperature-controlled buffer circulator (RE415, Lauda), a buffer filter (McMaster) and a power supply (Bio-Rad). The sample is electrophoresed by applying 20–60 volts to the electrodes. Buffer solution is circulated through the chamber to maintain temperature and the composition of the buffer solution constant throughout the clearing process. The cut-through view (bottom right) shows placement of the hydrogel-embedded tissue in the sample holder (Cell Strainer, BD Biosciences) located in the middle of the chamber between the two electrodes. The end of each electrode exposed outside the chamber is connected to a power supply. Image adapted from ref. 1. focus on mapping the brain a Step 1: hydrogel monomer infusion (days 1–3) Proteins DNA ER 4 °C Vesicle 510 | VOL.10 NO.6 | JUNE 2013 | nature methods + Formaldehyde Hydrogel monomer Plasma membrane Step 2: hydrogel-tissue hybridization (day 3) 37 °C Hydrogel b Step 3: ETC (days 5–8) Extracted lipids in SDS micelle SDS micelle Here we address opportunities and challenges for CLARITY methods in brain mapping, in combination with other emerging genetic and imaging technologies. Imaging methods for CLARITY CLARITY is compatible with most fluorescence microscopy techniques. Conventional laser-scanning approaches, such as confocal and multiphoton microscopies, are well-suited to image samples prepared using CLARITY because the excitation and emission wavelengths of light involved can penetrate deep into the transparent tissue. Objective working distance can be limiting in this setting, although already single-photon imaging of the intact clarified adult mouse brain with a 3.6-millimeter working distance objective has been achieved without noticeable degradation in resolution (Fig. 1)1. Even greater depth of imaging, for larger brains, tissues or organisms, is possible with longer working distance and high-numerical-aperture objectives (such as the 5–8-millimeter working distance versions available from several manufacturers), but it will be important for optics to be refined and for objectives to be developed that are matched to the refractive index of clarified tissue to minimize aberration and maintain both high resolution and long working distance. Single-photon laser-scanning confocal imaging is in many respects appropriate for CLARITY, given the broad range of suitable fluorescent labels available for multicolor confocal interrogation of clarified tissue. But confocal microscopy does expose the entire depth of tissue to excitation light and therefore induces substantial photobleaching of fluorescent molecules, a particularly acute problem in the setting of slow, high-resolution whole-brain imaging. Nonlinear imaging techniques such as two-photon microscopy address this out-of-focus photobleaching issue, but both single and two-photon imaging methods suffer from low image acquisition rate. With these scanning methods, imaging a mouse brain may take days at single-cell resolution and months at single-neurite resolution—a key practical issue when dealing with the larger samples that are characteristic of CLARITY approaches. For imaging large samples at high resolution and at high dataacquisition rates, we suggest that selective-plane illumination microscopy (SPIM) may be a method of choice. In this technique, + Temperature-controlled buffer circulator Top view: cut through Electrode ETC chamber Target tissue Buffer filter Inlet port Electrode Electrophoresis power supply Electrode connector the sample is illuminated with a thin sheet of light from the side and fluorescence emission is collected along an axis perpendicular to the plane of illumination by wide-field acquisition26–31. This unique illumination and imaging modality substantially reduces both photobleaching and imaging time (for imaging whole mouse brain, closer to hours at single-cell resolution, and days at single-neurite resolution). CLARITY will naturally work well with SPIM; for example, the high and uniform transparency of clarified tissue will minimize light-plane broadening that causes degradation in resolution toward the center of the tissue in SPIM and therefore enable uniform high-resolution imaging of large samples. Other rapidly developing imaging technologies, such as structured illumination and Bessel-beam illumination31, will enhance resolution and may potentially allow imaging in large and intact tissue specimens even beyond the diffraction limit. The CLARITY-SPIM combination of high resolution, independence from mechanical sectioning and reconstruction, and fast dataacquisition rate may become particularly useful for distinguishing and tracing neural projections with high throughput and accuracy in the course of generating maps of brain connectivity. When indicated, after global maps of intact systems are generated, individual portions of tissue such as the downstream target of an imaged projection may be extracted for analyses such as electron microscopy that require sectioning. CLARITY is compatible with subsequent electron microscopy and preserves some ultrastructural features such as postsynaptic densities1, although because of the absence of lipid, conventional electron microscopy staining does not currently provide enough contrast to identify all npg © 2013 Nature America, Inc. All rights reserved. focus on mapping the brain relevant ultrastructures and boundaries, pointing to the need for optimization of the electron microscopy preparation and staining process. Alternatively, when combined with array tomography and/or super-resolution imaging in neural applications10,32, clarified tissue may provide for nanoscopic molecular characterization of synaptic connections defined by paired presynaptic and postsynaptic proteins, so that the synaptic wiring logic of incoming projections to a brain region can be established in detail. Of course, no single imaging modality that may be used with CLARITY provides all potentially important information on functional connectivity. For example, conventional light microscopy may not readily provide definitive identification of synaptic contacts, whereas electron microscopy analysis does not typically reveal rich molecular information on detected synapses relevant to functional properties, nor would an absence of synaptic junctions between cells identified by any method demonstrate absence of direct cellular communication, which can occur via nonlocal and volume-transmission influences, as in the case of neuro modulators or extrasynaptic neurotransmitter action. Consistent with possible integrative strategies, clarified tissue appears compatible with diverse imaging readout modalities that complement each other; moreover, the multiround molecular phenotyping capacity of CLARITY (see below) may be suitable for providing enriched detail on the composition and relationships of sub cellular structures such as synapses33. Integrating circuit maps with molecular information A unique feature of CLARITY is its potential for intact-tissue molecular phenotyping studies. The hydrogel-tissue hybridization preserves endogenous biomolecules ranging from neurotransmitters to proteins and nucleic acids1; soluble proteins and cell-membrane proteins alike are secured by chemical tethering to the hydrogel mesh. Moreover, removal of the lipid membrane makes this retained molecular information accessible via passive diffusion of macromolecular probes into the tissue, and the enhanced structural integrity of clarified tissue allows multiple rounds of antibody staining, elution or destaining and restaining that are not typically feasible with conventionally fixed tissue1,34. Immunostaining of 1-millimeter-thick clarified tissue blocks takes days, and immunostaining of entire intact adult mouse brains is possible on practical timescales of several weeks (Fig. 2). This process could be accelerated by increasing probe diffusion rate with electrophoresis or other methods. The fact that CLARITY supports multiple (at least three) rounds of molecular phenotyping may be of value to define cell types and to link form and function in brain-mapping studies, allowing the integration of rich local and global morphological details (for example, type of synapse, shape of cell body, and information about dendritic arborization and axonal projections) with genetic fingerprints. In addition, the combination of CLARITY with genetic methods for identification of synapses or for labeling specific circuits is also of interest. First, GFP reconstitution across synaptic partners (GRASP) has emerged as a light microscopy–based synapse-detection technique35. In this method, two nonfluorescent split-GFP fragments can be virally expressed in the synaptic membrane of two separate neuronal populations; these two fragments reconstitute fluorescent GFP only across a synaptic cleft so that the location of synapses between the two populations perspective can be visualized35,36. This approach was validated with electron microscopy and was used to map synaptic connectivity in the nematode35 and the fruit fly37; most recently, applicability to mammalian systems was demonstrated: mammalian GRASP (mGRASP) revealed labeling of both excitatory and inhibitory synapses38. Second, in recent years many mouse lines and viral targeting techniques have been developed for intersectional labeling of spatially, genetically, synaptically and functionally defined neuronal circuit elements. For instance, an engineered rabies virus encoding EGFP can be injected into a particular brain region to label neurons that are presynaptic to the injection site39–41. More recently, activity-dependent expression of channelrhodopsin (ChR2) conjugated with EYFP was used to label and drive a subset of synaptically recruited neurons associated with fear memory42, and a technique has been developed and validated that allows permanent marking (by filling a cell with fluorescent protein labels) of neurons across the mouse brain that were active during a relatively restricted time window (on a time scale of hours)43. These versatile targeting approaches, in combination with CLARITY, could provide a high-throughput approach for globally mapping synaptically connected and synaptically activated populations across the brain that could complement the use of electron microscopy in some settings. Although new genetic strategies such as these are quite power ful, for studying and mapping the human brain, more conventional antibody-based (nongenetic) molecular phenotyping alone is particularly important. Lipophilic dyes (for example, 1,1′-dioctadecyl-3,3,3′3′-tetramethylindocarbocyanine perchlor ate) can be injected into postmortem human brain to effectively label local projections through passive diffusion44; however, in fixed tissue without active axonal transport, diffusion rates (and therefore tracing distance) are substantially limited and may not be ideally suited to mapping long-range connectivity. With CLARITY, molecular markers have already been used to identify individual structures and projections (cell bodies and fibers) in human tissue on the millimeter length scale1; although this approach has not yet been tested for tracking long-range projections in the human brain, neurofilament protein staining reliably highlights a major subset of axonal fibers, and because continuity of labeled projections is preserved in the intact tissue, fibers can be traced with diminished risk of alignment and/or reconstruction error. Cell type–specific markers (for example, parvalbumin and tyrosine hydroxylase) can also be used to label both cell bodies and projections in human brain tissue1. CLARITY may in this way help unlock a rich source of clinically relevant information, providing a means to interrogate and make accessible brain-bank samples that are unique or rare, for the purpose of understanding disease mechanisms as well as the native structure, organization and complexity of the human brain. Limitations, challenges and opportunities CLARITY is currently in the early stages of development; innovation will be needed over the coming years. First, we have noted that tissues can expand after electrophoretic tissue clearing and return to the original size after refractive-index matching1; although macroscopic observation indicates that the changes in volume are largely isotropic and reversible, quantitative monitoring at microscopic and nanoscopic resolution will be required to confirm that loss in structural connectivity or occurrence of nature methods | VOL.10 NO.6 | JUNE 2013 | 511 npg © 2013 Nature America, Inc. All rights reserved. perspective other tissue artifacts is minimal. Second, although adherence to the described protocol1 will minimize tissue damage, the precise extent to which CLARITY under best current practice secures specific molecular information represented in proteins, nucleic acids and small molecules must be further explored. Only ~8% of total protein content is lost in CLARITY1, substantially lower than with other methods over the same timescale (for example, ~24% loss with conventional fixative (paraformaldehyde) and detergent (0.1% Triton X-100) histological solution compositions 1). However, the nature of the remaining loss should be investigated, and we note that the current chemistry of CLARITY (by design) does not preserve lipids or other molecules lacking functional groups required for chemical tethering to the hydrogel mesh, such as phosphatidylinositol 4,5-bisphosphate, or PIP2. Additional chemical-treatment strategies may need to be explored for specialized experimental questions. It also remains to be determined how accessible different classes of biological information from clarified tissue may be, in clinical or animal settings, for high-content quantitative data-extraction pipelines. Mapping this landscape will be of value for developing strategies to maximize information extraction after global (brainwide) projection mapping has been completed, and may be carried out in the context of normal function, disease states and treatment regimens (in mapping effects of brain-stimulation treatments or in drug screening, for example). Another open question is how long clarified tissue may be maintained or stored for such analysis, assessment, imaging or remodeling in subsequent rounds of CLARITY processing. CLARITY can be seen as a prototype for a general approach for building new structures and installing new functions from within biological systems—an approach that may find other instantiations as the technology is developed for introducing components or monomers into biological tissue that are then triggered subsequently to form a polymer, gel, mesh, network or assembled structure with desired physical parameters (for example, stiffness, transparency, pore size, conductivity and permeability) or active properties (for interfacing, catalysis or functionalization). Relevant exogenous components may include proteins, oligonucleotides, stains, chemicals or even small mechanical, electronic or optical components. These introduced components could be designed for either constitutive or inducible functionality—in the latter case, such that they can be activated by external elements including heat, mechanical force, redox changes, electromagnetic triggers such as light, and other accelerators, thereby enabling temporally precise initiation of the structural and functional tissue transformation. The capacity to trigger functionality could help enable versions of this general approach (not involving lipid removal) that may be compatible with ongoing vital functions of the tissue. On the brain-mapping front, as pioneering technologies such as array tomography and serial block face scanning electron microscopy have already proven, turning large data sets into useful and tractable deliverables still poses an immense challenge. Relevant computational approaches for registration in three dimensions will need to be developed, and particularly in the setting of variable, quenching and/or bleaching fluorescent signals, automated image acquisition, segmentation and tracing algorithms will be helpful. We note that molecular landmarks are useful for navigating three-dimensional data, and here the intact-tissue 512 | VOL.10 NO.6 | JUNE 2013 | nature methods focus on mapping the brain olecular phenotyping capability of CLARITY becomes particm ularly important. As in conventional histology, broadly staining markers (which are dense yet distinct) can be useful as landmarks to navigate and identify brain regions of interest45. In this regard, DAPI staining is compatible with CLARITY1 and appears particularly useful as a Nissl-like registration method. Registration of CLARITY data with other data is both a challenge and an opportunity. For example, in an animalsubject brain in which a particular set of neurons becomes known (via imaging of activity 46–49 and/or via optogenetic control4) to be involved in a specific behavioral function or dysfunction, to then clarify the same preparation and obtain brain-wide wiring and molecular information on those same cells (and their connection partners) would be of substantial value. This approach could be used to address fundamental questions in both basic and clinical or preclinical neuroscience but will require the development of efficient workflows for the registration of the different types of experimental data. CLARITY may also help to conceptually link future high-resolution activity maps50 with structural or functional macroscale maps (for example, http:// www.neuroscienceblueprint.nih.gov/connectome/), providing an anatomical foundation that could help researchers decipher the meaning of brain-activity patterns linked to health and disease4,50. Beyond neuroscience, CLARITY is currently being explored for the evaluation, diagnosis and prognosis of pathological states including cancer, infection, autoimmune disease and other clinical conditions as well as for the study of normal tissue, organ and organism function, including development and relationships of cells and tissues. Resources that may help enable the general user to establish the methodology are available online (http://clarityresourcecenter.org/). Acknowledgments We acknowledge all members of the Deisseroth laboratory for discussions and support. This work was funded by a US National Institutes of Health Director’s Transformative Research Award (TR01) to K.D. from the National Institute of Mental Health, as well as by the National Science Foundation, the Simons Foundation, the President and Provost of Stanford University, and the Howard Hughes Medical Institute. K.D. is also funded by the National Institute on Drug Abuse and the Defense Advanced Research Projects Agency Reorganization and Plasticity to Accelerate Injury Recovery program, and the Wiegers, Snyder, Reeves, Gatsby, and Yu Foundations. K.C. is supported by the Burroughs Wellcome Fund Career Award at the Scientific Interface. COMPETING FINANCIAL INTERESTS The authors declare competing financial interests: details are available in the online version of the paper. Reprints and permissions information is available online at http://www.nature. com/reprints/index.html. 1. Chung, K. et al. Structural and molecular interrogation of intact biological systems. Nature advance online publication, doi:10.1038/nature12107 (10 April 2013). 2. Petersen, C.C.H. The functional organization of the barrel cortex. Neuron 56, 339–355 (2007). 3. Mombaerts, P. et al. Visualizing an olfactory sensory map. Cell 87, 675–686 (1996). 4. Deisseroth, K. Optogenetics and psychiatry: applications, challenges, and opportunities. Biol. Psychiatry 71, 1030–1032 (2012). 5. Kasthuri, N. & Lichtman, J.W. The rise of the ‘projectome’. Nat. Methods 4, 307–308 (2007). 6. Nicholson, C. Diffusion in brain extracellular space. Brain 6, 1277–1340 (2008). 7. Cheong, W., Prahl, S. & Welch, A. A review of the optical properties of biological tissues. IEEE J. Quantum Electronics 26, 2166–2185 (1990). npg © 2013 Nature America, Inc. All rights reserved. focus on mapping the brain 8. Helmchen, F. & Denk, W. Deep tissue two-photon microscopy. Nat. Methods 2, 932–940 (2005). 9. Rauschning, W. Surface cryoplanning: a technique for clinical anatomical correlations. Uppsala J. Med. Sci. 91, 251–255 (1986). 10. Toga, A.W., Ambach, K., Quinn, B., Hutchin, M. & Burton, J.S. Postmortem anatomy from cryosectioned whole human brain. J. Neurosci. Methods 54, 239–252 (1994). 11. Ewald, A.J., McBride, H., Reddington, M., Fraser, S.E. & Kerschmann, R. Surface imaging microscopy, an automated method for visualizing whole embryo samples in three dimensions at high resolution. Dev. Dyn. 225, 369–375 (2002). 12. McCormick, B.H. et al. Construction of anatomically correct models of mouse brain networks. Neurocomputing 58–60, 379–386 (2004). 13. Li, A. et al. Micro-optical sectioning tomography to obtain a highresolution atlas of the mouse brain. Science 330, 1404–1408 (2010). 14. Tsai, P. et al. All-optical histology using ultrashort laser pulses. Neuron 39, 27–41 (2003). 15. Ragan, T. et al. Serial two-photon tomography for automated ex vivo mouse brain imaging. Nat. Methods 9, 255–258 (2012). 16. Denk, W. & Horstmann, H. Serial block-face scanning electron microscopy to reconstruct three-dimensional tissue nanostructure. PLoS Biol. 2, e329 (2004). 17. Micheva, K.D. & Smith, S.J. Array tomography: a new tool for imaging the molecular architecture and ultrastructure of neural circuits. Neuron 55, 25–36 (2007). 18. Osten, P. & Margrie, T.W. Mapping brain circuitry with a light microscope. Nat. Methods 10, 515–523 (2013). 19. Livet, J. et al. Transgenic strategies for combinatorial expression of fluorescent proteins in the nervous system. Nature 450, 56–62 (2007). 20. Tsai, P. et al. Correlations of neuronal and microvascular densities in murine cortex revealed by direct counting and colocalization of cell nuclei and microvessels. J. Neurosci. 18, 14553–14570 (2009). 21. Kleinfeld, D. et al. Large-scale automated histology in the pursuit of connectomes. J. Neurosci. 31, 16125–16138 (2011). 22. Bock, D.D. et al. Network anatomy and in vivo physiology of visual cortical neurons. Nature 471, 177–182 (2011). 23. Briggman, K.L., Helmstaedter, M. & Denk, W. Wiring specificity in the direction-selectivity circuit of the retina. Nature 471, 183–188 (2011). 24. Dodt, H., Leischner, U. & Schierloh, A. Ultramicroscopy: three-dimensional visualization of neuronal networks in the whole mouse brain. Nat. Methods 4, 331–336 (2007). 25. Ertürk, A., Mauch, C., Hellal, F. & Förstner, F. Three-dimensional imaging of the unsectioned adult spinal cord to assess axon regeneration and glial responses after injury. Nat. Med. 18, 166–171 (2012). 26. Hama, H. et al. Scale: a chemical approach for fluorescence imaging and reconstruction of transparent mouse brain. Nat. Neurosci. 14, 1481–1488 (2011). 27. Keller, P.J. et al. Fast, high-contrast imaging of animal development with scanned light sheet-based structured-illumination microscopy. Nat. Methods 7, 637–642 (2010). 28. Truong, T.V., Supatto, W., Koos, D.S., Choi, J.M. & Fraser, S.E. Deep and fast live imaging with two-photon scanned light-sheet microscopy. Nat. Methods 8, 757–760 (2011). perspective 29. Tomer, R., Khairy, K., Amat, F. & Keller, P.J. Quantitative high-speed imaging of entire developing embryos with simultaneous multiview light-sheet microscopy. Nat. Methods 9, 755–763 (2012). 30. Krzic, U., Gunther, S., Saunders, T.E., Streichan, S.J. & Hufnagel, L. Multiview light-sheet microscope for rapid in toto imaging. Nat. Methods 9, 730–733 (2012). 31. Keller, P.J., Schmidt, A.D., Wittbrodt, J. & Stelzer, E.H.K. Reconstruction of zebrafish early embryonic development by scanned light sheet microscopy. Science 322, 1065–1069 (2008). 32. Planchon, T.A. et al. Rapid three-dimensional isotropic imaging of living cells using Bessel beam plane illumination. Nat. Methods 8, 417–423 (2011). 33. Dani, A., Huang, B., Bergan, J., Dulac, C. & Zhuang, X. Superresolution imaging of chemical synapses in the brain. Neuron 68, 843–856 (2010). 34. Micheva, K.D., Busse, B., Weiler, N.C., O’Rourke, N. & Smith, S.J. Single-synapse analysis of a diverse synapse population: proteomic imaging methods and markers. Neuron 68, 639–653 (2010). 35. Wählby, C. & Erlandsson, F. Sequential immunofluorescence staining and image analysis for detection of large numbers of antigens in individual cell nuclei. Cytometry 47, 32–41 (2002). 36. Feinberg, E.H. et al. GFP reconstitution across synaptic partners (GRASP) defines cell contacts and synapses in living nervous systems. Neuron 57, 353–363 (2008). 37. Wickersham, I.R. & Feinberg, E.H. New technologies for imaging synaptic partners. Curr. Opin. Neurobiol. 22, 121–127 (2012). 38. Gordon, M.D. & Scott, K. Motor control in a Drosophila taste circuit. Neuron 61, 373–384 (2009). 39. Kim, J. et al. mGRASP enables mapping mammalian synaptic connectivity with light microscopy. Nat. Methods 9, 96–102 (2012). 40. Wickersham, I. et al. Retrograde neuronal tracing with a deletion-mutant rabies virus. Nat. Methods 4, 47–49 (2007). 41. Miyamichi, K. et al. Cortical representations of olfactory input by trans-synaptic tracing. Nature 472, 191–196 (2011). 42. Liu, X. et al. Optogenetic stimulation of a hippocampal engram activates fear memory recall. Nature 484, 381–385 (2012). 43. Guenthner, C.J., Miyamichi, K., Yang, H., Heller, H.C. & Luo, L. Permanent genetic access to transiently active neurons using targeted recombination in active populations (TRAP). Neuron (in the press). 44. Godement, P., Vanselow, J., Thanos, S. & Bonhoeffer, F. A study in developing visual systems with a new method of staining neurones and their processes in fixed tissue. Development 101, 697–713 (1987). 45. Helmstaedter, M. Cellular-resolution connectomics: challenges of dense neural circuit reconstruction. Nat. Methods 10, 501–507 (2013). 46. Dombeck, D.A., Graziano, M.S. & Tank, D.W. Functional clustering of neurons in motor cortex determined by cellular resolution imaging in awake behaving mice. J. Neurosci. 29, 13751–13760 (2009). 47. Ahrens, M.B. et al. Brain-wide neuronal dynamics during motor adaptation in zebrafish. Nature 485, 471–477 (2012). 48. Ziv, Y. et al. Long-term dynamics of CA1 hippocampal place codes. Nat. Neurosci. 16, 264–266 (2013). 49. Ahrens, M.B., Orger, M.B., Robson, D.N., Li, J.M. & Keller, P.J. Whole-brain functional imaging at cellular resolution using light-sheet microscopy. Nat. Methods 10, 413–420 (2013). 50. Alivisatos, P. et al. The brain activity map. Science 339, 1284–1285 (2013). nature methods | VOL.10 NO.6 | JUNE 2013 | 513 FOCUS ON MAPPING THE BRAINreview Mapping brain circuitry with a light microscope npg © 2013 Nature America, Inc. All rights reserved. Pavel Osten1 & Troy W Margrie2,3 The beginning of the 21st century has seen a renaissance in light microscopy and anatomical tract tracing that together are rapidly advancing our understanding of the form and function of neuronal circuits. The introduction of instruments for automated imaging of whole mouse brains, new cell type–specific and trans-synaptic tracers, and computational methods for handling the whole-brain data sets has opened the door to neuroanatomical studies at an unprecedented scale. We present an overview of the present state and future opportunities in charting long-range and local connectivity in the entire mouse brain and in linking brain circuits to function. Since the pioneering work of Camillo Golgi and Santiago Ramón y Cajal at the turn of the last century1,2, advances in light microscopy (LM) and neurotracing methods have been central to the progress in our understanding of anatomical organization in the mammalian brain. The Golgi silver-impregnation method allowed the visualization of neuron morpho logy, providing the first evidence for cell type–based and connectivity-based organization in the brain. The introduction of efficient neuroanatomical tracers in the second half of the last century greatly increased the throughput and versatility of neuronal projection mapping, which led to the identification of many anatomical pathways and circuits, and revealed the basic principles of hierarchical and laminar connectivity in sensory, motor and other brain systems3,4. The beginning of this century has seen a methodsdriven renaissance in neuroanatomy, one that is distinguished by a focus on large-scale projects that yield unprecedented amounts of anatomical data. Instead of the traditional ‘cottage-industry’ approach to studying one anatomical pathway at a time, the new projects aim to generate complete data sets—so-called projectomes and connectomes—that can be used by the scientific community as resources for answering specific experimental questions. These efforts range in scale and resolution from the macroscopic to the microscopic: from studies of the human brain by magnetic resonance imaging to reconstructions of dense neural circuits in small volumes of brain tissue by electron microscopy (see Review5 and Perspective6 in this Focus). Advancements in LM methods, the focus of this Review, are being applied to the mapping of point-topoint connectivity between all anatomical regions in the mouse brain by means of sparse reconstructions of anterograde and retrograde tracers7. Taking advantage of the automation of LM instruments, powerful dataprocessing pipelines, and combinations of traditional and modern viral vector-based tracers, teams of scientists at Cold Spring Harbor Laboratory (CSHL), Allen Institute for Brain Science (AIBS) and University of California Los Angeles (UCLA) are racing to complete a connectivity map of the mouse brain—dubbed the ‘mesoscopic connectome’—which will provide the scientific community with online atlases for viewing entire anatomical data sets7. These efforts demonstrate the transformative nature of today’s LM-based neuroanatomy studies and the astonishing speed with which large amounts of data can be disseminated online, and have an immediate impact on research in neuroscience laboratories around the world. As the mouse mesoscopic connectomes are being completed, it is clear that LM methods will continue to impact the evolution of biological research and specifically neuroscience: new trans-synaptic viral tracers are being engineered to circumvent the need to resolve synapses, which has constrained the interpretation of cell-to-cell connectivity in LM studies, and new assays 1Cold Spring Harbor Laboratory, Cold Spring Harbor, New York, USA. 2Department of Neuroscience, Physiology and Pharmacology, University College London, London, UK. 3Division of Neurophysiology, The Medical Research Council National Institute for Medical Research, London, UK. Correspondence should be addressed to P.O. ([email protected]) or T.W.M. ([email protected]). Received 2 march; accepted 15 April; published online 30 may 2013; doi:10.1038/nmeth.2477 nature methods | VOL.10 NO.6 | JUNE 2013 | 515 review b Ti:sapphire PMT c PMT PMT Camera Piezo element Microtome y Camera 515-nm laser Detection objective Illumination objective Knife Light sheet y x y x x z z npg © 2013 Nature America, Inc. All rights reserved. Motorized x-y-z stage Motorized x-y-z stage Motorized x-y-z stage z J. Kuhl a FOCUS ON MAPPING THE BRAIN Figure 1 | Whole-brain LM methods. (a) In STP tomography, a two-photon microscope is used to image the mouse brain in a coronal plane in a mosaic grid pattern, and a microtome sections off the imaged tissue. Piezo objective scanner can be used for z-stack imaging (image adapted from ref. 15). (b) In fMOST, confocal line-scan is used to image the brain as 1-micrometer thin section cut by a diamond knife (image adapted from ref. 16). (c) In LSFM, the cleared brain is illuminated from the side with a light sheet (blue) through an illumination objective (or cylinder lens 19) and imaged in a mosaic grid pattern from the top (image adapted from ref. 20). In all instruments, the brain is moved under the objective on a motorized x-y-z stage; PMT, photomultiplier tube. combining anatomical and functional measurements are being applied to bridge the traditional structure-function divide in the study of the mammalian brain. In this Review, we aim to provide an overview of today’s state of the art in LM instrumentation and to highlight the opportunities for progress as well as the challenges that need to be overcome to transform neuronal-tracing studies into a truly quantitative science that yields comprehensive descriptions of long-range and local projections and connectivity in whole mouse brains. We also discuss present strategies for the integration of anatomy and function in the study of mouse brain circuits. Automated light microscopes for whole-brain imaging The field of neuroanatomy has traditionally been associated with labor-intensive procedures that greatly limit the throughput of data collection. Recent efforts to automate LM instrumentation have standardized and dramatically increased the throughput of anatomical studies. The main challenge for these methods is to maintain the rigorous quality of traditional neuroanatomical studies, which results from detailed visual analysis, careful data collection and expert data interpretation. There are currently two approaches to the automation of LM for imaging three-dimensional (3D) whole-brain data sets: one based on the integration of block-face microscopy and tissue sectioning and the other based on light-sheet fluorescence micros copy (LSFM) of chemically cleared tissue. The first approach has been developed for wide-field imaging, line-scan imaging, confocal microscopy and two-photon microscopy8–16. Common to all these instruments is the motorized movement of the sample under the microscope objective for top-view mosaic imaging, followed by mechanical removal of the imaged tissue before the next cycle of interleaved imaging and sectioning steps (Fig. 1a,b). As the objective is always near the tissue surface, it is possible to use high-numerical-aperture lenses to achieve submicrometer resolution close to the diffraction limits of LM. 516 | VOL.10 NO.6 | JUNE 2013 | nature methods Three instruments have been designed that combine twophoton microscopy17 with subsequent tissue sectioning by ultrashort laser pulses in all-optical histology10, by a milling machine in two-photon tissue cytometry12 or by a vibrating blade microtome in serial two-photon (STP) tomography15 (Fig. 1a). Whereas in both all-optical histology and two-photon tissue cytometry the sectioning obliterates the imaged tissue, the integration of vibratome-based sectioning in STP tomography allows the collection of the cut tissue for subsequent analysis by, for example, immunohistochemistry (see below). In addition, the tissue preparation by simple formaldehyde fixation and agar embedding in STP tomography has minimal detrimental effects on fluorescence and brain morphology. This makes STP tomography applicable to a broad range of neuroanatomical projects that use genetically encoded fluorescent protein–based tracers, which are sensitive to conditions used for fixation, dehydration and tissue clearing. This method is also versatile in terms of the mode and resolution of data collection. For example, imaging the mouse brain as a data set of 280 serial coronal sections, evenly spaced at 50 micrometers and at x-y resolution of 1 micrometer, takes about ~21 hours and generates a brain atlas–like data set of ~70 gigabytes. A complete visualization can be achieved by switching to 3D scanning of z-volume stacks between the mechanical sectioning steps, which allows the entire mouse brain to be imaged, for instance, at 1-micrometer x-y resolution and 2.5-micrometer z resolution in ~8 days, generating ~1.5 terabytes of data15. The instrument is commercially available from TissueVision Inc. The Allen Brain Institute is using this methodology for its Mouse Connectivity project (see below). Two instruments have been designed to combine bright-field linescan imaging and ultramicrotome sectioning of resin-embedded tissue in methods named knife-edge scanning microscopy (known as KESM)13 and micro-optical sectioning tomography (MOST)14 (Fig. 1b). The latter was used to image Golgi-stained mouse brain npg © 2013 Nature America, Inc. All rights reserved. FOCUS ON MAPPING THE BRAINreview at 0.33 × 0.33 × 1.0 micrometer x-y-z resolution, generating >8 terabytes of data in ~10 days13,14. The MOST instrument design was also recently built for fluorescence imaging (fMOST) by confocal laser scanning microscopy, with the throughput of one mouse brain at 1.0-micrometer voxel resolution in ~19 days16. Knife-edge scanning microscopy imaging is now also available as a commercial service from 3Scan. The second approach for automated whole-brain imaging is based on LSFM (also known as selective-plane illumination microscopy18 and ultramicroscopy19; Fig. 1c). This approach allows fast imaging of chemically cleared ‘transparent’ mouse brains without the need for mechanical sectioning19,20 but, at least until now, with some trade-offs for anatomical tracing applications. The chemical clearing procedures reduced the signal of fluorescent proteins, but this problem appears to be solved by a new hydrogel-based tissue transformation and clearing method termed CLARITY21 (see Perspective about this methodology in this Focus22). The spatial resolution of LSFM for the mouse brain also has been limited by the requirement for large field-of-view objectives with low power and low numerical aperture that were used to visualize the whole brain19,23. However, new objectives with long working distance and high numerical aperture, such as 8-millimeter working distance and 0.9 numerical aperture objective from Olympus, promise to enable LSFM of the whole mouse brain at submicrometer resolution. If necessary, LSFM can also be combined with one of several forms of structured illumination to reduce out-of-focus background fluorescence and improve contrast24–26. Taken together, these modifications are likely to enhance the applicability of LSFM to anterograde tracing of thin axons at high resolution in the whole mouse brain, as done by STP tomography in the AIBS Mouse Connectivity project (see below) and by fMOST in a recent report16. In addition, LSFM is well-suited for retrograde tracing in the mouse brain, which relies on detection of retrogradely fluorescence-labeled neuronal soma that are typically >10 micrometers in diameter. Such application was recently demonstrated for mapping retrograde connectivity of granule cells of the mouse olfactory bulb20 using rabies viruses that achieve high levels of fluorescent protein labeling27,28. Mesoscopic connectivity-mapping projects The labeling of neurons and subsequent neuroanatomical tract tracing by LM methods has been used for over a century to interrogate the anatomical substrate of the transmission of information in the brain. Throughout those years, the credo of neuroanatomy, ‘the gain in brain is mainly in the stain’, signified that progress was made mainly through the development of new anatomical tracers. Yet despite the decades of neuroanatomical research, the laborious nature of tissue-processing and datavisualization has kept the progress in our knowledge of brain circuitry at a disappointingly slow pace7. Today, neuroanatomy stands to greatly benefit from the application of high-throughput automated LM instruments and powerful informatics tools for the analysis of mouse brain data29,30. The high-resolution capacity LM methods afford, and the fact that an entire brain data set can be captured, makes these systems well-suited for the systematic charting of the spatial profile and the connectivity of populations of neurons and even individual cells projecting over long distances. The pioneering effort in the field of anatomical projects applied at the scale of whole animal brains was the Allen Mouse Brain Atlas of Gene Expression, which cataloged in situ hybridization maps for more than 20,000 genes in an online 3D digital mouse brain atlas29,31,32. The proposal by a consortium of scientists led by Partha Mitra to generate similar LM-based atlases of ‘brainwide neuroanatomical connectivity’ in several animal models7 has in short time spurred three independent projects, each promising to trace all efferent and afferent anatomical pathways in the mouse brain. The aim of the Mouse Brain Architecture Project (http://brainarchitecture.org/) from the Mitra team at CSHL is to image >1,000 brains; the Allen Mouse Brain Connectivity Atlas project (http://connectivity.brain-map.org/) led by Hongkui Zeng at AIBS has a goal of imaging >2,000 brains; and the Mouse Connectome Project (http://www.mouseconnectome.org/) led by Hong-Wei Dong at UCLA has a goal of imaging 500 brains, with each brain injected with four tracers. Whereas the CSHL and UCLA projects rely on automated wide-field fluorescence microscopy (Hamamatsu Nanozoomer 2.0 and Olympus VS110, respectively) to image manually sectioned brains, the Mouse Connectivity project at the Allen Institute is being done entirely by STP tomography15. The main strength of these efforts is in the broad range of tracers used. Given that each tracer has its own advantages and problems33, the information derived from all three projects will ensure generalizable interpretation of the projection results throughout the brain. The CSHL group uses a combination of traditional anterograde and retrograde tracers, fluorophoreconjugated dextran amine34 and cholera toxin B (CTB) subunit35, respectively, which are complemented by a combination of viral vector–based tracers, GFP-expressing adeno-associated virus (AAV)36 for anterograde tracing (Fig. 2a) and modified rabies virus27 for retrograde tracing. Although the virus-based methods are less well tested, they offer advantages in terms of the brightness of labeling and the possibility of cell type–specific targeting using Cre recombinase–dependent viral vectors37 and transgenic lines expressing Cre recombinase from cell type–specific promoters38–40. The AIBS team uses solely anterograde tracing by AAV-GFP viruses that label axonal arborizations with GFP 41 (Fig. 2b), in many cases taking advantage of transgenic ‘driver’ mouse lines expressing Cre recombinase from cell type–specific promoters to achieve anterograde tracing of specific neuronal cell types. Finally, the team at UCLA is using a strategy of two injections per brain, each with a mix of anterograde and retrograde tracers42: CTB together with Phaseolus vulgaris leucoagglutinin43 and FluoroGold44 together with biotinylated dextran amine42,45. This approach has an added advantage of enabling direct visualization of the convergence of inputs and outputs from across different areas in one brain42,46,47. The unprecedented amounts of data being collected in these projects means that the considerable person-hours historically spent performing microscopy have largely shifted toward data analysis. The first step of such data analysis comprises the compilation of the serial section images for viewing as whole-brain data sets at resolutions beyond the minimum geometric volume of the neuronal structures of interest: soma for retrograde tracing and axons for anterograde tracing. All three projects offer a convenient way to browse the data sets online, including high-resolution magnified views that in most cases are sufficient to visually determine labeled soma and axons. All three projects use the Allen nature methods | VOL.10 NO.6 | JUNE 2013 | 517 review a npg © 2013 Nature America, Inc. All rights reserved. b Figure 2 | Primary motor cortex projection maps. (a) Mouse Brain Architecture data of AAV-GFP injected into the supragranular layers and AAV–red fluorescent protein injected in the infragranular layers (F. Mechler and P. Mitra; unpublished data). Front (left) and lateral (right) views of the volume-rendered brain (top); and coronal section image from the area marked by the dashed line (center) with magnification of the lower boxed region showing axonal fibers in the cerebral peduncle (left) and magnification of the upper boxed region showing projections to the midbrain reticular nucleus (right). Scale bars, 1,000 µm (top) and 20 µm (bottom). (b) Mouse Connectivity data of a similar AAV-GFP injection show the primary motor cortex projectome reconstructed in the Allen Brain Explorer48 (H. Zeng; unpublished data). Inset, magnified view and coronal section overview of projections in the ventral posteromedial (VPM) nucleus of the thalamus. Mouse Brain Atlas for the registration of the coronal sections, which will help in the cross-validation of results obtained from the different tracers. The Allen Mouse Brain Connectivity Atlas website also offers the option to view the data after projection segmentation, which selectively highlights labeled axons, as well as in 3D in the Brain Explorer registered to the Allen Mouse Brain Atlas48 (Fig. 2b). The second step of data analysis requires the development of informatics methods for quantitation of the data sets, which will facilitate the interpretation of the data available online. The Allen Mouse Brain Connectivity Atlas online tools allow the user to search the projections between injected regions and display the labeled pathways as tracks in three dimensions in the Brain Explorer. The CSHL and UCLA connectomes can currently be viewed online as serial section data sets. The data from the Cre 518 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN recombinase driver mouse lines in the AIBS project provide a unique feature of cell-type specificity for the interpretation of the anterograde projections. The main strength of the CSHL and UCLA efforts lies in the multiplicity of the anatomical tracers used. The use of multiple retrograde tracers in particular will yield useful information, as retrogradely labeled soma (>10 micrometers in diameter) are easier to quantify than thin (<1 micrometer) axon fibers. These experiments will also provide an important comparison between the traditional CTB and FluoroGold tracers and the rabies virus tracer that is also being used in trans-synaptic labeling (see below) but is less well studied and may show some variation in transport affinity at different types of synapses. In summary, the LM-based mesoscopic mapping projects are set to transform the study of the circuit wiring of the mouse brain by providing online access to whole-brain data sets from several thousand injections of anterograde and retrograde tracers. The informatics tools being developed to search the databases will greatly aid in parsing the large amounts of data and in accessing specific brain samples for detailed scholarly analyses by the neuroscience community. Mapping connectivity using trans-synaptic tracers In contrast to electron microscopy methods, which provide a readout of neuronal connectivity with synapse resolution over small volumes of tissue, the whole-brain LM methods permit the assessment of projection-based connectivity between brain regions and in some cases between specific cell types in those regions but without the option of visualizing the underlying synaptic contacts. Trans-synaptic viruses that cross either multiple or single synapses can help to circumvent the requirement to confirm connectivity at resolution achieved by electron microscopy because such connectivity may be inferred from the known direction and mechanism of spread of the trans-synaptic tracer. Transsynaptic tracers based on rabies virus, pseudorabies virus and herpes simplex virus, which repeatedly cross synaptic connections in a retrograde or anterograde direction, are powerful tools for studying multistep pathways upstream and downstream from the starter cell population49–51. Furthermore, modified trans-synaptic rabies viruses have been developed that are restricted in their spread to a single synaptic jump and thus can be used to identify monosynaptic connections onto and downstream of specific neuronal populations and even individual cells27,52–58 (Fig. 3). Rabies virus spreads from the initially infected cells in a transsynaptic retrograde manner49,59. Rabies virus infection does not occur via spurious spread or uptake by fibers of passage and, because it cannot cross via electrical synapses, it is an effective tool for unidirectional anatomical tracing60. In a modified rabies virus system, the infection can also be cell type–targeted by encapsulating glycoprotein-deficient rabies virus with an avian virus envelope protein (referred to as ‘SAD-∆G-EnvA’). This restricts infection to only those cells that express an avian receptor protein TVA that is natively found in birds but not in mammals61,62. Thus, the delivery of vectors driving the expression of both TVA and rabies virus glycoprotein (RV-G) into a single cell28,54,56 (see below) or a specific population of cells55,63, ensures that only the targeted cell or cells will (i) be susceptible to initial infection and (ii) provide the replication-incompetent virus with RV-G required for trans-synaptic infection64. In this system, the virus can spread from the primarily infected cell or cells to the presynaptic FOCUS ON MAPPING THE BRAINreview npg © 2013 Nature America, Inc. All rights reserved. Figure 3 | Mapping the function and a b c Action potential response 2 Hz connectivity of single cells in the mouse brain Synaptic response 24 V × s in vivo. (a) Experimental setup for combined 0 315 45 10 mV single-cell physiology and trans-synaptic 100 ms connectivity mapping. Patch pipettes have 270 90 TVA RV-G Rabies virus solutions containing DNA vectors used to drive 135 225 –78 mV the expression of the TVA and RV-G proteins. Synaptic tuning Local synaptic Intrinsic properties (b) Patch pipettes are used to perform a connectivity 20 mV V2ML whole-cell recording of the intrinsic and V1 200 ms 330° sensory-evoked synaptic properties of a V2L 300° single layer-5 neuron in primary visual cortex. 240° Synaptic responses are averages of five traces. 210° DLG (c) Then the encapsulated modified rabies virus 150° DLG 120° is injected into the brain in close proximity 60° to the recorded neuron. After a period of up 30° to 12 days to ensure retrograde spread of the Retrograde labeling of local Synaptic properties and long-range inputs modified rabies from the recorded neuron, the brain is removed and imaged for identification of the local and long-range presynaptic inputs underlying the tuning of the recorded neuron to the direction of visual motion (polar plot). Fluorescence image on the right shows injection site with the starter cell (yellow) in the middle. Scheme (bottom left) and imaged data (bottom right) of long-range retrograde tracing. V1, primary visual cortex; V2L, secondary visual cortex (lateral); V2ML secondary visual cortex (medio-lateral). Inset shows long-range inputs from the dorsal lateral geniculate nucleus (DLG). Scale bars, 300 µm (top) and 50 µm (bottom), respectively. Images modified from ref. 54. input cells, which become labeled by expression of a fluorescent protein. However, as the presynaptic cells do not express RV-G28, the virus cannot spread further. This approach thus allows the discovery of the identity and location of the upstream input network relative to a defined population of neurons57,58. Brain-region specificity and cell-type specificity for mapping connectivity by the modified rabies virus system can be achieved by using a Cre recombinase–dependent helper virus driving expression of TVA and RV-G and transgenic driver mouse lines that express Cre recombinase in specific cell types or cortical layers38,39,63. This strategy is particularly useful for brain regions comprising many different cell types that could not be otherwise selectively targeted. Moreover, the engineering of other neurotropic trans-synaptic viruses is adding new tools for anatomical tracing, including Cre recombinase–dependent anterograde tracers based on a modified H129 strain of herpes simplex virus65 and vesicular stomatitis virus66, and retrograde tracers based on a modified pseudorabies virus (H. Oyibo and A. Zador, personal communication). The use of retrograde and anterograde trans-synaptic viruses, in combination with wholebrain LM methods, thus promises to afford unprecedented access to the upstream and downstream connectivity of specific cell types in the mouse brain. Present challenges and opportunities for whole-brain LM As highlighted above, LM instruments for whole-brain imaging are expected to make a substantial contribution in large-scale projects that focus on anatomical connectivity at the level of the whole mouse brain. It has also become clear that the use of these instruments will have an impact in many experimental applications in different neuroscience laboratories. It is therefore imperative that there exist broadly applicable image-processing, warping and analytical tools that will facilitate data sharing and across-laboratory collaboration and validation in future neuroscience studies focusing on, for example, mapping whole-brain anatomical changes during development and in response to experience. One practical problem arising from the choice to scan entire mouse brains at high resolution relates to the handling of large data sets (up to several terabytes per brain), which necessitates automated analytical pipelining. STP tomography is currently the most broadly used method among the whole-brain LM approaches, and there are freely available informatics tools for compiling STP tomography image stacks and viewing them as three-dimensional data, including algorithms that automate seamless stitching15. Another key challenge for charting the distribution of the labeled elements in the whole mouse brain is the process of accurate registration of the individual brain data sets onto an anatomical reference atlas. To this end, scientists at AIBS have generated the open-source segmented Allen Mouse Brain Atlas for the adult C57BL/6 mouse29,31,32,48, which is also available for registration of data sets generated by STP tomography (Figs. 2b and 4b). In addition, the so-called Waxholm space for standardized digital atlasing67 allows comparisons of registered mouse brain data using multiple brain atlases, including the Allen Mouse Brain Atlas, the digital Paxinos and Franklin Mouse Brain Atlas68, and several magnetic resonance imaging reference mouse brains. The continuing development of the Waxholm space and other online data-analysis platforms30,69,70 will be essential for standardized comparisons of mouse brain data collected in different laboratories using different instruments. The completion of the three mesoscopic connectome projects in the next several years will yield a comprehensive map of pointto-point connectivity between anatomical regions in the mouse brain7. Determining the cell-type identity of the neurons sending and receiving the connections in the brain regions will be essential for interpreting the function of the brain-wide neural circuits. Immunohistochemical analyses of labeled circuits have proven invaluable for ascertaining the identity of specific classes of neurons71–73 and synaptic connections52,74. The combination of immunohistochemical analysis by array tomography75,76 and anatomical tracing by the whole-brain LM instruments promises to be particularly powerful, as it will bring together two largely automated methodologies with complementary focus on synaptic and mesoscopic connectivity, respectively. STP tomography outputs sectioned tissue (typically 50-micrometer-thick sections15), which can be further resectioned, processed and reimaged by array nature methods | VOL.10 NO.6 | JUNE 2013 | 519 review npg © 2013 Nature America, Inc. All rights reserved. Figure 4 | Imaging induction of c-fos as a means to map whole-brain activation. (a) A 3D visualization of 367,378 c-fos–GFP cells detected in 280 coronal sections of an STP tomography data set of a mouse brain after the mouse was allowed to explore a novel object for 90 seconds. (b) Examples of anatomical segmentation of the brain volume with the Allen Mouse Brain Reference Atlas labels48 modified for the 280-section STP tomography data sets: hippocampus (blue), prelimbic cortex (aqua), infralimbic cortex (orange) and piriform cortex (green). (c) Visualization of c-fos–GFP cells in the hippocampus (38,170 cells), prelimbic cortex (3,305 cells), infralimbic cortex (3,827 cells) and piriform cortex (10,910 cells) (P.O., Y. Kim and K. Umadevi Venkataraju; unpublished data). FOCUS ON MAPPING THE BRAIN a tomography for integrating cell type–specific information into the whole-brain data sets. Industrial-level automation of slice capture and immunostaining can be developed to minimize manual handling and enhance the integration of immunohistochemistry and STP tomography. In addition, sectioning and immunostaining can also be applied to LSFM-imaged mouse brains20. A related, cell type–focused application of whole-brain LM imaging will be to quantitatively map the distribution (the cell counts) of different neuronal cell types in all anatomical regions in the mouse brain. Several such cell count–based anatomical studies have been done previously at smaller scales, revealing, for example, cell densities with respect to cortical vasculature77 or the density of neuronal cell types per layer in a single cortical column78–80. Using the whole-brain LM methods, a comprehensive anatomical atlas of different GABAergic inhibitory interneurons81 can now be generated by imaging cell type–specific Cre recombinase–mediated knock-in mouse lines38,39 crossed with Cre recombinase–dependent reporter mice expressing nuclear GFP. These and similar data sets for other neuronal cell types will complement the mesoscopic brain region connectivity data and help the interpretation of the immunohistochemistry data by providing a reference for total numbers of specific cell types per anatomical brain region. Integrating brain anatomy and function The anterograde, retrograde and trans-synaptic tracing approaches described above will yield the structural scaffold of anatomical projections and connections throughout the mouse brain. However, such data will not be sufficient to identify how specific brain regions connect to form functional circuits driving different behaviors. Bridging whole-brain structure and function is the next frontier in systems neuroscience, and the development of new technologies and methods will be crucial in achieving progress. The structure-function relationship of single neurons can be examined by in vivo intracellular delivery of the DNA vectors required for targeting and driving trans-synaptic virus expression via patch pipettes in loose cell-attached mode for electroporation56 or via whole-cell recording54 (Fig. 3). Used in combination with two-photon microscopy, this single-cell delivery technique may also be targeted at fluorescently labeled neurons of specific cell types56,82,83. The whole-cell method is particularly informative, as 520 | VOL.10 NO.6 | JUNE 2013 | nature methods b c its intracellular nature permits recording the intrinsic biophysical profile of the target cell, which, in turn, may reflect its functional connectivity status in the local network 84. In addition, by recording sensory-evoked inputs, it is possible to compare single-cell synaptic receptive fields and anatomical local and long-range connectivity traced by LM methods54. This combinatorial approach, involving single-cell electrophysiology and genetic manipulation designed for connection mapping, makes it possible to test long-standing theories regarding the extent to which emergent features of sensory cortical function manifest via specific wiring motifs85. As has recently been achieved for serial electron microscopy– based reconstruction86,87, it will also be valuable to functionally characterize larger local neuronal populations for registration against LM-based connectivity data. In this sense, genetically encoded calcium indicators, which permit physiological characterization of neuronal activity in specific cell types88–90, along with viral vectors for trans-synaptic labeling and LM-based tracing, will have critical complementary roles. Large-volume in vivo twophoton imaging of neuronal activity before ex vivo whole-brain imaging will establish the extent to which connectivity patterns relate to function91 at the level of single cells, and local and longrange circuits. Interpolation of such experiments will rely on the ability to cross-register in vivo functional imaging with complete ex vivo LM connectivity data. Preliminary experiments, which already hint at the spatial spread of monosynaptic connectivity of individual principal cortical cells, suggest that combination of functional imaging and traditional anatomical-circuit reconstruction may only be feasible at the local network level where connection probability is the highest92–94. Given the broad, sparse expanse of connectivity in most brain regions and especially in cortical areas, high-throughput whole-brain LM methods will be imperative for complete anatomical-circuit reconstruction of functionally characterized local networks. The amalgamation of whole-brain LM and physiological methods for single neurons and small networks offers a powerful means to study the mouse brain. An exciting application of this approach will be to trace the synaptic circuits of neurons functionally characterized in head-fixed behaving animals engaged in tasks related to spatial navigation, sensorimotor integration and other complex brain functions95–97. This research will lead to the generation of whole-brain structure-function hypotheses for specific npg © 2013 Nature America, Inc. All rights reserved. FOCUS ON MAPPING THE BRAINreview behaviors, which can then be tested for causality by optogenetic methods targeted to the identified cell types and brain regions98. Furthermore, LM, physiological and optogenetic methods can be applied to interrogate entire brain systems in large-scale projects, as is currently being done for the mouse visual cortex in an effort led by C. Koch and R.C. Reid at AIBS99. Finally, a discussion in the neuroscience community has been initiated regarding the feasibility of mapping activity at cellular resolution in whole brains and linking the identified activity patterns to brain anatomy100. Today, such experiments are possible in small, transparent organisms, as was demonstrated by twophoton microscopy and LSFM-based imaging of brain activity in larval zebrafish expressing the calcium indicator GCaMP89,101,102. Understandably, LM-based approaches will not be useful for in vivo whole-brain imaging in larger, nontransparent animals, and the invention of new disruptive technologies will likely be needed to achieve the goal of mapping real-time brain activity at cellular resolution in, for example, the mouse. In contrast, LM methods can be used to map patterns of whole-brain activation indirectly, by post-hoc visualization of activity-induced expression of immediate early genes, such as mouse Fos (c-fos), Arc or Homer1a (ref. 103). Transgenic fluorescent immediate early gene reporter mice, such as c-fos–GFP or Arc-GFP mice104–106, can be trained in a specific behavior, their brains subsequently can be imaged ex vivo, and the exact distribution of GFP-positive neurons can be mapped and analyzed by computational methods (Fig. 4). In this approach, a statistical analysis of the counts of GFP-labeled neurons can be used to identify brain regions and cell types activated during behaviors but without providing information on the temporal sequence of brain region activation or the firing patterns of the activated cells. However, the development of more sensitive, for instance, fluorescent RNA–based methods, may allow calibration of the cellular signal with respect to the temporal window and the pattern of activity related to the induction of immediate early genes. Such calibration would considerably enhance the power of LM-based whole-brain mapping of the induction of immediate early genes, which, in combination with the connectomics data, could then be used to begin to build cellularresolution models of function-based whole-brain circuits. Conclusions The advances in automated LM methods, anatomical tracers, physiological methods and informatics tools have begun to transform our understanding of the circuit wiring in the mouse brain. The focus on the mouse as an animal model is, of course, not accidental. In addition to the generation of cell type–specific driver mouse lines38–40 that allow the study of specific neuronal populations in the normal brain, mouse genetics are used in hundreds of laboratories to model gene mutations linked to heritable human disorders, including complex cognitive disorders such as autism and schizophrenia. Without a doubt, understanding the relationships between brain structure and function in the genetic mouse models will be crucial to understanding the underlying brain circuit mechanisms of these disorders. The toolbox of LM methods described here, and the continuing development of new methods, promise to transform the study of brain circuits in animal models and to decipher the structure-function relationships essential to the understanding of complex brain functions and their deficits in human brain disorders. Acknowledgments We thank P. Mitra, H. Zeng and Christian Niedworok for comments on the manuscript and J. Kuhl for the graphics. P.O. is supported by the US National Institute of Mental Health grant 1R01MH096946-01, McKnight Foundation, Technological Innovations in Neuroscience Award and Simons Foundation for Autism Research grants 204719 and 253447. T.W.M. is supported as a Wellcome Trust Investigator and by the Medical Research Council MC U1175975156. COMPETING FINANCIAL INTERESTS The authors declare competing financial interests: details are available in the online version of the paper. Reprints and permissions information is available online at http://www.nature. com/reprints/index.html. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. Golgi, C. Sulla Struttura Della Sostanza Grigia del Cervello. Gazz. Med. Ital. (Lombardia) 33, 244–246 (1873). Ramón y Cajal, S. Textura del Sistema Nervioso del Hombre y de los Vertebrados. Vol. 2 (Moya, 1904). Felleman, D.J. & Van Essen, D.C. Distributed hierarchical processing in the primate cerebral cortex. Cereb. Cortex 1, 1–47 (1991). Rockland, K.S. & Pandya, D.N. Laminar origins and terminations of cortical connections of the occipital lobe in the rhesus monkey. Brain Res. 179, 3–20 (1979). Craddock, R.C. et al. Imaging human connectomes at the macroscale. Nat. Methods 10, 524–539 (2013). Helmstaedter, M. Cellular-resolution connectomics: challenges of dense neural circuit reconstruction. Nat. Methods 10, 501–507 (2013). Bohland, J.W. et al. A proposal for a coordinated effort for the determination of brainwide neuroanatomical connectivity in model organisms at a mesoscopic scale. PLoS Comput. Biol. 5, e1000334 (2009). This paper describes the rationale for mapping connectivity in the whole mouse brain at the mesoscale level by LM. Odgaard, A., Andersen, K., Melsen, F. & Gundersen, H.J. A direct method for fast three-dimensional serial reconstruction. J. Microsc. 159, 335–342 (1990). Ewald, A.J., McBride, H., Reddington, M., Fraser, S.E. & Kerschmann, R. Surface imaging microscopy, an automated method for visualizing whole embryo samples in three dimensions at high resolution. Dev. Dyn. 225, 369–375 (2002). Tsai, P.S. et al. All-optical histology using ultrashort laser pulses. Neuron 39, 27–41 (2003). This study pioneered the approach of serial imaging by two-photon microscopy and tissue sectioning for ex vivo collection of neuroanatomical data. Sands, G.B. et al. Automated imaging of extended tissue volumes using confocal microscopy. Microsc. Res. Tech. 67, 227–239 (2005). Ragan, T. et al. High-resolution whole organ imaging using two-photon tissue cytometry. J. Biomed. Opt. 12, 014015 (2007). Mayerich, D., Abbott, L. & McCormick, B. Knife-edge scanning microscopy for imaging and reconstruction of three-dimensional anatomical structures of the mouse brain. J. Microsc. 231, 134–143 (2008). Li, A. et al. Micro-optical sectioning tomography to obtain a highresolution atlas of the mouse brain. Science 330, 1404–1408 (2010). Ragan, T. et al. Serial two-photon tomography for automated ex vivo mouse brain imaging. Nat. Methods 9, 255–258 (2012). This study introduces the method of STP tomography and demonstrates its use for anterograde and retrograde tracing in the mouse brain. Gong, H. et al. Continuously tracing brain-wide long-distance axonal projections in mice at a one-micron voxel resolution. Neuroimage 74, 87–98, (2013). This study demonstrates the first long-range tracing of individual axons in the mouse brain by fMOST. Denk, W., Strickler, J.H. & Webb, W.W. Two-photon laser scanning fluorescence microscopy. Science 248, 73–76 (1990). Huisken, J., Swoger, J., Del Bene, F., Wittbrodt, J. & Stelzer, E.H. Optical sectioning deep inside live embryos by selective plane illumination microscopy. Science 305, 1007–1009 (2004). Dodt, H.U. et al. Ultramicroscopy: three-dimensional visualization of neuronal networks in the whole mouse brain. Nat. Methods 4, 331–336 (2007). This study is the first to use LSFM for imaging the entire mouse brain. nature methods | VOL.10 NO.6 | JUNE 2013 | 521 npg © 2013 Nature America, Inc. All rights reserved. review 20. Niedworok, C.J. et al. Charting monosynaptic connectivity maps by twocolor light-sheet fluorescence microscopy. Cell Rep. 2, 1375–1386 (2012). 21. Chung, K. et al. Structural and molecular interrogation of intact biological systems. Nature doi:10.1038/nature12107 (2013). 22. Chung, K. & Deisseroth, K. CLARITY for mapping the nervous system. Nat. Methods 10, 508–513 (2013). 23. Leischner, U., Zieglgansberger, W. & Dodt, H.U. Resolution of ultramicroscopy and field of view analysis. PLoS ONE 4, e5785 (2009). 24. Mertz, J. & Kim, J. Scanning light-sheet microscopy in the whole mouse brain with HiLo background rejection. J. Biomed. Opt. 15, 016027 (2010). 25. Kalchmair, S., Jahrling, N., Becker, K. & Dodt, H.U. Image contrast enhancement in confocal ultramicroscopy. Opt. Lett. 35, 79–81 (2010). 26. Keller, P.J. et al. Fast, high-contrast imaging of animal development with scanned light sheet-based structured-illumination microscopy. Nat. Methods 7, 637–642 (2010). 27. Wickersham, I.R., Finke, S., Conzelmann, K.K. & Callaway, E.M. Retrograde neuronal tracing with a deletion-mutant rabies virus. Nat. Methods 4, 47–49 (2007). 28. Wickersham, I.R. et al. Monosynaptic restriction of transsynaptic tracing from single, genetically targeted neurons. Neuron 53, 639–647 (2007). This study describes a genetically modified rabies virus designed to specifically label direct presynaptic input onto a given cell population. 29. Ng, L. et al. An anatomic gene expression atlas of the adult mouse brain. Nat. Neurosci. 12, 356–362 (2009). 30. Jones, E.G., Stone, J.M. & Karten, H.J. High-resolution digital brain atlases: a Hubble telescope for the brain. Ann. NY Acad. Sci. 1225 (suppl. 1), E147–E159 (2011). 31. Lein, E.S. et al. Genome-wide atlas of gene expression in the adult mouse brain. Nature 445, 168–176 (2007). This study pioneered large-scale LM-based mouse brain anatomy and introduced the Allen Mouse Brain Atlas and online data portal. 32. Dong, H.W. The Allen Reference Atlas: A Digital Color Brain Atlas of the C57Bl/6J Male Mouse (John Wiley & Sons Inc., 2008). 33. Lanciego, J.L. & Wouterlood, F.G. A half century of experimental neuroanatomical tracing. J. Chem. Neuroanat. 42, 157–183 (2011). 34. Glover, J.C., Petursdottir, G. & Jansen, J.K. Fluorescent dextran-amines used as axonal tracers in the nervous system of the chicken embryo. J. Neurosci. Methods 18, 243–254 (1986). 35. Llewellyn-Smith, I.J., Martin, C.L., Arnolda, L.F. & Minson, J.B. Tracertoxins: cholera toxin B-saporin as a model. J. Neurosci. Methods 103, 83–90 (2000). 36. Grinevich, V., Brecht, M. & Osten, P. Monosynaptic pathway from rat vibrissa motor cortex to facial motor neurons revealed by lentivirus-based axonal tracing. J. Neurosci. 25, 8250–8258 (2005). 37. Atasoy, D., Aponte, Y., Su, H.H. & Sternson, S.M. A FLEX switch targets Channelrhodopsin-2 to multiple cell types for imaging and long-range circuit mapping. J. Neurosci. 28, 7025–7030 (2008). 38. Taniguchi, H. et al. A resource of Cre driver lines for genetic targeting of GABAergic neurons in cerebral cortex. Neuron 71, 995–1013 (2011). 39. Madisen, L. et al. A robust and high-throughput Cre reporting and characterization system for the whole mouse brain. Nat. Neurosci. 13, 133–140 (2010). 40. Madisen, L. et al. A toolbox of Cre-dependent optogenetic transgenic mice for light-induced activation and silencing. Nat. Neurosci. 15, 793–802 (2012). 41. Harris, J.A., Wook Oh, S. & Zeng, H. Adeno-associated viral vectors for anterograde axonal tracing with fluorescent proteins in nontransgenic and cre driver mice. Curr. Protoc. Neurosci. 1.20, 21–18 (2012). 42. Thompson, R.H. & Swanson, L.W. Hypothesis-driven structural connectivity analysis supports network over hierarchical model of brain architecture. Proc. Natl. Acad. Sci. USA 107, 15235–15239 (2010). 43. Gerfen, C.R. & Sawchenko, P.E. An anterograde neuroanatomical tracing method that shows the detailed morphology of neurons, their axons and terminals: immunohistochemical localization of an axonally transported plant lectin, Phaseolus vulgaris leucoagglutinin (PHA-L). Brain Res. 290, 219–238 (1984). 44. Naumann, T., Hartig, W. & Frotscher, M. Retrograde tracing with FluoroGold: different methods of tracer detection at the ultrastructural level and neurodegenerative changes of back-filled neurons in long-term studies. J. Neurosci. Methods 103, 11–21 (2000). 45. Reiner, A. et al. Pathway tracing using biotinylated dextran amines. J. Neurosci. Methods 103, 23–37 (2000). 522 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN 46. Hintiryan, H. et al. Comprehensive connectivity of the mouse main olfactory bulb: analysis and online digital atlas. Front. Neuroanat. 6, 30 (2012). 47. Conte, W.L., Kamishina, H. & Reep, R.L. Multiple neuroanatomical tracttracing using fluorescent Alexa Fluor conjugates of cholera toxin subunit B in rats. Nat. Protoc. 4, 1157–1166 (2009). 48. Sunkin, S.M. et al. Allen Brain Atlas: an integrated spatio-temporal portal for exploring the central nervous system. Nucleic Acids Res. 41, D996–D1008 (2013). 49. Ugolini, G. Advances in viral transneuronal tracing. J. Neurosci. Methods 194, 2–20 (2010). 50. Ekstrand, M.I., Enquist, L.W. & Pomeranz, L.E. The alpha-herpesviruses: molecular pathfinders in nervous system circuits. Trends Mol. Med. 14, 134–140 (2008). 51. Song, C.K., Enquist, L.W. & Bartness, T.J. New developments in tracing neural circuits with herpesviruses. Virus Res. 111, 235–249 (2005). 52. Callaway, E.M. Transneuronal circuit tracing with neurotropic viruses. Curr. Opin. Neurobiol. 18, 617–623 (2008). 53. Wickersham, I.R. & Feinberg, E.H. New technologies for imaging synaptic partners. Curr. Opin. Neurobiol. 22, 121–127 (2012). 54. Rancz, E.A. et al. Transfection via whole-cell recording in vivo: bridging single-cell physiology, genetics and connectomics. Nat. Neurosci. 14, 527–532 (2011). This study was the first to combine intracellular neuronal recording with DNA delivery. The authors use this method to map the synaptic function of a single cell in vivo and then target rabies-based retrograde labeling of the cells’ synaptic input. 55. Miyamichi, K. et al. Cortical representations of olfactory input by transsynaptic tracing. Nature 472, 191–196 (2011). 56. Marshel, J.H., Mori, T., Nielsen, K.J. & Callaway, E.M. Targeting single neuronal networks for gene expression and cell labeling in vivo. Neuron 67, 562–574 (2010). This paper describes an electroporation method for single-cell delivery of DNA for targeted infection of modified rabies virus. 57. Takatoh, J. et al. New modules are added to vibrissal premotor circuitry with the emergence of exploratory whisking. Neuron 77, 346–360 (2013). 58. Arenkiel, B.R. et al. Activity-induced remodeling of olfactory bulb microcircuits revealed by monosynaptic tracing. PLoS ONE 6, e29423 (2011). 59. Finke, S. & Conzelmann, K.K. Replication strategies of rabies virus. Virus Res. 111, 120–131 (2005). 60. Ugolini, G. Specificity of rabies virus as a transneuronal tracer of motor networks: transfer from hypoglossal motoneurons to connected secondorder and higher order central nervous system cell groups. J. Comp. Neurol. 356, 457–480 (1995). 61. Federspiel, M.J., Bates, P., Young, J.A., Varmus, H.E. & Hughes, S.H. A system for tissue-specific gene targeting: transgenic mice susceptible to subgroup A avian leukosis virus-based retroviral vectors. Proc. Natl. Acad. Sci. USA 91, 11241–11245 (1994). 62. Young, J.A., Bates, P. & Varmus, H.E. Isolation of a chicken gene that confers susceptibility to infection by subgroup A avian leukosis and sarcoma viruses. J. Virol. 67, 1811–1816 (1993). 63. Wall, N.R., Wickersham, I.R., Cetin, A., De La Parra, M. & Callaway, E.M. Monosynaptic circuit tracing in vivo through Cre-dependent targeting and complementation of modified rabies virus. Proc. Natl. Acad. Sci. USA 107, 21848–21853 (2010). 64. Etessami, R. et al. Spread and pathogenic characteristics of a G-deficient rabies virus recombinant: an in vitro and in vivo study. J. Gen. Virol. 81, 2147–2153 (2000). 65. Lo, L. & Anderson, D.J.A. Cre-dependent, anterograde transsynaptic viral tracer for mapping output pathways of genetically marked neurons. Neuron 72, 938–950 (2011). 66. Beier, K.T. et al. Anterograde or retrograde transsynaptic labeling of CNS neurons with vesicular stomatitis virus vectors. Proc. Natl. Acad. Sci. USA 108, 15414–15419 (2011). 67. Hawrylycz, M. et al. Digital atlasing and standardization in the mouse brain. PLoS Comput. Biol. 7, e1001065 (2011). 68. Paxinos, G. & Franklin, K.B. The Mouse Brain in Stereotaxic Coordinates (Gulf Professional Publishing, 2004). 69. Swanson, L.W. & Bota, M. Foundational model of structural connectivity in the nervous system with a schema for wiring diagrams, connectome, and basic plan architecture. Proc. Natl. Acad. Sci. USA 107, 20610–20617 (2010). npg © 2013 Nature America, Inc. All rights reserved. FOCUS ON MAPPING THE BRAINreview 70. Moene, I.A., Subramaniam, S., Darin, D., Leergaard, T.B. & Bjaalie, J.G. Toward a workbench for rodent brain image data: systems architecture and design. Neuroinformatics 5, 35–58 (2007). 71. Klausberger, T. & Somogyi, P. Neuronal diversity and temporal dynamics: the unity of hippocampal circuit operations. Science 321, 53–57 (2008). 72. O’Rourke, N.A., Weiler, N.C., Micheva, K.D. & Smith, S.J. Deep molecular diversity of mammalian synapses: why it matters and how to measure it. Nat. Rev. Neurosci. 13, 365–379 (2012). 73. DeFelipe, J. et al. New insights into the classification and nomenclature of cortical GABAergic interneurons. Nat. Rev. Neurosci. 14, 202–216 (2013). 74. Emes, R.D. & Grant, S.G. Evolution of synapse complexity and diversity. Annu. Rev. Neurosci. 35, 111–131 (2012). 75. Micheva, K.D., Busse, B., Weiler, N.C., O’Rourke, N. & Smith, S.J. Singlesynapse analysis of a diverse synapse population: proteomic imaging methods and markers. Neuron 68, 639–653 (2010). 76. Micheva, K.D. & Smith, S.J. Array tomography: a new tool for imaging the molecular architecture and ultrastructure of neural circuits. Neuron 55, 25–36 (2007). 77. Tsai, P.S. et al. Correlations of neuronal and microvascular densities in murine cortex revealed by direct counting and colocalization of nuclei and vessels. J. Neurosci. 29, 14553–14570 (2009). 78. Oberlaender, M. et al. Cell type-specific three-dimensional structure of thalamocortical circuits in a column of rat vibrissal cortex. Cereb. Cortex 22, 2375–2391 (2012). 79. Meyer, H.S. et al. Inhibitory interneurons in a cortical column form hot zones of inhibition in layers 2 and 5A. Proc. Natl. Acad. Sci. USA 108, 16807–16812 (2011). 80. Meyer, H.S. et al. Number and laminar distribution of neurons in a thalamocortical projection column of rat vibrissal cortex. Cereb. Cortex 20, 2277–2286 (2010). 81. Ascoli, G.A. et al. Petilla terminology: nomenclature of features of GABAergic interneurons of the cerebral cortex. Nat. Rev. Neurosci. 9, 557–568 (2008). 82. Margrie, T.W. et al. Targeted whole-cell recordings in the mammalian brain in vivo. Neuron 39, 911–918 (2003). 83. Kitamura, K., Judkewitz, B., Kano, M., Denk, W. & Hausser, M. Targeted patch-clamp recordings and single-cell electroporation of unlabeled neurons in vivo. Nat. Methods 5, 61–67 (2008). 84. Angelo, K. et al. A biophysical signature of network affiliation and sensory processing in mitral cells. Nature 488, 375–378 (2012). 85. Reid, R.C. From functional architecture to functional connectomics. Neuron 75, 209–217 (2012). 86. Briggman, K.L., Helmstaedter, M. & Denk, W. Wiring specificity in the direction-selectivity circuit of the retina. Nature 471, 183–188 (2011). 87. Bock, D.D. et al. Network anatomy and in vivo physiology of visual cortical neurons. Nature 471, 177–182 (2011). 88. Wallace, D.J. et al. Single-spike detection in vitro and in vivo with a genetic Ca2+ sensor. Nat. Methods 5, 797–804 (2008). 89. Akerboom, J. et al. Optimization of a GCaMP calcium indicator for neural activity imaging. J. Neurosci. 32, 13819–13840 (2012). 90. Mank, M. et al. A genetically encoded calcium indicator for chronic in vivo two-photon imaging. Nat. Methods 5, 805–811 (2008). 91. Ko, H. et al. Functional specificity of local synaptic connections in neocortical networks. Nature 473, 87–91 (2011). 92. Holmgren, C., Harkany, T., Svennenfors, B. & Zilberter, Y. Pyramidal cell communication within local networks in layer 2/3 of rat neocortex. J. Physiol. 551, 139–153 (2003). 93. Thomson, A.M., West, D.C., Wang, Y. & Bannister, A.P. Synaptic connections and small circuits involving excitatory and inhibitory neurons in layers 2–5 of adult rat and cat neocortex: triple intracellular recordings and biocytin labelling in vitro. Cereb. Cortex 12, 936–953 (2002). 94. Song, S., Sjostrom, P.J., Reigl, M., Nelson, S. & Chklovskii, D.B. Highly nonrandom features of synaptic connectivity in local cortical circuits. PLoS Biol. 3, e68 (2005). 95. Harvey, C.D., Collman, F., Dombeck, D.A. & Tank, D.W. Intracellular dynamics of hippocampal place cells during virtual navigation. Nature 461, 941–946 (2009). This study introduced the method of physiological recording in headrestrained mice on a spherical treadmill while the mice performed spatial tasks in a virtual environment. 96. Harvey, C.D., Coen, P. & Tank, D.W. Choice-specific sequences in parietal cortex during a virtual-navigation decision task. Nature 484, 62–68 (2012). 97. Huber, D. et al. Multiple dynamic representations in the motor cortex during sensorimotor learning. Nature 484, 473–478 (2012). 98. Fenno, L., Yizhar, O. & Deisseroth, K. The development and application of optogenetics. Annu. Rev. Neurosci. 34, 389–412 (2011). 99. Koch, C. & Reid, R.C. Neuroscience: observatories of the mind. Nature 483, 397–398 (2012). 100. Alivisatos, A.P. et al. Neuroscience. The brain activity map. Science 339, 1284–1285 (2013). 101. Ahrens, M.B., Orger, M.B., Robson, D.N., Li, J.M. & Keller, P.J. Whole-brain functional imaging at cellular resolution using light-sheet microscopy. Nat. Methods 10, 413–420 (2013). 102. Ahrens, M.B. et al. Brain-wide neuronal dynamics during motor adaptation in zebrafish. Nature 485, 471–477 (2012). 103. Herrera, D.G. & Robertson, H.A. Activation of c-fos in the brain. Prog. Neurobiol. 50, 83–107 (1996). 104. Barth, A.L., Gerkin, R.C. & Dean, K.L. Alteration of neuronal firing properties after in vivo experience in a FosGFP transgenic mouse. J. Neurosci. 24, 6466–6475 (2004). 105. Grinevich, V. et al. Fluorescent Arc/Arg3.1 indicator mice: a versatile tool to study brain activity changes in vitro and in vivo. J. Neurosci. Methods 184, 25–36 (2009). 106. Reijmers, L.G., Perkins, B.L., Matsuo, N. & Mayford, M. Localization of a stable neural correlate of associative memory. Science 317, 1230–1233 (2007). nature methods | VOL.10 NO.6 | JUNE 2013 | 523 review Focus on mapping the brain Imaging human connectomes at the macroscale © 2013 Nature America, Inc. All rights reserved. R Cameron Craddock1,2, Saad Jbabdi3,9, Chao-Gan Yan1,2,4,9, Joshua T Vogelstein1,5–7, F Xavier Castellanos2,4, Adriana Di Martino4, Clare Kelly4, Keith Heberlein8, Stan Colcombe2 & Michael P Milham1,2 At macroscopic scales, the human connectome comprises anatomically distinct brain areas, the structural pathways connecting them and their functional interactions. Annotation of phenotypic associations with variation in the connectome and cataloging of neurophenotypes promise to transform our understanding of the human brain. In this Review, we provide a survey of magnetic resonance imaging–based measurements of functional and structural connectivity. We highlight emerging areas of development and inquiry and emphasize the importance of integrating structural and functional perspectives on brain architecture. First introduced in 2005 (ref. 1), the term ‘connectome’ embodies the advances of over a century of neuroscientific innovation and reflects an agenda for a new era. Initially defined as a complete map of neural connections in the brain, the connectome is a multiscale construct that can be examined at varying resolutions. At the extremes are the microscale, which encompasses individual neurons and their synaptic connections, and the macroscale, which encompasses cortical tissues (commonly a cubic centimeter or larger). The intermediate resolution is the mesoscale, which, in humans, encompasses vertical columns of 80–120 neurons (commonly referred to as micro- or mini-columns)1,2. Although all scales of resolution are intimately associated, each provides unique perspectives on the connectome. At the macroscale, and particularly in studies of the human brain, conceptualizations of the connectome have grown to also include information about function3. In this Review, we will use the term connectome to refer to brain areas, their anatomical connections and their functional interactions. At present, methodologies for analysis at macro scale resolution are best positioned for mapping and annotating human connectomes with cognitive and behavioral associations. The higher-order, albeit lowerresolution, representations captured at the macroscale most directly relate to regulatory, cognitive and affective processes. Interpretation of macroscale-resolution findings is most amenable to guidance from lesion and brain-imaging studies. Comprehensive mapping and annotation of the connectome is most feasible at the lower-resolution macroscale, owing to lower computational and analytical demands. Moreover, noninvasive tools for in vivo imaging the human connectome are only available for analyses at the macroscale; in vivo microscale-resolution studies are currently limited to model organisms and neurosurgical patients. Among the modalities used for macroconnectomics, magnetic resonance imaging (MRI) is dominant, partly because of widespread availability, safety and spatial resolution. Diffusion-weighted MRI (dMRI) and functional MRI (fMRI) are widely used for inferring structural and functional connectivity, respectively4. dMRI provides cubic-millimeter-resolution portrayals of white-matter tracts and insights into organizing principles that guide their orientation and 1Center for the Developing Brain, Child Mind Institute, New York, New York, USA. 2Nathan Kline Institute for Psychiatric Research, Orangeburg, New York, USA. 3The Oxford Centre for Functional MRI of the Brain, University of Oxford, Oxford, UK. 4The Phyllis Green and Randolph Cowen Institute for Pediatric Neuroscience, Child Study Center, New York University Langone Medical Center, New York, New York, USA. 5Department of Statistical Science, Duke University, Durham, North Carolina, USA. 6Institute for Brain Sciences, Duke University, Durham, North Carolina, USA. 7Institute for Data Intensive Engineering and Sciences, John Hopkins University, Baltimore, Maryland, USA. 8Siemens Medical Solutions USA, Charlestown, Massachusetts, USA. 9These authors contributed equally to this work. Correspondence should be addressed to S.C. ([email protected]) or M.P.M. ([email protected]). RECEIVED 22 FEBRUARY; ACCEPTED 22 APRIL; published online 30 may 2013; doi:10.1038/nmeth.2482 524 | VOL.10 NO.6 | JUNE 2013 | nature methods Defining nodes Defining the nodes of a macroscale connectome is a complex task as we lack agreement on how best to define the constituent brain units. Depending on the scope of the investigation, the specific brain subunits represented by nodes can range from the small patches of cortex contained in individual MRI voxels to larger brain areas (for example, dorsolateral prefrontal cortex). Early parcellation efforts used postmortem architectonic measurements (for example, cell morphology) of single individuals. Although these resulting atlases are central to neuroscience, no functional or structural connectivity–based information was used to construct them, thus limiting their capacity to accurately represent connectomes. For example, despite being represented as a single area in anatomical atlases 6, the anterior cingulate region contains subregions that are each characterized by dramatically different functional7 and structural8 connectivity patterns. Although the large-scale human brain patterns captured using different strategies of parceling data may bear a gross similarity to one another, the specific details conveyed vary substantially (Fig. 1). Ideally, both brain-function and structural-connectivity information should be used to delineate brain areas. Metaanalytic approaches can be used to define nodes on the basis of task-based fMRI (T-fMRI) studies9. Alternatively, data-driven clustering techniques can be used to subdivide the brain into areas based on homogeneity of functional time series10,11, or functional or structural connectivity profiles8,11,12. Blinded source-separation techniques can also be used to define network nodes using spatial independence13. These methods commonly involve pooling of information across individuals, and most enforce specific properties on resulting brain areas11. One drawback is the need to prespecify the number of areas to be generated, which can be estimated based on homogeneity, accuracy, reproducibility or stability of the brain areas 10,11. EZ HO TT CC200 trajectories; fMRI reveals a universal functional architecture, with variations among individuals meaningfully related to phenotypic variables5 (for example, behavioral and psychiatric). In this Review, we focus on the mapping, characterization and analysis of macroscale connectomes. We structured our presentation in terms of a mathematical perspective that treats the connectome as a graph of interactions among brain areas. Nodes in the graph are abstract representations of brain areas, and edges represent pairwise relationships between nodes. We first review approaches and challenges to subdividing the brain into discrete subunits represented by nodes (here referred to as ‘parcellation’ efforts) and then review the imaging and analytic methodologies used to map and quantify patterns of structural and functional connectivity that are represented by edges in the connectome. CC400 © 2013 Nature America, Inc. All rights reserved. Figure 1 | Different parcellations of the human brain. Atlases of brain areas generated using anatomical (top four rows) and functional (bottom two rows) parcellation schemes show a lateral view (right) and top views of the human brain. AAL (automated anatomical labeling)109 and Harvard Oxford (HO)110 are derived from anatomical landmarks (sulci and gyral). The EZ (Eickhoff–Zilles)111 and TT (Talariach Daemon)112 atlases are derived from postmortem cyto- and myelo-architectonic segmentations. The CC200 and CC400 atlases are derived from 200- and 400-unit functional parcellations11. review AAL Focus on mapping the brain Optimal comparison of connectomes is likely to require parcellation strategies that incorporate information across individuals and modalities, potentially at the cost of quality of fit for single subjects and modalities. Estimating structural connectivity Structural connectivity encompasses the collection of axonal and dendritic connections among neurons 1. Despite definitional simplicity, structural connectivity is difficult to measure with noninvasive, in vivo imaging approaches. Before dMRI, our knowledge was primarily derived from lesions and blunt dissection in humans or invasive tracing in nonhumans. These methodologies remain the gold standard for establishing connectivity, but the noninvasive nature of dMRI makes it the de facto standard for studies of human structural connectivity. Acquisition. The basic principle underlying the inference of structural connectivity from dMRI data is that water diffusion in white matter is hindered and occurs primarily along the path of axons. In contrast, water diffusion in gray matter and cerebral spinal fluid occurs (almost) equally in all directions. By following the motion of water, it is possible to map the orientation(s) of fibers passing through each voxel of white matter. In dMRI, a series of images are acquired, each sensitive to diffusion along a specific direction. The number of images for unique directions acquired varies from six to hundreds. When combined, these images contain the information necessary to estimate the orientation(s) of fibers passing through each voxel; this information is used to reconstruct large-scale tracts of white matter (tractography). We describe dMRI acquisition in more detail in Box 1 and Table 1. nature methods | VOL.10 NO.6 | JUNE 2013 | 525 review Focus on mapping the brain Box 1 Optimizing fMRI and dMRI image acquisition Echo-planar imaging (EPI) is the most common method for acquiring functional magnetic resonance imaging (fMRI) and diffusion MRI (dMRI) data. An EPI volume is a set of slices, each acquired after a single excitation. EPI volumes are collected in sequence, representing time points in fMRI data and diffusion directions in dMRI data. Gradient echo EPI in fMRI only involves excitation and readout. However, in dMRI an additional 180° rephasing (spin echo) radiofrequency pulse and bipolar diffusion–encoding gradients are applied. © 2013 Nature America, Inc. All rights reserved. SNR ∝ I0 × ∆ x × ∆ y × ∆ z × N acq × N x × N y × N z × 1 Pf (1) Signal-to-noise ratio (SNR), spatial distortions and artifacts, and image contrast determine image quality. Image SNR can be represented in terms of parameters such as available signal (I0), voxel dimensions (∆x, ∆y, ∆z), number of image acquisitions (Nacq), number of samples in each dimension (Nx, Ny, Nz), time between samples (∆t) and parallel imaging factor (Pf) (equation (1))117. These parameters are described in Table 1. Equation (1) demonstrates how different parameter settings affect the SNR; for example, acquiring and averaging two volumes (Nacq = 2) improves SNR by 2 . SNR for modern imaging technology is greater than required for most imaging applications and can be traded-off to optimize other imaging aspects (for example, to reduce spatial distortions). Beyond spatial image considerations, parameter optimizations must be evaluated in terms of impact on the temporal signal in fMRI or diffusion signal in dMRI. General considerations for EPI Spatial distortion. The EPI readout scheme leads to extended readout times and asymmetric bandwidth between directions that result in prolonged spin-spin (T2) relaxation, which increases in-plane smoothing. Bandwidth corresponds to the number of frequencies allocated to a voxel; any shift in frequency resulting from magnetic field imperfections spatially shifts the signal in proportion to the inverse of the bandwidth118. The bandwidth in the phase-encoding direction is a fraction of the bandwidth in the readout direction, making it particularly susceptible to spatial distortions. Increasing bandwidth can minimize spatial distortions, though costing SNR. Parallel imaging techniques increase effective bandwidth and image acquisition rate119. These techniques require calibration scans for proper reconstruction, and any misalignment (for example, head motion) between these scans and image data produce reconstruction errors119. Alternatively, spatial distortions can be mathematically corrected using maps of the spatial variation in the magnetic field15. Image resolution. Image resolution can be increased by reducing the field of view (FOV) or increasing the number of samples along a dimension. Decreasing the FOV produces a proportional decrease in the SNR; the FOV must remain larger than the head to avoid excessive ghosting. Increasing the number of samples has less impact on SNR (reduced by square root of the increase117). However, increasing the number of 526 | VOL.10 NO.6 | JUNE 2013 | nature methods samples in the phase encoding direction substantially increases readout time, which increases the amount of T2 decay experienced during slice acquisition and leads to greater in-plane smoothness118; parallel imaging mitigates these effects119. Considerations for fMRI BOLD contrast. Spin phase (T2*) contrast is the most sensitive to the BOLD effect and is optimized using an echo time equal to the T2* of gray matter (echo time of ~30 milliseconds at 3 tesla). BOLD can also be measured with spin-echo sequences, which are sensitive to T2 rather than T2* and hence have better spatial specificity, although less sensitivity than gradient echo techniques120. Minimizing dropout. Tilting and positioning imaging slices to prevent slices from intersecting air-tissue interfaces can minimize signal dropout resulting from differences in magnetic susceptibility121. Slice positioning should focus on areas most vital to the experiment. Reducing echo time, which decreases BOLD contrast, and increasing resolution, which decreases SNR, should be avoided. Specialty sequences (for example, spiral-in/out122 and z-shim techniques123) result in the acquisition of multiple echoes with different dephasing characteristics to reduce susceptibility effects. Alternatively, spin-echo EPI avoids susceptibility artifacts but is much less sensitive to BOLD contrast120. Motion sensitivity. Beyond image misalignment, head motion induces fMRI signal fluctuations owing to partial voluming and spin-history effects39. Lower flip angles124, longer repetition time125 and interleaved slices minimize spin-history effects by allowing the signal in a slice to fully relax before acquisition of the neighboring slice. Spiral sequences are less sensitive to motion during slice acquisition126. The 3D imaging techniques minimize the impact of within-volume motion, but betweenvolume motions remain problematic unless a long repetition time is used; long readouts increase the interaction between magnetic field inhomogeneity and motion127. Parallel imaging can improve acquisition time, resolution and bandwidth but requires calibration scans119. Misalignments between the acquired data and calibration scans induce reconstruction errors and motion-correlated imaging artifacts119. Temporal resolution. Number of slices acquired determines temporal resolution. Echo time fundamentally limits slice- acquisition time. Increasing bandwidth, partial Fourier imaging and parallel imaging can result in marginal speed gains118. Considerations for dMRI Diffusion contrast. Minimizing echo time, and therefore T2 decay, is essential to dMRI. T2 reduces overall magnetic resonance signal and image SNR, and increases in-plane image blurring. Increasing bandwidth, partial Fourier imaging, parallel imaging and using stronger diffusion gradients (b values; high b values require longer diffusion gradient pulses) can minimize echo time118. Higher-strength gradient systems enable high b values (desirable for fiber tracking) in shorter continued Focus on mapping the brain review Box 1 Optimizing fMRI and dMRI image acquisition (continued) time, though require a hardware upgrade for most systems. Echo shifting also reduces echo time, by relaxing the spin echo condition and shifting the echo train earlier in time118. Scan time. Echo time and number of slices determine dMRI volume repetition time. The number of diffusion directions, b values and averages acquired determine total scan time. © 2013 Nature America, Inc. All rights reserved. Emerging acquisition sequences Multiband and multi-echo imaging are emerging acquisition strategies. In multiband imaging, multiple slices are acquired simultaneously, substantially improving temporal resolution, Preprocessing. After acquisition, dMRI data must be preconditioned before directional information can be extracted and tractography can be performed. Little debate exists regarding dMRI data preprocessing (see ref. 14 for an exception), though this may reflect the difficulties of preprocessing and its complex impact rather than consensus in the field. Correcting image distortions. In vivo dMRI data are plagued by spatial distortions, which are a central focus of preprocessing. In particular, magnetic field inhomogeneities are a major contributor. Areas where materials that differ with respect to magnetic susceptibility (that is, extent of magnetization achieved in the MRI) interface with one another (for example, air-tissue interfaces) are particularly prone to such inhomogeneities. These local variations in the magnetic field generate spatial distortions, which can be reduced through parallel imaging techniques or mathematically corrected using estimates of the field variations15. Another cause of spatial distortion is the interaction between the static magnetic field and the currents induced by rapid switching of gradients with the magnetic field, known as eddy currents. These artifacts can be or increases in spatial resolution per unit time128, though at the cost of SNR and spatial smoothness. This improvement enables higher fMRI sampling rates and more diffusion directions per b-value acquisitions in dMRI129, substantially improving estimation of macroscale connectomes130. In multi-echo fMRI, two echoes of a slice are acquired at each acquisition (one with short echo time and one with an echo time optimized for BOLD imaging131). Systematic noise (heartbeat, respiration, head motion and scanner instability) are measured in the first echo (with the short echo time), and removed from the second, enabling signal denoising with practically no SNR cost. reduced using bipolar gradients16. Image coregistration, or spatial alignment of brain scans, is commonly used in preprocessing to correct for eddy-current distortions and subject head motion17. However, this method is ineffective for images acquired using very strong diffusion gradients. Model-based approaches that explicitly account for the effects of eddy currents during image acquisition are emerging as the preferred option18. Overlooked issues. The above preprocessing steps, combined with visual inspection, constitute standard preprocessing. A few important details are often overlooked (see ref. 14 for examples). First, modern scanners are often equipped with antennas that acquire data in parallel through multiple channels that are subsequently combined. Such parallel imaging techniques can increase noise levels if the data are submitted to standard reconstruction19. This can artifactually lower diffusivity estimates along the axons, which increases the tendency of tractography approaches to overfit the data and generate false positive connections. An alternative reconstruction method has recently been proposed to address this important issue19. Second, any correction of distortion must account Table 1 | Magnetic resonance imaging parameters Parameter Echo time Repetition time Bandwidth (1/∆t) Definition Time between slice excitation and acquiring the center of k space The time between acquisitions of adjacent fMRI volumes (sampling period) The range of frequencies mapped to a voxel Flip angle (α) The amount of rotation applied to proton spins by the excitation pulse Spatial resolution (∆x, ∆y, ∆z, Nx, Ny) The volume of tissue sampled in a given voxel; determined by field of view and the number of points sampled in a slice Parallel imaging (Pf) Methods such as GRAPPA and SENSE can decrease repetition time and reduce spatial distortions by sampling k-space lines in parallel; decreases in repetition time are rate-limited by echo time required for BOLD contrast Number of acquisitions that are acquired and subsequently averaged Nacq Impact Determines the impact of spin-spin relaxation (T2) and spin dephasing (T2*) on image Impacts the signal available for imaging (I0), impacted by the number of slices; longer repetition time durations reduce motion sensitivity Lower bandwidth settings can increase artifacts owing to inadequate shimming or susceptibility and distortions in the phase-encoding direction (for example, ‘scalloping’) Impacts I0. Flip angles larger or smaller that the Ernst angle for a given repetition time will reduce I0. Low flip angles may reduce motion sensitivity and in-flow effects and improve spin-lattice relaxation (T1) contrast of images. Signal-to-noise ratio (SNR) is substantially impacted by voxel volume; higher spatial resolution have lower SNR, for example, a 2-mm isovoxel has only ~30% of the SNR of a 3-mm isovoxel, holding all other factors constant Can allow faster image acquisition but at a reduction of SNR by 1/parallelization factor, holding all other factors constant; will increase temporal noise resulting from head motion, respiration or pulsatile effects Improves SNR; does not make sense for fMRI but is commonly used for dMRI nature methods | VOL.10 NO.6 | JUNE 2013 | 527 review Figure 2 | Diffusion imaging of structural connectivity maps for a human brain. (a) dMRIbased map of principal tensor orientations of the human brain viewed from the front. Blue, superior-inferior; green, anteriorposterior; and red, medial-lateral. (b) DTI fiber orientation estimates (red lines) from a region of corpus callosum superimposed on a fractional anisotropy image. Sample streamline is in yellow. Image is magnified 4× relative to that in a. (c) Pyramidal tract streamlines based on deterministic (left) and probabilistic (right) approaches (results superimposed on fractional anisotropy image). Color bar indicates confidence about the presence of the tract. Images courtesy of S. Sotiropoulos (http:// etheses.nottingham.ac.uk/1164/). Focus on mapping the brain a b 2 cm c © 2013 Nature America, Inc. All rights reserved. High for voxel-wise compression or expansion during image coregistration. This is particularly important for dMRI, where distortions vary in direction and magnitude between different gradient orientations, but is largely ignored by current software packages. Finally, image coregistrations contain a rotation component that is applied to each volume and must therefore be applied to the concurrent gradient orientation20. Estimating fiber orientation. Before delineating tracts and bundles via tractography, fiber orientation(s) must be inferred for white-matter voxels individually. Sensitivity of the dMRI signal to water-diffusion properties (such as rate and direction) enables the estimation of fiber orientation at each voxel. As each voxel contains thousands of axonal fibers, not just one, the goal of dMRI analysis is to infer a probability function for each voxel, which captures the different fiber orientations present and their relative proportions21. Estimation of this function— referred to as the fiber orientation density function (fODF)—at each voxel is the first step in estimating structural connectivity. The diffusion tensor is a simplistic but viable model for the diffusion profile that provides a simple approximation to the fODF22. Diffusion tensor imaging (DTI) uses a 3 × 3 matrix to provide an abstract ellipsoid representation of the water-diffusion profile for a given voxel. Mathematical decomposition of this matrix yields information regarding the directions (x, y and z) of maximum and minimum water motion (eigenvectors) as well as the amount of diffusion that occurs along each direction (eigenvalues). The direction of maximal diffusion, referred to as the principal diffusion direction, is taken as the best estimate of fiber orientation within a voxel. Formally speaking, in the case of DTI, the fODF is approximated using a delta function or peak-aligned with the principal diffusion direction. The diffusion tensor provides a good estimate of fiber orientations when axons are homogeneously aligned within a voxel. However, this is not always the case. Fibers are known to disperse (fan), cross, merge and kiss (temporarily run adjacent to one another)—all of which can happen within a single voxel and lead to heterogeneity not accounted for by a simple delta function. More complex approximations of fODF can account for such heterogeneity within a voxel, though they require greater angular coverage23,24 (that is, more directions) and models that either 528 | VOL.10 NO.6 | JUNE 2013 | nature methods Low explicitly or implicitly account for interactions between fiber orientation and the diffusion signal (see ref. 21 for example methods). Complex fODF models better estimate fiber trajectories, particularly when several white-matter tracts intersect and allow recovery of nondominant pathways invisible to DTI25. Estimating edges. After estimation of voxel-wise fiber orientations, tractography approaches are used to establish structural connectivity between connectome nodes. Three-dimensional (3D) trajectories, referred to as ‘streamlines’, are used to trace putative white-matter paths. Local fiber orientation information guides the construction of streamlines along the fODF, allowing us to trace major white-matter bundles26. Results are typically visualized as 3D renderings of thin curves grouped into bundles (Fig. 2), reminiscent of postmortem dissection photographs. The individual streamlines do not represent actual axons; they depict estimates of the average trajectories of axon bundles, given our assumption that diffusion is least hindered along axons. The specific process by which streamlines are developed varies depending on the complexity of the fODF approximations available. With diffusion-tensor modeling, the principal diffusion direction at each voxel guides the formation of the streamline; specifically, it provides a candidate for the tangent to the streamline at each voxel. For more complex fODF models, streamlining follows the same principle, though with multiple peak orientations available at each voxel rather than a single principal diffusion direction. This allows streamlines with differing orientations to pass through the same voxel, which is crucial when heterogeneous fibers are present. Whereas traditional tractography approaches are deterministic, probabilistic approaches account for uncertainty in estimates of local fiber orientations, allowing for estimation of probabilities for any given streamline. Using streamlining methodologies, it is theoretically possible to measure all connections between gray-matter areas. We can estimate both the trajectories and the end points of anatomical pathways. In practice, however, inference of point-to-point © 2013 Nature America, Inc. All rights reserved. Focus on mapping the brain connectivity using streamlining is imprecise and error-prone27; improvements in both data quality and modeling are needed to yield more accurate structural connectomes. Ideally, we should not only be able to infer the existence or absence of connections between nodes, but estimate connection (edge) strengths as well. Anatomical connections are made up of axons; features of these axons, such as density, size, length and myelination, have important consequences on the propagation of action potentials, and hence information transfer. Measures of microstructural features based on more complex dMRI experiments are emerging28, and may become an important component of connectomics. Related measures, such as fractional anisotropy and diffusivity, serve as common proxies for these microstructural complexities, but are extremely sensitive to confounding factors such as partial volume (for example, voxels containing a mixture of white matter and gray matter) and axonal dispersion27,29. Accordingly, these measures should be used with caution to quantify connection strength. Other anatomical factors such as dendrite densities, spine densities, number of synapses at axon terminals and synaptic efficacy are much harder to determine noninvasively, though potentially are more relevant. Unfortunately, tractography does not result in quantification of any of the above properties. Often, probabilistic tractography, which provides an estimate in the uncertainty of streamline trajectories, is used to quantify connection strength. Strong connections are expected to have a more discernible trace in the diffusion data and therefore lower uncertainty in their trajectories. This approximation, however, can easily break. For instance, locally nondominant pathways (for example, those that cross larger bundles), have greater uncertainty. Uncertainty is also affected by nonrelevant factors such as signal-to-noise ratio and partial volume effects. Another issue specific to streamlining is that uncertainty in the streamline’s path increases with the length of tract. Because streamlining operates by propagating uncertainty spatially, connection probabilities inevitably decrease with distance. As a result, tractography-based structural connections are difficult to quantify, threshold, compare between groups and use for other types of statistical analyses27,29. Interpretation and considerations. Pitfalls of tractography can be divided into two categories: accuracy (correctness) and precision (reproducibility)14. Accuracy refers to our ability to infer axonal organization from measurements of water diffusion. In the ideal case, there is no instrument-based or physiological noise, yet we can still make erroneous inferences regarding microstructure because of inaccurate modeling. A white-matter voxel contains hundreds of thousands of axons, which do not necessarily align30. The fODF models, which account for multiple directions in a voxel (crossing fibers) are replacing tensor models for tractography. However, the subvoxel organization of axons can be more complex than a simple crossing and may not always be easily recovered from the diffusion profile. For instance, a collection of axons that bend in a voxel will create a diffusion pattern that may not be easily distinguishable from that of fiber dispersion. One can easily imagine even more complex situations where all these configurations (for example, bending, dispersion and crossing) happen in the same voxel. Diffusion data from a single voxel cannot be used to unambiguously resolve these complexities. Future approaches may benefit from semiglobal review models that aggregate diffusion data across multiple adjacent voxels to infer subvoxel features. With regard to precision, measurement noise (for example, instrument-based or physiological) and inadequate waterdiffusion modeling can compromise tensor and fODF estimation, inducing spurious variations in the generated streamlines. Additionally, use of a fixed step size in the generation of streamlines (despite local variations in anatomy) and the discrete nature of voxels (when tracts are continuous) increase measurement error. Probabilistic tractography algorithms try to quantify these errors by estimating the uncertainty in the entire process. Uncertainty in voxelwise fiber orientation can be quantified31 and propagated into uncertainties regarding the location of streamlines. This process turns 3D point estimates of streamline trajectories into spatial histograms of their locations (Fig. 2c). Beyond concerns regarding accuracy and precision, identification of fibers in their entirety requires knowledge of where tracts terminate throughout cortex. This remains a challenge for tractography algorithms, which are very good at estimating the location of bundles in deep white matter but not good (yet) at identifying where they project into gray matter27. These difficulties result from a lack of detail in white-matter architecture modeled through diffusion and biases in cortical projections. Estimating functional connectivity Although the concept of structural connectivity is relatively intuitive given the presence of physical connections between brain areas, its functional counterpart can be more challenging to define. The macroconnectomics field has adopted a neurophysio logical perspective of functional connectivity, defining it as the synchronization of neurophysiological events between spatially remote brain areas32. First quantified in early electroencephalography and multiunit recording studies, functional connectivity analyses were adopted for positron emission tomography and fMRI in 1993 (ref. 32). Although functional connectivity can be measured noninvasively using a variety of neuroimaging modalities (for example, positron emission tomography, fMRI and magnetoencephalography) and different indices related to physiological function (for example, blood oxygenation level– dependent (BOLD), cerebral blood flow and glucose metabolism analyses), BOLD-based fMRI is the most widely used technique for inferring functional connectivity. Studies of functional connectivity may be dichotomized on the basis of the presence or absence of a task (that is, T-fMRI versus task-free or ‘resting state’ fMRI (R-fMRI)). Task-based approaches focus on the detection of synchronous responses to extrinsic stimulation or tasks, referred to here as evoked functional connectivity (eFC) or coactivation33. Evoked functional connectivity can be quantified across the entire period of task performance or in response to specific types of events. Approaches using R-fMRI focus on the detection of synchronized spontaneous activity occurring in the absence of experimenter-controlled tasks or stimuli, referred to as intrinsic functional connectivity (iFC)34. Although iFC and eFC patterns can be notably similar, especially when eFC is assessed using meta-analytic techniques35 or broad comparisons (for example, task versus rest), these analyses probe different aspects of the functional architecture33. eFC patterns obtained using one task will not necessarily generalize to another, and aspects of iFC obtained during one state may not necessarily nature methods | VOL.10 NO.6 | JUNE 2013 | 529 review Focus on mapping the brain Box 2 iFC and Conscious States © 2013 Nature America, Inc. All rights reserved. The impact of cognitive97, physiological132 and pathological states133 on functional connectivity has been recently demonstrated in the literature. Consciousness has received particular attention, with many studies examining physiological states (for example, sleep), induced states (for example, anesthesia and hypnosis) and pathological states (for example, coma, vegetative syndrome and minimally conscious state133). iFC patterns detected in these states are grossly similar to those observed during wakefulness, though direct comparison reveals state-related iFC changes. For example, the default and lateral networks each exhibit decreased functional connectivity among network components (that is, within-network connectivity) during sleep, suggesting decreased integration generalize to another (for example, wakefulness and sleep; see Box 2 for a discussion of states other than wakeful rest). Acquisition. BOLD is the predominant fMRI technique used in studies of functional connectivity (see ref. 36 for alternative cerebral blood flow–based technique). BOLD is measured using ultrafast imaging sequences that are sensitive to relative concentrations of deoxyhemoglobin, which is paramagnetic (that is, dephases the magnetic resonance signal), in contrast to oxyhemoglobin, which is diamagnetic37; the resulting measurement is an indirect measure of neural activation. Data sets for functionalconnectivity analysis using fMRI are typically obtained in 5–30 minutes, as a participant either performs an experimental task (T-fMRI) or rests quietly in the scanner (typically while awake; R-fMRI). We discuss determinants of BOLD imaging acquisition quality in Box 1. Preprocessing. The aim of preprocessing is to remove confounding variation from data and facilitate comparison across subjects. Structured nuisance signals and anatomical variation can obscure functional connectivity measurements if left unaccounted for. Despite considerable effort, we lack consensus regarding the optimal set of preprocessing steps, their ordering and their implementation. Most preprocessing steps originated with task-activation approaches. However, their use and implications are greater for functional-connectivity approaches because of the greater risk of spurious findings given that the independent and dependent variables can be contaminated by the same noise signals (for example, motion and respiration). Comprehensive comparison of preprocessing strategies and their implications for eFC and iFC analyses remain elusive, in part owing to a lack of objective benchmarks. The preprocessing steps described below are post-hoc corrections. However, optimization of acquisition strategies to minimize the impact of noise sources is preferred (Box 1). Slice timing correction. The slices of an fMRI volume are acquired at different times, creating effective shifts in time series obtained at different slices. Although some question its necessity38, correction by temporal interpolation is recommended to avoid the potential for deleterious impact of these lags on signal denoising and time-series extraction from brain areas. 530 | VOL.10 NO.6 | JUNE 2013 | nature methods of information. Complementary findings of decreased negative iFC between these networks and others may suggest decreased segregation as well134. These results suggest that although sleep-based studies may be useful in populations not amenable to examination in wakeful states (for example, toddlers), comparison of findings across states may be problematic. Of note, changes in thalamocortical connectivity are reported during anesthesia, non–rapid eye movement (non-REM) sleep and vegetative states135, but the specific role of thalamo cortical circuitry in consciousness remains underexplored. These studies also suggest potential clinical applications of iFC, such as improving recognition of consciousness after recovery from coma. Motion correction. Head motion results in a misalignment of brain areas between volumes typically accounted for using 3D imageregistration techniques. Additionally, head motion induces artifactual fMRI signal fluctuations resulting from changes in slice tissue composition (partial voluming) and residual magnetization from prior slice excitations (spin-history effects)39. These motion artifacts are typically modeled and removed in a regression framework, containing predictors calculated from motion parameters estimated during coregistration40. Although effective, modeling-based approaches do not completely remove motion-related fluctuations in the fMRI signal41–43. To address this issue, the ‘scrubbing’ of offending volumes via removal41 or spike regression44 has been proposed. Excluding time points alters the temporal structure of the data, thereby compromising analyses that rely on this structure (for example, temporal dynamics, spectral analysis and estimation of temporal autocorrelation). Regardless of the motion-correction scheme used, it is necessary to account for motion in group-level analyses42,45. Physiological noise correction. Cardiac pulsation and respiration can induce fMRI signal fluctuations, which had led to early criticisms attributing iFC to these physiological signals rather than neural signals46. The cardiac cycle generates pulsatile motion throughout the brain47. Respiratory movement of the chest and abdomen induce changes in the magnetic field, producing intensity fluctuations in fMRI images47. Additionally, changes in cardiac rhythm as well as rate and depth of breathing create longer-term effects. Respiration and pulse can be recorded to model and subsequently remove their impact47. Although this is accepted as ideal, it is not commonly performed. Instead, signals present in white matter and cerebrospinal fluid are taken as surrogates for respiration and cardiac effects, and regressed from the fMRI time series. Incorporating spatial variation in the noise captured by the white-matter signal provides superior denoising (for example, anatomy-based correlation corrections (ANATICOR)48). Blinded source-separation techniques provide another means of physio logical correction (for example, Corsica49). Global signal regression. The mean time series across the whole brain is commonly regressed from the data. In this model, the global signal is considered a nonspecific measure of noise, whose removal improves the specificity of iFC50, decreases motion effects44, and Focus on mapping the brain © 2013 Nature America, Inc. All rights reserved. removes intersession and intersite effects. However, awareness that global signal regression centers the correlation distribution at zero, and thus introduces negative connections51 and can alter inter-individual differences52, has made its use controversial. In this regard, it is important to note that functional correlation coefficients obtained after global signal regression are relative values, not absolute. Additionally, electrophysiological demonstrations of globally synchronous neural signals in gray matter53 call into question the interpretation of the global signal as simply noise. Temporal filtering. Bandpass filtering is usually performed to remove frequencies below 0.001 hertz and greater than 0.08 hertz from the fMRI time series. This frequency range targets removal of low-frequency scanner drift and frequencies above those traditionally associated with functional connectivity34,54. However, complete removal of physiological noise is unlikely because of artifacts induced by the low-temporal-resolution of fMRI (for example, aliasing)46. Concerns about temporal filtering include reductions in the degrees of freedom for the time series and recent demonstrations of functional connectivity at frequencies greater than 0.1 hertz for several brain areas, suggesting that low-pass filtering is removing valuable signal55. Thus, despite historical precedent, inclusion of low-pass filtering merits additional consideration. Spatial normalization and smoothing. Another aspect of preprocessing is conditioning the data for comparison across subjects. Spatial normalization addresses morphological variation across individuals by transforming subject data to a common stereotactic space; population- and study-specific templates are increasingly used to optimize correspondence. Spatial smoothing additionally improves the correspondence of brain areas across individuals and increases the signal-to-noise ratio56. Estimating edges. Several mathematical modeling techniques can be used to define functional relationships, differing primarily in the stringency with which they define functional connectivity. Functional connectivity simply implies a statistical dependency between activities observed in brain areas and is an umbrella term for a wide range of dependency measures, each providing a different perspective32. For example, mutual information measures statistical dependency from the joint-probability distribution function and is sensitive to linear and nonlinear relationships, whereas Pearson’s correlation is primarily sensitive to linear relationships57. Effective connectivity, in contrast, requires a mathe matically precise (directional) description of the interactions between brain areas58. This leads to a plurality of graphs that can be derived from functional connectivity, with each graph characterizing functional interactions from a different perspective. Estimating iFC from R-fMRI data typically begins with extraction of the mean time series across voxels in each brain node (that is, parcellation-defined brain area). Intrinsic functional connectivity is commonly estimated from bivariate tests for statistical dependency (for example, Pearson’s correlation, mutual information and spectral coherence) between every possible pairing of time series59. Although these approaches perform well in simple simulations59, the limited number of observations results in noisy estimates of statistical relationships, which can be reduced using regularization (shrinkage) methods60. A limitation of bivariate approaches is that they do not account for information from review multiple brain areas simultaneously. Hence they cannot be used to distinguish direct from indirect interactions (mediated by common relationships with other areas). Partial correlation (related to the inverse covariance matrix) results in estimates of the conditional linear dependency between two brain areas, after accounting for interactions with every other area61. Although this approach is preferred to bivariate approaches, the number of brain areas commonly exceeds the degrees of freedom, preventing unique specification of partial correlations. In these cases, regularization techniques (for example, graphical lasso and elastic net) can be used to find a solution59. Additionally, information can be pooled across individuals to optimize estimation parameters62. Note that some, but not all, of these approaches enforce symmetry; in other words, one can obtain dependency in one direction but not the other (Box 3 ). Many approaches exist for estimating eFC. Several authors have borrowed approaches used to examine iFC; such approaches are based on the assumption that the time series spans the entire task63 or concatenated blocks of specific task conditions64. Psychophysiological interaction39 analyses directly model interactions between patterns of functional connectivity and the experimental stimulus design, potentially offering greater specificity of findings. Others have measured eFC from ‘coactivation’ using fitted regression coefficients65 or binarized (by applying a threshold to regression coefficients) time series66 generated from a first-level task analysis. Regression coefficient series are then compared using correlation or partial correlation33,65. Binarized time series can be compared from the joint distribution of the two values using measures akin to mutual information66. Finally, meta-analytical approaches provide a means of measuring eFC across studies and often tasks, enabling detection of patterns of coactivation across statistical maps generated from data in the literature35,67. A variety of data-driven techniques are also used for identifying iFC and eFC patterns. Examples include self-organizing maps68, principal component analysis, normalized cut clustering 69 and independent component analysis70. These methods are more appropriate for identifying nodes of connectome graphs than edges, although exceptions exist59. Once functional connectivity is estimated, some applications require it to be thresholded or binarized (that is, to determine whether a connection is present or not). Threshold selection is not straightforward but can be accomplished by applying a test of statistical significance to each edge. When using parametric statistics, care must be taken to adjust the degrees of freedom for temporal autocorrelation. Alternatively, this can be addressed using nonparametric tests of significance such as wavestrapping71 or circular-block bootstrap 72. Sparse covariance estimation methods can also be used. Interpretations and considerations. A common pursuit of T-fMRI and R-fMRI studies is to fractionate the connectome into a set of spatially and functionally distinct networks that can each be annotated in terms of the specific functionality domain they subserve (for example, cognitive, affective or visceral). It is impressive that these T-fMRI and R-fMRI studies have converged on similar definitions of 8–20 spatially and functionally distinct networks, though studies have suggested the actual number of networks is substantially greater13. The concordance of T-fMRI and R-fMRI findings suggests the brain’s intrinsic functional architecture nature methods | VOL.10 NO.6 | JUNE 2013 | 531 review Focus on mapping the brain Box 3 Directionality in functional connectomics © 2013 Nature America, Inc. All rights reserved. Commonly referred to as effective connectivity, the mapping of directional relationships is essential for characterizing information flow in functional-connectivity studies136. Identification of neural drivers can facilitate our understanding of control systems as well as ectopic foci leading to pathological conditions (for example, epilepsy137). Although invasive tracing and stimulation techniques are powerful tools for mapping directional relationships in nonhuman populations138, they are not generally applicable in humans. Here we provide an overview of noninvasive approaches to establishing directionality. Statistical techniques Before reviewing statistical effective connectivity approaches, we note that despite the nomenclature, their findings should not be interpreted as indicating causality but rather the directionality of information flow. Structural equation modeling (SEM) and dynamic causal modeling (DCM) approaches evaluate the fit of hypothesized models of directional inter actions among nodes with measured fMRI data. SEM is a covariance-based approach that represents each node as an exogenous variable (predicting the activity of another node), endogenous variable (activity is predicted by another node) or both139; effective connectivity is modeled at the hemodynamic (BOLD response) level. In contrast, DCM is a generative approach, modeling effective connectivity at the neuronal level based on a given biophysical model and uses a forward model to produce the downstream fMRI activity. Models can be evaluated individually; however, common practice is to compare models representing competing hypotheses regarding causality to identify the ‘best-fit’ model. Concerns exist regarding the feasibility of successfully identifying a ‘best fit’ model from a large population of putative models140. SEM and DCM approaches are limited in the number of nodes and interactions they can model efficiently. Primarily intended for confirmatory analysis, exploratory implementations of SEM141 and DCM136 are emerging for use with R-fMRI. Granger causality analysis (GCA) is a model-free, data-driven effective connectivity approach142 that investigates whether past values of the time series for one node could improve the prediction of the current value in another143. In contrast to SEM and DCM, GCA can be used to assess many nodes simultaneously. Application of GCA to R-fMRI144 is particularly controversial because of (i) limitations imposed by the sluggish and variable provides a framework for moment-to-moment responses to the external world. As summarized in ref. 67, it appears that “the full repertoire of functional networks utilized by the brain in action is continuously and dynamically ‘active’ even when at ‘rest’.” When considering the visualization of functional connectivity (Fig. 3), an important question is: ‘what are we missing?’. Whereas functional connectivity is often represented with static graphs, neurophysiological models have long asserted the transient nature of many functional interactions. Specifically, distributed neural assemblies appear to change their patterns of interaction with one another from one cognitive act or state to another. Consistent with this notion, eFC studies have noted 532 | VOL.10 NO.6 | JUNE 2013 | nature methods hemodynamic response function (HRF), (ii) slow sampling rate (for example, repetition time ≥ 2 s)59 and (iii) assumptions that HRF characteristics are constant across nodes59. Attempts to rehabilitate GCA for R-fMRI include HRF deconvolution before GCA137, accounting for regional variation in the HRF using breath-hold scans and faster sampling rates145. Nonetheless, using GCA with R-fMRI remains problematic. Lesion studies Lesion studies directly perturb the system and therefore can be compelling in establishing causality when carried out in a controlled manner. From this perspective, changes in the activity of a node after disruption of inputs from another are taken to infer causal influences. In the ideal paradigm, scans are obtained before and after occurrence of a given lesion. For example146, R-fMRI scans from the brain of a 6-year-old child before and after callosotomy revealed disruptions specific to interhemispheric connectivity. Lesion studies are, however, commonly limited to scans after lesion, taken after natural occurrence of lesions (for example, stroke, congenital abnor mality, neoplasm and seizure focus). Although useful for testing predictions or generating hypotheses, findings of such studies cannot be considered definitive. Brain stimulation The gold standard for establishing directional influences is the demonstration that direct stimulation of node A impacts node B but not the converse. In this regard, intraoperative studies of corticocortically evoked potentials—which involve the tracking of electrical signals from stimulation sites to other locations—are powerful tools147, though with limited applicability (that is, in neurosurgical patients). Noninvasive stimulation techniques with a broader range of applications are slowly evolving. Transcranial magnetic stimulation (TMS) can be used to produce temporary and reversible neural excitation or suppression in targeted cortical areas. Concurrent TMS-fMRI studies have revealed causal relationships in motor circuits148 and the visual system149. Transcranial direct-current stimulation and transcranial alternating-current stimulation approaches deliver weak currents to targeted brain areas, and are emerging as less expensive and technically demanding alternatives to TMS, though delivering focal stimulation is challenging150. substantial task-dependency in their findings, even when looking at the same regions33,39. Perhaps most exciting, recent iFC studies have observed dynamic changes in iFC patterns over a 5-minute scan73. These findings suggest that commonly used metrics of iFC are incomplete, only capturing the ‘mean’ connectivity over time. If true, the implications would be multifold: (i) findings of hypo- or hyper-connectivity in population studies would need to be reassessed, as they may reflect a different distribution of time spent in the various iFC configurations between populations, and (ii) the detection of changing eFC patterns over the course of task performance may prove to be a means of explaining observed behavioral variability. More and Figure 3 | Visualizing the connectome. (a) Classical anatomical-tracing–style depiction of iFC for posteromedial cortex subdivisions (image reproduced from ref. 113). (b) Matrix representation of functional brain connections predictive of diagnostic status in depression. dmThalamus, dorsomedial thalamus; vMF10, ventral medial prefrontal cortex (Brodmann area (BA) 10); sgCg24|25, subgenual cingulate (BA 24 and 25); aINS, anterior insula; rACC24, right anterior cingulate cortex (BA 24); dlPFC9, dorsolateral prefrontal cortex (BA 9); OFC11, orbitofrontal cortex (BA 11); dMF10, dorsal medial prefrontal cortex (BA 10); MCC24, midcingulate cortex (BA 24); vPCC, ventral posterior cingulate cortex; scACC25, subcallosal cingulate cortex (BA 25) (image reproduced from ref. 90). (c) Flatmap-based representation of the CoCoMac atlas of the macaque connectome (image reproduced from ref. 114). (d) Connectogram depicts brain areas (nodes) as columns in the circular band, differing connectivity metrics in separate layers and connections with lines; lobes are differentiated by color, and left or right halves corresponds to hemispheres (reproduced with permission from ref. 115). review a b dmThalamus vMF10 sgCg24|25 Hypothalamus Nucleus accumbens aINS rACC24 dlPFC9 OFC11 dMF10 MCC24 vPCC Amygdala Hippocampus scACC25 c 6.30 5.91 5.51 5.12 4.73 4.33 3.94 3.54 3.15 2.76 2.36 1.97 1.58 1.18 0.79 0.39 0.00 –0.39 –0.79 –1.18 –1.58 –1.97 –2.36 –2.76 –3.15 –3.54 –3.94 –4.33 –4.72 –5.12 –5.51 –5.91 –6.30 Connection strength (z) © 2013 Nature America, Inc. All rights reserved. Focus on mapping the brain d CCR CCP PCM CCS CCA S1 PCS M1 PMCM PCIP PFCM MACD FCD MCD PCI V1 PFCD FCPCL FEF FCC A2 S2 MCV FCV V2 more studies are highlighting the potential value of examining transition zones between functional areas in the brain5. Examination of temporal dynamics may inform such efforts, by mapping changes in the boundary over time and providing greater clarity for findings. In addition to naturally occurring variations in iFC over time, studies have suggested that iFC can be systematically impacted by cognitive demands before R-fMRI data acquisition. By comparing iFC during R-fMRI scans collected before and after task performance74, studies show that iFC strength within and between networks is altered in a task-dependent manner. For example, R-fMRI–based functional connectivity between the inferior frontal gyrus and visual areas varied depending on the category of stimuli viewed before the R-fMRI scan75. Across participants, the extent to which iFC was modulated correlated with subsequent memory for the stimuli. In addition to exhibiting plasticity related to tasks performed close in time to the measurements, iFC is modulated by direct-stimulation protocols including median nerve stimulation 76, heat pain77, transcranial magnetic stimulation78 and transcranial direct current stimulation79. This suggests that R-fMRI may have utility in the identification of targets for stimulation protocols as well as assessment of their efficacy (for example, in the context of the treatment of depression80). This iFC-based evidence of experience-induced plasticity provides strong support for the hypothesis that iFC reflects a history of coactivation among areas. However, this also suggests a corollary: correlated intrinsic activity has a role in learning and memory consolidation81. The demonstration of brain-behavior correlations between task-related modulations of iFC and subsequent behavior (for example, recall) supports this hypothesis. If short-term iFC alterations reflect experience-induced plasticity, MACV TCV A1 IP TCC FCOR8 TCS IA TCI PHC CPO then enduring changes would be expected after extended practice or training. Several studies suggest this is the case82. Studies of long-term training–induced plasticity have the potential to inform our understanding of mechanisms involved in remediation-based recovery of function or even to index the efficacy of treatment interventions. For example, in a preliminary retrospective study, differences in iFC were observed between children with dyslexia who were remediated by reading interventions versus children who received no treatment83. Statistical analysis of the connectome Once connectome graphs are estimated, the next goal is to annotate them in terms of their relevance to higher-order cognitive processes, neuropsychiatric diagnoses or other phenotypic variables5. These associations are most often inferred by performing a categorical or dimensional statistical analysis that compares connectivity across a population of individuals or within an individual across time or treatments 84. Many of the same statistical approaches are appropriate for the analysis of connectome graphs regardless of whether they were constructed with functional or structural data sets. However, the interpretation of the results must always account for the idiosyncratic differences between structural and functional connectivity. For example, functional connections are typically weighted and can be positive or negative. Structural connectivity graphs tend to be unweighted and are strictly nonnegative27. Additionally, structural connections can be thought of as pathways along nature methods | VOL.10 NO.6 | JUNE 2013 | 533 review © 2013 Nature America, Inc. All rights reserved. which information can flow, but functional connections cannot be interpreted in the same manner85. A bag of edges. The simplest approach to compare graphs is to treat them as a bag, or collection, of edges and perform statistical analyses at each edge one at a time, without taking into account interactions or relationships between them (that is, edgewise statistics)84. Such univariate approaches (for example, t-tests, F-tests or regression) allow researchers to identify easily interpretable relationships between categorical or dimensional variables and edge weights. However, this approach results in the need to perform many statistical tests, which require correction for multiple comparisons to adequately control for the number of false positives. Standard correction techniques such as false discovery rate86 that do not model the dependencies between edges may result in overly liberal or conservative corrections87. Alternate correction techniques such as the network-based statistic88 or group Benjamini-Hochberg89 leverage information about the group structure of connectome graphs to increase statistical power, while maintaining control of false positives. Alternatively, multivariate regression and classification techniques evaluate the relationship between the entire connectome graphs and their associated phenotypic variables with a single statistical test90,91. Although powerful for the analysis of connectome-phenotype relationships, they obscure information about the involvement of individual edges. Extracting this information, if desired, requires a return to edge-specific tests, and the need for multiple-comparison correction90. Although these multivariate techniques tend to be applied to bag-of-edges representations, which ignore graph structure, they can also be performed using graph-distance measures that preserve topological information when comparing graphs92. Node and graph-level statistics: invariants. Graphical representations of connectomes contain a wealth of information about brain architecture beyond the presence and strength of bivariate connections, which can be described using a variety of node-level and graph-level statistics. These measures are called ‘invariants’ in graph-theory parlance or ‘topological measures’ in networktheory parlance because they are not unique to particular representations of the graph. The most commonly used node invariants are centrality measures that indicate a node’s relative influence in a graph. Several different centrality metrics are available that mea sure a node’s importance on the basis of the number and strength of direct connections (degree centrality93), the importance of neighboring nodes (eigenvector94 or Page Rank95 centrality) and their role in connecting other pairs of nodes (betweenness85). The various measures provide different perspectives on a node’s role in the graph and, when combined, can lead to a more holistic understanding of connectome-phenotype relationships95. Similarly, a range of graph-level invariants is used for studying structural and functional connectivity. In particular, graphs are commonly assessed in terms of their local and global efficiency. Local efficiency assesses the extent to which neighbors are densely interconnected, whereas global efficiency captures the number of connections that must be traversed to connect any two nodes93. The relationships of these two measures to what would be obtained from random graphs with similar properties can be combined to assess the ‘small-worldness’ of a graph93. Small-world graphs balance 534 | VOL.10 NO.6 | JUNE 2013 | nature methods Focus on mapping the brain i ntegration and segregation to obtain fast and cost-efficient propagation of information through the graph as well as robustness to single-node failures96. The cost-efficiency of a graph can be inferred from the difference between global efficiency and the number of edges in the graph93. An additional invariant is modularity, which quantifies the extent to which a graph can be segregated into densely intraconnected but sparsely interconnected modules and allows direct comparison of module membership between graphs85. Each of the previously described node and graph invariants can be statistically evaluated to identify relationships with categorical and dimensional phenotypes. Although invariants can increase statistical power by decreasing the number of multiple comparisons, the resulting relationships can be more difficult to interpret. When comparing invariants between graphs, it is important to consider the impact of potential differences in graph properties (for example, number of edges) that can systematically differ between individuals or groups and confound interpretation of findings93. Additionally, as the distribution properties of most invariants are poorly characterized, nonparametric statistical tests are preferred93. Predictive modeling Finally, researchers frequently aim to identify connectivity patterns predictive of a phenotypic variable (for example, diagnosis90, age91 or brain state97). Predictive modeling can be used to directly assess the ability of a connectivity pattern to predict the phenotype of an individual, in contrast to inferential statistics, which evaluate improbability of a set of relationships arising by chance98. Predictive modeling is typically supervised, with the training set consisting of connectivity graphs and their associated phenotypes99. One can assess the predictive accuracy via cross-validation99 or other model-selection techniques. Predictive modeling has primarily focused on invariants and bag-of-edges–style90,100 approaches. Translational connectomics MRI-based approaches to connectomics research are rapidly transforming neuroscience in animal models as well, by removing barriers to longitudinal examinations associated with invasive techniques (for example, animal killing and injection of toxic chemicals). The recent Mouse Biomedical Informatics Research Network initiative (http://www.loni.ucla.edu/BIRN/Projects/ Mouse/) provides an initial demonstration of the potential to complement cross-sectional atlases of the developing brain generated using histology approaches with longitudinal atlases obtained using dMRI. Simultaneously, R-fMRI is emerging as a powerful tool for comparative functional neuroanatomy studies. Initial work has demonstrated impressive correspondence between the iFC observed in humans and macaques for homologous functional networks supporting an array of functions, including those that are putatively ‘human’ (for example, language, self-referential processes and cognition)101. Evidence of homologies with patterns of iFC in lower mammals, such as rats, underscores this translational potential102. Armed with increasingly powerful imaging-based tools, macroscale connectomics studies in animal models are poised to provide a mechanistic understanding of brain function through the combination of noninvasive imaging with direct structural, pharmacological, molecular and genetic manipulations that are impossible in humans. Focus on mapping the brain review Table 2 | Initiatives promising to accelerate macroscale connectomics research © 2013 Nature America, Inc. All rights reserved. Initiative (weblink when available; location) Goals Brain Genomics Superstruct (United States) Aims to collect a large-scale imaging data set to explore brain–behavior relationships and their genetic influences. The initiative has collected R-fMRI, dMRI and saliva samples from over 3,000 adults, along with comprehensive phenotyping data (cognition, personality and lifestyle), and the resultant repository containing 1,500 completed, quality-pass data sets is expected to be publicly available in 2013. Brainnetome (http://www.brainnetome.org/; China) Attempts to characterize brain networks with multimodal neuroimaging techniques, from the microscale (microtechnique, ultramicrotomy, staining and visualization techniques) to the macroscale (electroencephalography, fMRI and dMRI). R-fMRI and diffusion-imaging data sets, along with behavioral and blood data from more than 1,000 patients with schizophrenia, 300 patients with Alzheimer’s disease and mild cognitive impairment, 120 patients who had a stroke, 50 patients with glioma and 2,000 healthy controls collected from 11 hospitals and imaging centers. Consortium of Neuroimagers for the Noninvasive Exploration of Brain Connectivity and Tracts (http://www.brain-connect.eu/; European Union) Consortium focused on studying the brain’s microstructure, tracts and connectivity using dMRI. Target deliverables include optimized acquisition protocols, analytic tools and a connectivity atlas. Developing Human Connectome Project (European Union) Initiative to comprehensively map and model the human connectome for 1,000 babies, including in utero and in vivo imaging (20–44 weeks after conception). US National Institutes of Health Human Connectome Project: Washington University in Saint Louis–University of Minnesota consortium (http://humanconnectome.org/; United States) State-of-the-art multimodal imaging initiative (R-fMRI, T-fMRI, dMRI and magnetoencephalography) that makes use of a twin design (1,200 healthy adults, including twin pairs and their siblings from 300 families) to provide insights into relationships between brain connectivity, behavior and genetics. The project uses multiband imaging sequences for R-fMRI (high spatial and temporal resolution) and dMRI (high spatial resolution), which it has refined and is currently distributing to interested centers. All data and tools developed through the initiative will be openly shared. US National Institutes of Health Human Connectome Project: MGH-Harvard-UCLA consortium (http://humanconnectomeproject.org/; United States) Initiative focusing on unraveling the full connectivity map using the first ‘Connectome Scanner’, which is designed to carry out diffusion using ultrahigh gradient strength (4–8 times the strength of conventional systems). Efforts to optimize dMRI technology will focus on increasing the spatial resolution, quality and speed of acquisition. 1000 Functional Connectomes Project (FCP) (http://fcon_1000.projects.nitrc.org/; global) Grass-roots data-sharing initiative that brought together over 1,200 previously collected R-fMRI data sets from 33 independent sites around the world and released them openly to the scientific community via the Neuroimaging Informatics Tool Resources Clearinghouse (http://www.nitrc.org/). International Neuroimaging Data-sharing Initiative (INDI) (global) Second FCP initiative that was founded in an attempt to (i) expand the scope of open data sharing in the functional neuroimaging community to include phenotypic data beyond age and sex (major INDI data releases: ADHD-200 Consortium (http://fcon_1000.projects.nitrc.org/indi/adhd200/), the Autism Brain Imaging Data Exchange (ABIDE; http://con_1000.projects.nitrc.org/indi/abide/)) and (ii) provide a model for prospective, prepublication data sharing (major release: the Nathan Kline Institute-Rockland Sample http://fcon_1000.projects.nitrc.org/indi/enhanced/). More than 5,000 R-fMRI data sets are available through the FCP and INDI efforts combined as well as a growing number of dMRI data sets. UK Biobank Imaging (http://www.ukbiobank. ac.uk/; UK) Building on an existing long-term prospective epidemiological study that has collected genetics, blood samples and lifestyle information from a cohort of 500,000 subjects, the UK Biobank Imaging Extension aims to resample 100,000 of the cohort using multimodal neuroimaging (including but not limited to R-fMRI and dMRI), as well as cardiac MRI and rich phenotyping. Large-scale initiatives from around the world that are promising to accelerate the pace of macroscale connectomics research through either the advancement of macroconnectomics research through the generation and sharing of large-scale imaging data sets with phenotyping or innovation of data-acquisition and/or analysis techniques (see ref. 116 for additional information regarding these initiatives and others). Despite the rich promise of translational connectomics, methodological issues must also be addressed. For example, iFC can be examined in awake rats that have been habituated to restraint in the loud MRI environment103. However, most studies are conducted under anesthesia—in particular, using the general anesthetic isoflurane104, which can confound findings owing to its effects on neural excitability. The sedating alpha-2 adrenergic agonist medetomidine may be preferable as it avoids such confounds105. Dose-response studies of anesthesia are few104 and are essential. Initial translational studies in monkeys, rats and mice have relied on preprocessing and analytical approaches identical to those developed in humans103. Although their success is encouraging, differences in physiological (for example, cardiac activity and respiration) and imaging parameters must be explored to arrive at optimal strategies. Finally, we note that the many questions raised regarding the interpretation of dMRI and R-fMRI techniques in humans also apply to animal studies. Toward neurophenotypes and clinical applications An overarching goal of the connectomics era is the derivation of ‘neurophenotypes’106, a concept that remains poorly specified despite increasing enthusiasm from investigators. An individual’s macroscale connectome and its subgraphs contribute to the specification of that individual’s neurophenotype. A central goal of connectomics is to catalog neurophenotypes and relate them to phenotypic profiles4. This can be accomplished through datadriven approaches focused on the detection of commonalities and distinctions in connectomes or by differentiating populations of neurophenotypes based on their phenotypic profiles. The breadth of phenotyping can vary depending on the application, though it typically consists of some combination of cognitive, affective, behavioral, neurological or psychiatric variables. When cataloging neurophenotypes based upon macroscale connectomes, the specificity of findings will depend on their nature, granularity of node definitions and quality of neuroimaging data used. Similarly, nature methods | VOL.10 NO.6 | JUNE 2013 | 535 review Focus on mapping the brain Table 3 | Connectomics studies in clinical populations © 2013 Nature America, Inc. All rights reserved. Neuropsychiatric diagnoses Schizophrenia Alzheimer’s disease Depression Epilepsy and seizures Mild cognitive impairment Other neurological disorders Substance dependence Attention deficit hyperactivity disorder Autism spectrum disorders Multiple sclerosis Parkinson’s disease Traumatic brain injury Sleep disorders Stroke Anxiety disorders Coma and vegetative state Obsessive compulsive disorder Amyotrophic lateral sclerosis Bipolar disorder R-fMRI count dMRI count 100 89 81 62 58 45 40 32 26 22 19 18 17 15 15 15 12 10 10 283 201 101 179 115 50 94 37 86 211 73 259 7 182 16 9 27 84 61 Neuropsychiatric disorders most commonly studied using macroscale functional and structural connectivity. Determined by the number (count) of R-fMRI and dMRI papers dedicated to the neuropsychiatric diagnosis. Calculated from the Child Mind Institute Librarian (Resting State and DTI libraries; (http://www.mendeley.com/profiles/cmi-librarian/). field through the introduction of spurious, irreproducible findings associated with suboptimal methodologies. Conversely, increased attention to the acquisition of high-quality data, combined with optimized preprocessing and analytic methodologies can serve to accelerate the pace at which connectomes can be meaningfully annotated and their variations cataloged. Acknowledgments This work was supported by grants from US National Institute of Mental Health (BRAINS R01MH094639 to M.P.M. and K23MH087770 to A.D.M.), the Stavros Niarchos Foundation (M.P.M.), the Brain and Behavior Research Foundation (R.C.C.) and the Leon Levy Foundation (C.K. and A.D.M.). J.T.V. receives funding from the London Institute for Mathematical Sciences HDTRA1-11-1-0048 and US National Institutes of Health R01ES017436. Additional support was provided by a gift from Joseph P. Healey to the Child Mind Institute (M.P.M.). We thank D. Lurie for his assistance in the preparation of the manuscript and references as well as Z. Shehzad, Z. Yang and S. Urchs for their helpful comments. We acknowledge our colleagues who allowed us to reproduce their figures. COMPETING FINANCIAL INTERESTS The authors declare competing financial interests: details are available in the online version of the paper. Reprints and permissions information is available online at http://www.nature. com/reprints/index.html. 1. when sorting neurophenotypes on the basis of phenotypic profiles, specificity will be determined by the precision and comprehensiveness (that is, number and breadth of independent features) of the phenotyping available to statistical analysis. Future work will need to find a balance between categorical and dimensional perspectives of neurophenotypes. Beyond the derivation of a fundamental understanding of brain architecture and its implications for behavior and cognition, a major reason for the excitement surrounding connectomics is the promise of clinical utility because of the ability to obtain individual-relevant reliable brain indices (see Table 2 for initiatives that are accelerating the pace of macroconnectomics research). Recent years have witnessed an explosion in the number of neurological and psychiatric disorders studied with dMRI and R-fMRI (Table 3). Hopes of attaining clinically useful diagnostic tools are increasingly espoused in the literature. However, leaders in the field have recently suggested that the attainment of tools capable of stratifying individuals based upon disease risk, prognosis and treatment response may prove to be a more fruitful goal than focusing on diagnosis107. Regardless, a key requirement remains: attaining large-scale data sets representative of the human population. In this regard, the macroconnectomics community has supported several large-scale data-sharing initiatives dedicated to rapidly aggregating the necessary data108. 2. Conclusion The connectomics era is the culmination of more than a century of conceptual and methodological innovation. MRI-based approaches to mapping and annotating the connectome at the macroscale are transforming basic, translational and clinical neuroscience research by overcoming barriers to progress faced by more traditional invasive methodologies. In this Review we broadly surveyed the many challenges that remain in the acquisition, preprocessing and analysis of brain-imaging data. Failure to consider the many complexities could jeopardize this burgeoning 15. 536 | VOL.10 NO.6 | JUNE 2013 | nature methods 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 16. 17. 18. Sporns, O., Tononi, G. & Kötter, R. The human connectome: A structural description of the human brain. PLoS Comput. Biol. 1, e42 (2005). Varela, F., Lachaux, J.P., Rodriguez, E. & Martinerie, J. The brainweb: phase synchronization and large-scale integration. Nat. Rev. Neurosci. 2, 229–239 (2001). Biswal, B.B. et al. Toward discovery science of human brain function. Proc. Natl. Acad. Sci. USA 107, 4734–4739 (2010). Behrens, T.E.J. & Sporns, O. Human connectomics. Curr. Opin. Neurobiol. 22, 144–153 (2012). Kelly, C., Biswal, B.B., Craddock, R.C., Castellanos, X.F. & Milham, M.P. Characterizing variation in the functional connectome: promise and pitfalls. Trends Cogn. Sci. 16, 181–188 (2012). Talairach, J. & Tournoux, P. Co-planar Stereotaxic Atlas of the Human Brain (Thieme Classics, 1988). Margulies, D.S. et al. Mapping the functional connectivity of anterior cingulate cortex. Neuroimage 37, 579–588 (2007). Beckmann, M., Johansen-Berg, H. & Rushworth, M.F.S. Connectivity-based parcellation of human cingulate cortex and its relation to functional specialization. J. Neurosci. 29, 1175–1190 (2009). Dosenbach, N.U. et al. A core system for the implementation of task sets. Neuron 50, 799–812 (2006). Bellec, P. et al. Identification of large-scale networks in the brain using fMRI. Neuroimage 29, 1231–1243 (2006). Craddock, R.C., James, G.A., Holtzheimer, P.E., Hu, X.P. & Mayberg, H.S. A whole brain fMRI atlas generated via spatially constrained spectral clustering. Hum. Brain Mapp. 33, 1914–1928 (2012). Cohen, A.L. et al. Defining functional areas in individual human brains using resting functional connectivity MRI. Neuroimage 41, 45–57 (2008). Kiviniemi, V. et al. Functional segmentation of the brain cortex using high model order group PICA. Hum. Brain Mapp. 30, 3865–3886 (2009). Jones, D.K. Challenges and limitations of quantifying brain connectivity in vivo with diffusion MRI. Imaging 2, 341–355 (2010). Andersson, J.L.R., Skare, S. & Ashburner, J. How to correct susceptibility distortions in spin-echo echo-planar images: application to diffusion tensor imaging. Neuroimage 20, 870–888 (2003). Alexander, A.L., Tsuruda, J.S. & Parker, D.L. Elimination of eddy current artifacts in diffusion-weighted echo-planar images: the use of bipolar gradients. Magn. Reson. Med. 38, 1016–1021 (1997). Haselgrove, J.C. & Moore, J.R. Correction for distortion of echo-planar images used to calculate the apparent diffusion coefficient. Magn. Reson. Med. 36, 960–964 (1996). Anderson, J. et al. A comprehensive Gaussian Process framework for correcting distortions and movements in diffusion images. in Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM) 20th Annual Meeting and Exhibition (Melbourne, 2012). © 2013 Nature America, Inc. All rights reserved. Focus on mapping the brain 19. Sotiropoulos, S.N. et al. Effects of image reconstruction on fibre orientation mapping from multichannel diffusion MRI: reducing the noise floor using SENSE. Magn. Reson. Med. advance online publication, doi:10.1002/mrm.24623 (7 February 2013). 20. Leemans, A. & Jones, D.K. The B-matrix must be rotated when correcting for subject motion in DTI data. Magn. Reson. Med. 61, 1336–1349 2009). 21. Seunarine, K. & Alexander, D. Multiple fibers: beyond the diffusion tensor. in Diffusion MRI: From Quantitative Measurement to in vivo Neuroanatomy (eds., Johansen-Berg, H. and Behrens, T.E.J.) 55–72 (Academic Press, 2009). 22. Basser, P.J., Mattiello, J. & LeBihan, D. Estimation of the effective self-diffusion tensor from the NMR spin echo. J. Magn. Reson. B 103, 247–254 (1994). 23. Tuch, D.S. et al. High angular resolution diffusion imaging reveals intravoxel white matter fiber heterogeneity. Magn. Reson. Med. 48, 577–582 (2002). 24. Wedeen, V.J., Hagmann, P., Tseng, W.Y., Reese, T.G. & Weisskoff, R.M. Mapping complex tissue architecture with diffusion spectrum magnetic resonance imaging. Magn. Reson. Med. 54, 1377–1386 (2005). 25. Behrens, T.E., Berg, H.J., Jbabdi, S., Rushworth, M.F. & Woolrich, M.W. Probabilistic diffusion tractography with multiple fibre orientations: What can we gain? Neuroimage 34, 144–155 (2007). 26. Catani, M., Howard, R.J., Pajevic, S. & Jones, D.K. Virtual in vivo interactive dissection of white matter fasciculi in the human brain. Neuroimage 17, 77–94 (2002). 27. Jbabdi, S. & Johansen-Berg, H. Tractography: where do we go from here? Brain Connect. 1, 169–183 (2011). 28. Assaf, Y., Blumenfeld-Katzir, T., Yovel, Y. & Basser, P.J. AxCaliber: a method for measuring axon diameter distribution from diffusion MRI. Magn. Reson. Med. 59, 1347–1354 (2008). 29. Jones, D.K., Knosche, T.R. & Turner, R. White matter integrity, fiber count, and other fallacies: The do’s and don’ts of diffusion MRI. Neuroimage 73, 239–254 (2013). 30. Mikula, S., Binding, J. & Denk, W. Staining and embedding the whole mouse brain for electron microscopy. Nat. Methods 9, 1198–1201 (2012). 31. Behrens, T. & Jbabdi, S. MR diffusion tractography. in Diffusion MRI: From Quantitative Measurement to in vivo Neuroanatomy (eds., Johansen-Berg, H. and Behrens, T.E.J.) 333–352 (Academic Press, 2009). 32. Friston, K.J., Frith, C.D., Liddle, P.F. & Frackowiak, R.S. Functional connectivity: the principal-component analysis of large (PET) data sets. J. Cereb. Blood Flow Metab. 13, 5–14 (1993). 33. Mennes, M., Kelly, C., Colcombe, S., Castellanos, F.X. & Milham, M.P. The extrinsic and intrinsic functional architectures of the human brain are not equivalent. Cereb. Cortex 23, 223–229 (2013). 34. Biswal, B., Yetkin, F.Z., Haughton, V.M. & Hyde, J.S. Functional connectivity in the motor cortex of resting human brain using echoplanar MRI. Magn. Reson. Med. 34, 537–541 (1995). 35. Toro, R., Fox, P.T. & Paus, T. Functional coactivation map of the human brain. Cereb. Cortex 18, 2553–2559 (2008). 36. Chuang, K.-H. et al. Mapping resting-state functional connectivity using perfusion MRI. Neuroimage 40, 1595–1605 (2008). 37. Ogawa, S. et al. Intrinsic signal changes accompanying sensory stimulation: functional brain mapping with magnetic resonance imaging. Proc. Natl. Acad. Sci. USA 89, 5951–5955 (1992). 38. Wu, C.W. et al. Empirical evaluations of slice-timing, smoothing, and normalization effects in seed-based, resting-state functional magnetic resonance imaging analyses. Brain Connect. 1, 401–410 (2011). 39. Friston, K.J. et al. Psychophysiological and modulatory interactions in neuroimaging. Neuroimage 6, 218–229 (1997). 40. Friston, K.J., Williams, S., Howard, R., Frackowiak, R.S. & Turner, R. Movement-related effects in fMRI time-series. Magn. Reson. Med. 35, 346–355 (1996). 41. Power, J.D., Barnes, K.A., Snyder, A.Z., Schlaggar, B.L. & Petersen, S.E. Spurious but systematic correlations in functional connectivity MRI networks arise from subject motion. Neuroimage 59, 2142–2154 (2012). 42. Van Dijk, K.R.A., Sabuncu, M.R. & Buckner, R.L. The influence of head motion on intrinsic functional connectivity MRI. Neuroimage 59, 431–438 (2012). 43. Satterthwaite, T.D. et al. Impact of in-scanner head motion on multiple measures of functional connectivity: Relevance for studies of neurodevelopment in youth. Neuroimage 60, 623–632 (2012). 44. Satterthwaite, T.D. et al. An improved framework for confound regression and filtering for control of motion artifact in the preprocessing of resting-state functional connectivity data. Neuroimage 64, 240–256 (2013). review 45. Yan, C.G. et al. A comprehensive assessment of regional variation in the impact of head micromovements on functional connectomics. Neuroimage 76C, 183–201 (2013). 46. Lund, T.E. fcMRI–mapping functional connectivity or correlating cardiacinduced noise? Magn. Reson. Med. 46, 628–629 (2001). 47. Birn, R.M. The role of physiological noise in resting-state functional connectivity. Neuroimage 62, 864–870 (2012). 48. Jo, H.J., Saad, Z.S., Simmons, W.K., Milbury, L.A. & Cox, R.W. Mapping sources of correlation in resting state fMRI, with artifact detection and removal. Neuroimage 52, 571–582 (2010). 49. Perlbarg, V. et al. CORSICA: correction of structured noise in fMRI by automatic identification of ICA components. Magn. Reson. Imaging 25, 35–46 (2007). 50. Fox, M.D., Zhang, D., Snyder, A.Z. & Raichle, M.E. The global signal and observed anticorrelated resting state brain networks. J. Neurophysiol. 101, 3270–3283 (2009). 51. Murphy, K., Birn, R.M., Handwerker, D.a., Jones, T.B. & Bandettini, P.a. The impact of global signal regression on resting state correlations: are anti-correlated networks introduced? Neuroimage 44, 893–905 (2009). 52. Saad, Z. et al. Trouble at rest: how correlation patterns and group differences become distorted after global signal regression. Brain Connect. 2, 25–32 (2012). 53. Schölvinck, M.L., Maier, A., Ye, F.Q., Duyn, J.H. & Leopold, D.A. Neural basis of global resting-state fMRI activity. Proc. Natl. Acad. Sci. USA 107, 10238–10243 (2010). 54. Cordes, D. et al. Frequencies contributing to functional connectivity in the cerebral cortex in “resting-state” data. AJNR Am. J. Neuroradiol. 22, 1326–1333 (2001). 55. Niazy, R.K., Xie, J., Miller, K., Beckmann, C.F. & Smith, S.M. Spectral characteristics of resting state networks. Prog. Brain Res. 193, 259–276 (2011). 56. Worsley, K.J., Marrett, S., Neelin, P. & Evans, A.C. Searching scale space for activation in PET images. Hum. Brain Mapp. 4, 74–90 (1996). 57. Marrelec, G. et al. Regions, systems, and the brain: hierarchical measures of functional integration in fMRI. Med. Image Anal. 12, 484–496 (2008). 58. Friston, K.J. & Frith, C.D. Time dependent changes in effective connectivity measured with PET. Hum. Brain Mapp. 1, 69–79 (1993). 59. Smith, S.M. et al. Network modelling methods for fMRI. Neuroimage 54, 875–891 (2011). 60. Varoquaux, G., Gramfort, A. & Poline, J.B. Advances in Neural Information Processing Systems. (Vancouver, Canada, 2010). 61. Marrelec, G. et al. Partial correlation for functional brain interactivity investigation in functional MRI. Neuroimage 32, 228–237 (2006). 62. Varoquaux, G. et al. A group model for stable multi-subject ICA on fMRI datasets. Neuroimage 51, 288–299 (2010). 63. Lowe, M.J., Dzemidzic, M., Lurito, J.T., Mathews, V.P. & Phillips, M.D. Correlations in low-frequency BOLD fluctuations reflect cortico-cortical connections. Neuroimage 12, 582–587 (2000). 64. Sun, F.T., Miller, L.M. & D’Esposito, M. Measuring interregional functional connectivity using coherence and partial coherence analyses of fMRI data. Neuroimage 21, 647–658 (2004). 65. Rissman, J., Gazzaley, A. & D’Esposito, M. Measuring functional connectivity during distinct stages of a cognitive task. Neuroimage 23, 752–763 (2004). 66. Patel, R.S., Bowman, F.D. & Rilling, J.K. A Bayesian approach to determining connectivity of the human brain. Hum. Brain Mapp. 27, 267–276 (2006). 67. Smith, S.M. et al. Correspondence of the brain’s functional architecture during activation and rest. Proc. Natl. Acad. Sci. USA 106, 13040–13045 (2009). 68. Peltier, S.J., Polk, T.A. & Noll, D.C. Detecting low-frequency functional connectivity in fMRI using a self-organizing map (SOM) algorithm. Hum. Brain Mapp. 20, 220–226 (2003). 69. van den Heuvel, M., Mandl, R. & Hulshoff Pol, H. Normalized cut group clustering of resting-state fMRI data. PLoS ONE 3, e2001 (2008). 70. Beckmann, C.F. Modelling with independent components. Neuroimage 62, 891–901 (2012). 71. Breakspear, M., Brammer, M.J., Bullmore, E.T., Das, P. & Williams, L.M. Spatiotemporal wavelet resampling for functional neuroimaging data. Hum. Brain Mapp. 23, 1–25 (2004). 72. Bellec, P., Marrelec, G. & Benali, H. A bootstrap test to investigate changes in brain connectivity for functional MRI. Stat. Sin. 18, 1253–1268 (2008). nature methods | VOL.10 NO.6 | JUNE 2013 | 537 © 2013 Nature America, Inc. All rights reserved. review 73. Smith, S.M. et al. Temporally-independent functional modes of spontaneous brain activity. Proc. Natl. Acad. Sci. USA 109, 3131–3136 (2012). 74. Waites, A.B., Stanislavsky, A., Abbott, D.F. & Jackson, G.D. Effect of prior cognitive state on resting state networks measured with functional connectivity. Hum. Brain Mapp. 24, 59–68 (2005). 75. Stevens, W.D., Buckner, R.L. & Schacter, D.L. Correlated low-frequency BOLD fluctuations in the resting human brain are modulated by recent experience in category-preferential visual regions. Cereb. Cortex 20, 1997–2006 (2010). 76. Klingner, C.M., Hasler, C., Brodoehl, S., Axer, H. & Witte, O.W. Perceptual plasticity is mediated by connectivity changes of the medial thalamic nucleus. Hum. Brain Mapp. advance online publication, doi:10.1002/ hbm.22074 (25 March 2012). 77. Riedl, V. et al. Repeated pain induces adaptations of intrinsic brain activity to reflect past and predict future pain. Neuroimage 57, 206–213 (2011). 78. Vercammen, A., Knegtering, H., Liemburg, E.J., den Boer, J.A. & Aleman, A. Functional connectivity of the temporo-parietal region in schizophrenia: effects of rTMS treatment of auditory hallucinations. J. Psychiatr. Res. 44, 725–731 (2010). 79. Keeser, D. et al. Prefrontal transcranial direct current stimulation changes connectivity of resting-state networks during fMRI. J. Neurosci. 31, 15284–15293 (2011). 80. Fox, M.D., Buckner, R.L., White, M.P., Greicius, M.D. & Pascual-Leone, A. Efficacy of transcranial magnetic stimulation targets for depression is related to intrinsic functional connectivity with the subgenual cingulate. Biol. Psychiatry 72, 595–603 (2012). 81. Tambini, A., Ketz, N. & Davachi, L. Enhanced brain correlations during rest are related to memory for recent experiences. Neuron 65, 280–290 (2010). 82. Lewis, C.M., Baldassarre, A., Committeri, G., Romani, G.L. & Corbetta, M. Learning sculpts the spontaneous activity of the resting human brain. Proc. Natl. Acad. Sci. USA 106, 17558–17563 (2009). 83. Koyama, M.S. et al. Cortical signatures of dyslexia and remediation: an intrinsic functional connectivity approach. PLoS ONE 8, e55454 (2013). 84. Varoquaux, G. & Craddock, R.C. Learning and comparing functional connectomes across subjects. Neuroimage advance online publication, doi:10.1016/j.neuroimage.2013.04.007 (11 April 2013). 85. Rubinov, M. & Sporns, O. Complex network measures of brain connectivity: uses and interpretations. Neuroimage 52, 1059–1069 (2010). 86. Genovese, C.R., Lazar, N.A. & Nichols, T. Thresholding of statistical maps in functional neuroimaging using the false discovery rate. Neuroimage 15, 870–878 (2002). 87. Efron, B. When should hypothesis testing problems be combined? Ann. Appl. Stat. 2, 197–223 (2008). 88. Zalesky, A., Fornito, A. & Bullmore, E.T. Network-based statistic: identifying differences in brain networks. Neuroimage 53, 1197–1207 (2010). 89. Hu, J.X., Zhao, H. & Zhou, H.H. False Discovery Rate Control With Groups. J. Am. Stat. Assoc. 105, 1215–1227 (2010). 90. Craddock, R.C., Holtzheimer, P.E., Hu, X.P. & Mayberg, H.S. Disease state prediction from resting state functional connectivity. Magn. Reson. Med. 62, 1619–1628 (2009). 91. Dosenbach, N.U.F. et al. Prediction of individual brain maturity using fMRI. Science 329, 1358–1361 (2010). 92. Bunke, H. A graph distance metric based on the maximal common subgraph. Pattern Recognit. Lett. 19, 255–259 (1998). 93. Bullmore, E. & Sporns, O. Complex brain networks: graph theoretical analysis of structural and functional systems. Nat. Rev. Neurosci. 10, 186–198 (2009). 94. Bullmore, E.T. & Bassett, D.S. Brain graphs: graphical models of the human brain connectome. Annu. Rev. Clin. Psychol. 7, 113–140 (2011). 95. Zuo, X.N. et al. Network centrality in the human functional connectome. Cereb. Cortex 22, 1862–1875 (2012). 96. Bassett, D.S., Meyer-Lindenberg, A., Achard, S., Duke, T. & Bullmore, E. Adaptive reconfiguration of fractal small-world human brain functional networks. Proc. Natl. Acad. Sci. USA 103, 19518–19523 (2006). 97. Richiardi, J., Eryilmaz, H., Schwartz, S., Vuilleumier, P. & Van De Ville, D. Decoding brain states from fMRI connectivity graphs. Neuroimage 56, 616–626 (2011). 98. Hansen, L.K. Multivariate strategies in functional magnetic resonance imaging. Brain Lang. 102, 186–191 (2007). 538 | VOL.10 NO.6 | JUNE 2013 | nature methods Focus on mapping the brain 99. Hastie, T., Tibshirani, R. & Friedman, J. The Elements of Statistical Learning (Springer, 2001). 100. Zhu, C.Z. et al. Discriminative analysis of brain function at resting-state for attention-deficit/hyperactivity disorder. Med. Image Comput. Comput. Assist. Interv. 8, 468–475 (2005). 101. Vincent, J.L. et al. Intrinsic functional architecture in the anaesthetized monkey brain. Nature 447, 83–86 (2007). 102. Pawela, C.P. et al. Resting-state functional connectivity of the rat brain. Magn. Reson. Med. 59, 1021–1029 (2008). 103. Becerra, L., Pendse, G., Chang, P.-C., Bishop, J. & Borsook, D. Robust reproducible resting state networks in the awake rodent brain. PLoS ONE 6, e25701 (2011). 104. Wang, K. et al. Temporal scaling properties and spatial synchronization of spontaneous blood oxygenation level-dependent (BOLD) signal fluctuations in rat sensorimotor network at different levels of isoflurane anesthesia. NMR Biomed. 24, 61–67 (2011). 105. Pawela, C.P. et al. A protocol for use of medetomidine anesthesia in rats for extended studies using task-induced BOLD contrast and resting-state functional connectivity. Neuroimage 46, 1137–1147 (2009). 106. Sörös, P. & Stanton, S.G. On variability and genes: inter-individual differences in auditory brain function. Front. Hum. Neurosci. 6, 150 (2012). 107. Kapur, S., Phillips, A.G. & Insel, T.R. Why has it taken so long for biological psychiatry to develop clinical tests and what to do about it? Mol. Psychiatry 17, 1174–1179 (2012). 108. Mennes, M., Biswal, B.B., Castellanos, F.X. & Milham, M.P. Making data sharing work: the FCP/INDI experience. Neuroimage advance online publication, doi:10.1016/j.neuroimage.2012.10.064 (30 October 2012). 109. Tzourio-Mazoyer, N. et al. Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. Neuroimage 15, 273–289 (2002). 110. Desikan, R.S. et al. An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage 31, 968–980 (2006). 111. Eickhoff, S.B. et al. A new SPM toolbox for combining probabilistic cytoarchitectonic maps and functional imaging data. Neuroimage 25, 1325–1335 (2005). 112. Lancaster, J.L. et al. Automated Talairach atlas labels for functional brain mapping. Hum. Brain Mapp. 10, 120–131 (2000). 113. Margulies, D.S. et al. Precuneus shares intrinsic functional architecture in humans and monkeys. Proc. Natl. Acad. Sci. USA 106, 20069–20074 (2009). 114. Knock, S.A. et al. The effects of physiologically plausible connectivity structure on local and global dynamics in large scale brain models. J. Neurosci. Methods 183, 86–94 (2009). 115. Van Horn, J.D. et al. Mapping connectivity damage in the case of Phineas Gage. PLoS ONE 7, e37454 (2012). 116. Jiang, T. Brainnetome: a new -ome to understand the brain and its disorders. Neuroimage advance online publication, doi:10.1016/j. neuroimage.2013.04.002 (6 April 2013). 117. Haacke, E.M., Bornw, R.W., Thompson, M.R. & Venkatesan, R. Magnetic Resonance Imaging: Physical Principles and Sequence Design (Wiley-Liss, 1999). 118. Bernstein, M.A., King, K.F. & Zhou, X.J. Handbook of MRI Pulse Sequences (Academic Press, 2004). 119. Blaimer, M. et al. SMASH, SENSE, PILS, GRAPPA: how to choose the optimal method. Top. Magn. Reson. Imaging 15, 223–236 (2004). 120. Yacoub, E. et al. Spin-echo fMRI in humans using high spatial resolutions and high magnetic fields. Magn. Reson. Med. 49, 655–664 (2003). 121. Weiskopf, N., Hutton, C., Josephs, O. & Deichmann, R. Optimal EPI parameters for reduction of susceptibility-induced BOLD sensitivity losses: a whole-brain analysis at 3 T and 1.5 T. Neuroimage 33, 493–504 (2006). 122. Glover, G.H. & Law, C.S. Spiral-in/out BOLD fMRI for increased SNR and reduced susceptibility artifacts. Magn. Reson. Med. 46, 515–522 (2001). 123. Heberlein, K.A. & Hu, X. Simultaneous acquisition of gradient-echo and asymmetric spin-echo for single-shot z-shim: Z-SAGA. Magn. Reson. Med. 51, 212–216 (2004). 124. Gonzalez-Castillo, J., Roopchansingh, V., Bandettini, P.A. & Bodurka, J. Physiological noise effects on the flip angle selection in BOLD fMRI. Neuroimage 54, 2764–2778 (2011). 125. Grootoonk, S. et al. Characterization and correction of interpolation effects in the realignment of fMRI time series. Neuroimage 11, 49–57 (2000). © 2013 Nature America, Inc. All rights reserved. Focus on mapping the brain 126. Noll, D.C., Cohen, J.D., Meyer, C.H. & Schneider, W. Spiral K-space MR imaging of cortical activation. J. Magn. Reson. Imaging 5, 49–56 (1995). 127. Rabrait, C. et al. High temporal resolution functional MRI using parallel echo volumar imaging. J. Magn. Reson. Imaging 27, 744–753 (2008). 128. Moeller, S. et al. Multiband multislice GE-EPI at 7 tesla, with 16-fold acceleration using partial parallel imaging with application to high spatial and temporal whole-brain fMRI. Magn. Reson. Med. 63, 1144–1153 (2010). 129. Setsompop, K. et al. Improving diffusion MRI using simultaneous multi-slice echo planar imaging. Neuroimage 63, 569–580 (2012). 130. Feinberg, D.A. et al. Multiplexed echo planar imaging for sub-second whole brain FMRI and fast diffusion imaging. PLoS ONE 5, e15710 (2010). 131. Bright, M.G. & Murphy, K. Removing motion and physiological artifacts from intrinsic BOLD fluctuations using short echo data. Neuroimage 64, 526–537 (2013). 132. Heine, L. et al. Resting state networks and consciousness: alterations of multiple resting state network connectivity in physiological, pharmacological, and pathological consciousness States. Front. Psychol. 3, 295 (2012). 133. Boly, M. et al. Brain connectivity in disorders of consciousness. Brain Connect. 2, 1–10 (2012). 134. Horovitz, S.G. et al. Decoupling of the brain’s default mode network during deep sleep. Proc. Natl. Acad. Sci. USA 106, 11376–11381 (2009). 135. Spoormaker, V.I. et al. Development of a large-scale functional brain network during human non-rapid eye movement sleep. J. Neurosci. 30, 11379–11387 (2010). 136. Friston, K. Dynamic causal modeling and Granger causality Comments on: the identification of interacting networks in the brain using fMRI: model selection, causality and deconvolution. Neuroimage 58, 303–305 (2011). 137. David, O. et al. Identifying neural drivers with functional MRI: an electrophysiological validation. PLoS Biol. 6, 2683–2697 (2008). 138. Scannell, J.W., Burns, G.A., Hilgetag, C.C., O’Neil, M.A. & Young, M.P. The connectional organization of the cortico-thalamic system of the cat. Cereb. Cortex 9, 277–299 (1999). review 139. Mclntosh, A.R. & Gonzalez-Lima, F. Structural equation modeling and its application to network analysis in functional brain imaging. Hum. Brain Mapp. 2, 2–22 (1994). 140. Lohmann, G., Erfurth, K., Müller, K. & Turner, R. Critical comments on dynamic causal modelling. Neuroimage 59, 2322–2329 (2012). 141. Zhuang, J., LaConte, S., Peltier, S., Zhang, K. & Hu, X. Connectivity exploration with structural equation modeling: an fMRI study of bimanual motor coordination. Neuroimage 25, 462–470 (2005). 142. Goebel, R. Investigating directed cortical interactions in time-resolved fMRI data using vector autoregressive modeling and Granger causality mapping. Magn. Reson. Imaging 21, 1251–1261 (2003). 143. Granger, C.W. Investigating causal relations by econometric models and cross-spectral methods. Econometrica 37, 424–438 (1969). 144. Sridharan, D., Levitin, D.J. & Menon, V. A critical role for the right fronto-insular cortex in switching between central-executive and default-mode networks. Proc. Natl. Acad. Sci. USA 105, 12569–12574 (2008). 145. Deshpande, G. & Hu, X. Investigating effective brain connectivity from fMRI data: past findings and current issues with reference to Granger causality analysis. Brain Connect. 2, 235–245 (2012). 146. Johnston, J.M. et al. Loss of resting interhemispheric functional connectivity after complete section of the corpus callosum. J. Neurosci. 28, 6453–6458 (2008). 147. Matsumoto, R. et al. Functional connectivity in the human language system: a cortico-cortical evoked potential study. Brain 127, 2316–2330 (2004). 148. Bohning, D.E. et al. Echoplanar BOLD fMRI of brain activation induced by concurrent transcranial magnetic stimulation. Invest. Radiol. 33, 336–340 (1998). 149. Ruff, C.C. et al. Distinct causal influences of parietal versus frontal areas on human visual cortex: evidence from concurrent TMS-fMRI. Cereb. Cortex 18, 817–827 (2008). 150. Datta, A. et al. Gyri-precise head model of transcranial direct current stimulation: improved spatial focality using a ring electrode versus conventional rectangular pad. Brain Stimulat. 2, 201–207 (2009). nature methods | VOL.10 NO.6 | JUNE 2013 | 539 RESOURCE Improved tools for the Brainbow toolbox npg © 2013 Nature America, Inc. All rights reserved. Dawen Cai1,2, Kimberly B Cohen1,2, Tuanlian Luo1,2, Jeff W Lichtman1,2 & Joshua R Sanes1,2 In the transgenic multicolor labeling strategy called ‘Brainbow’, Cre-loxP recombination is used to create a stochastic choice of expression among fluorescent proteins, resulting in the indelible marking of mouse neurons with multiple distinct colors. This method has been adapted to non-neuronal cells in mice and to neurons in fish and flies, but its full potential has yet to be realized in the mouse brain. Here we present several lines of mice that overcome limitations of the initial lines, and we report an adaptation of the method for use in adenoassociated viral vectors. We also provide technical advice about how best to image Brainbow-expressing tissue. The discovery that recombinant jellyfish GFP fluoresces when expressed in heterologous cells1 has led to a vast array of powerful methods for marking and manipulating cells, subcellular compartments and molecules. The discovery or design of numerous spectral variants2–13 (fluorescent proteins collectively called XFPs; ref. 14) expanded the scope of the ‘GFP revolution’ by enabling discrimination of nearby cells or processes labeled with contrasting colors. At least for the nervous system, however, two or three colors are far too few because each axon or dendrite approaches hundreds or thousands of other processes in the crowded neuropil of the brain. Several years ago, we developed a transgenic strategy called Brainbow15 that addresses this problem by marking neurons with many different colors. In this method, three or four XFPs are incorporated into a transgene, and the Cre-loxP recombination system16 is used to make a stochastic ‘choice’ of a single XFP to be expressed from the cassette. Because multiple cassettes are integrated at a single genomic site, and the choice within each cassette is made independently, combinatorial expression can endow individual neurons with 1 of ~100 colors, providing nearby neurons with distinct spectral identities. If Cre recombinase is expressed transiently, descendants of the initially marked cell inherit the color of their progenitor. Accordingly, the Brainbow method has been adapted for use in lineage analysis in non-neural tissues of mice17–21. In addition, it has been adapted for analyses of neuronal connectivity, cell migration and lineage in fish22,23 and flies24,25. In contrast, the method has been little used in the mouse nervous system26. We believe that the main reasons for this are limitations of the initial set of Brainbow transgenic mice. These include suboptimal fluorescence intensity, failure to fill all axonal and dendritic processes, and disproportionate expression of the ‘default’ (that is, nonrecombined) XFP in the transgene. We have now addressed several of these limitations, and we present here a new set of Brainbow reagents. In addition, we provide guidelines for imaging Brainbowexpressing tissue. RESULTS Design of Brainbow 3.0 transgenes As a first step in improving Brainbow methods, we sought XFPs with minimal tendency to aggregate in vivo, high photostability and maximal stability with respect to paraformaldehyde fixation. Because some XFPs that ranked highly in cultured cells performed poorly in vivo, we generated transgenic lines from 15 XFPs (Supplementary Table 1 and refs. 2–13). Of the XFPs tested in this way, seven were judged suitable: mTFP1, EGFP, EYFP, mOrange2, TagRFPt, tdTomato and mKate2. From these XFPs, we chose three according to the criteria of minimal spectral overlap and minimal sequence homology. These XFPs were mOrange2 from a coral (excitation peak (Ex) = 549 nm, emission peak (Em) = 565 nm), EGFP from a jellyfish (Ex = 488 nm, Em = 507 nm) and mKate2 from a sea anemone (Ex = 588 nm, Em = 635 nm)2,9,11. Our reason for minimizing sequence homology was to ensure that the XFPs would be antigenically distinct, in contrast to spectrally distinct but antigenically indistinguishable jellyfish variants (EBFP, ECFP, EGFP and EYFP). Exploiting this property, we generated antibodies to the XFPs in different host species (rabbit anti-mCherry, antimOrange2 and anti-tdTomato; chicken anti-EGFP, anti-EYFP and anti-ECFP; and guinea pig anti-mKate2, anti-TagBFP and antiTagRFP; Supplementary Table 1). Tests in transfected cultured cells confirmed a lack of cross-reactivity (data not shown). Next we addressed the need to fill all parts of the cell evenly. Unmodified XFPs labeled somata so strongly that nearby processes were difficult to resolve, whereas palmitoylated derivatives, which targeted the XFPs to the plasma membrane, were selectively transported to axons and labeled dendrites poorly15. We therefore generated farnesylated derivatives27, which were trafficked to membranes of all neurites (see below). On the basis of these results, we generated ‘Brainbow 3.0’ transgenic lines incorporating farnesylated derivatives of mOrange2, EGFP and mKate2. We retained the Brainbow 1 format15, in which 1Center for Brain Science, Harvard University, Cambridge, Massachusetts, USA. 2Department of Molecular and Cellular Biology, Harvard University, Cambridge, Massachusetts, USA. Correspondence should be addressed to J.R.S. ([email protected]). Received 3 October 2012; accepted 30 March 2013; published online 5 May 2013; doi:10.1038/nmeth.2450 540 | VOL.10 NO.6 | JUNE 2013 | nature methods npg © 2013 Nature America, Inc. All rights reserved. RESOURCE Figure 1 | Brainbow 3 transgenic mice. (a) Brainbow constructs and transgenic mice. Brainbow 1.0 is described in ref. 15. Brainbow 3.0 incorporates farnesylated (‘f’), antigenically distinct XFPs: mOrange2f (mO2f), EGFPf and mKate2f (mK2f). Brainbow 3.1 incorporates a nuclear-targeted nonfluorescent XFP (ØNFPnls) in the first (default) position. Brainbow 3.2 incorporates a woodchuck hepatitis virus posttranscriptional regulatory element (W) into Brainbow 3.1. P, loxP site; 2, lox2272; N, loxN; pA, polyadenylation sequence. (b,c) Low- (b) and high-power (c) views of muscles from a Brainbow 3.0 (line D) Islet-Cre mouse, showing terminal axons and neuromuscular junctions in extraocular muscle. (d) Rotated image along dashed bar in c showing five motor axons labeled in distinct colors. The open circles show that farnesylated XFPs mark plasma membranes more than cytoplasm. (e) Cerebellum from a Brainbow 3.1 (line 3) L7-Cre mouse. The ten Purkinje cells in this field are labeled by at least seven distinct colors (antibody amplified and numbered i–vii). Because Cre is selectively expressed by Purkinje cells in the L7-Cre line, no other cell types are labeled. (f) Cerebellum from a Brainbow 3.2 (line 7) CAGGS-CreER mouse showing granule native fluorescence in red, pink, yellow, green, cyan, blue and brown. P, parallel fibers in molecular layer. Purkinje cell bodies, which are unlabeled, are outlined. Scale bars: 50 µm (b), 20 µm (c,e), 5 µm (d), 10 µm (f). a Brainbow 1.0 Thy1 P 2 RFP pA P YFP pA 2 CFP pA mO2f pA 2 EGFPf pA N mK2f pA Brainbow 3.0 Thy1 2 N Brainbow 3.1 Thy1 P 2 N ∅NFPnls pA P mO2f pA 2 ∅NFPnls pA P mO2f W pA 2 EGFPf pA N pA mK2f Brainbow 3.2 Thy1 P 2 N b c e EGFPf W pA N W pA mK2f d f vii v P iv iii ii i incompatible wild-type and mutant loxP sites are concatenated so that Cre recombinase yields a stochastic choice among XFPs (Fig. 1a). We also retained two other features of the Brainbow 1 strategy. First, we used neuron-specific regulatory elements from the Thy1 gene14 because it promotes high levels of transgene expression in many, although not all, neuronal types; other promoter-enhancer sequences that we tested support considerably lower levels of expression21. Second, we generated transgenic lines by injection into oocytes because this method leads to integration of multiple copies of the cassette and, thus, a broad spectrum of outcomes15,28; by contrast, knock-in lines generated by homologous recombination contain one or two copies of the cassette (as heterozygotes or homozygotes, respectively) and, consequently, a smaller number of possible color combinations17–20. Design of Brainbow 3.1 and 3.2 transgenes In Brainbow 1, 2 and 3.0 (Fig. 1a) and ref. 15, one XFP is expressed ‘by default’ in Cre-negative cells. The presence of a default XFP has both drawbacks and advantages. In cases of limited Cre expression, this XFP is expressed in a majority of cells, reducing spectral diversity among recombined neurons. On the other hand, incorporation of a default XFP allows one to screen numerous lines in the absence of a Cre reporter to assess the number and types of cells in which XFPs could be expressed following recombination. To eliminate the default XFP while retaining the ability to assess expression in the absence of Cre, we adopted the following strategy. First, we incorporated three rather than two pairs of incompatible loxP sites, which allowed insertion of a ‘stop’ cassette29 vi i ii iii iv v vi vii into a first position. With this modification, the three XFPs were expressed only in Cre-positive neurons, giving more spectral diversity. Second, we inserted a fourth XFP, Phi-YFP (from the hydrozoon Phialidium)4, which is antigenically distinct from the other three, into the stop cassette. We mutated Phi-YFP to eliminate its endogenous fluorescence (PhiYFP Y65A), fused it to a nuclear localization sequence and generated antibodies against it in rat. In Brainbow 3.1 mice generated from this cassette (Fig. 1a), one can screen sections with rat anti–Phi-YFP before Cre recombination (Supplementary Fig. 1). Finally, we inserted a sequence that stabilizes mRNAs, called a woodchuck hepatitis virus post-transcriptional regulatory element (WPRE), into the 3′ untranslated sequences following each XFP. The WPRE has been used in many cases to increase protein levels produced by viral vectors and transgenes30,31. We call lines incorporating the WPRE ‘Brainbow 3.2’ (Fig. 1a). Characterization of Brainbow 3 mice We generated 31 lines of transgenic mice from the Brainbow 3.0, 3.1 and 3.2 cassettes. Offspring were crossed with several Cre transgenic lines32–37, leading to multicolor spectral labeling of neuronal populations in numerous regions including cerebral cortex, brainstem, cerebellum, spinal cord and retina (Fig. 1b–f, Supplementary Table 2, Supplementary Fig. 2 and Supplementary Videos 1 and 2). The intensity of expression varied markedly among lines, making quantitative comparison of dubious value, but the most nature methods | VOL.10 NO.6 | JUNE 2013 | 541 RESOURCE npg a c d e f h g i strongly expressing lines were those that incorporated the WPRE (Brainbow 3.2). Analysis of these lines confirmed their advantages over Brainbow 1 and 2 lines15. First, the use of farnesylated XFPs led the XFPs to concentrate at the plasma membrane (Fig. 1d). As a consequence, somata were less intensely labeled in Brainbow 3 than in Brainbow 1 mice, so processes could be visualized without saturation of the somata (Fig. 2a,b and Supplementary Fig. 3). Use of farnesylated XFPs also improved labeling of fine processes and dendritic spines (Fig. 2a,b). Second, the ability to immunostain all three XFPs led to enhancement of the intrinsic fluorescence without loss of color diversity (Fig. 2c,d and Supplementary Fig. 4). Third, XFP fluorescence was visible only in Crepositive cells, thereby correcting the color imbalance caused by default XFP expression in cells that were Cre negative or cells that expressed Cre at low levels in Brainbow 1 and 2 lines (Fig. 2e–i and Supplementary Fig. 5). Thus, Brainbow 3 lines are likely to be more useful than Brainbow 1 and 2 lines for multicolor labeling. Brainbow with self-excising Cre recombinase In Brainbow 1–3 lines, the cassette encodes XFPs separated by loxP sites; Cre recombinase is supplied from a separate transgene. For analysis of connectivity in mouse mutants, breeding mice with two transgenes (Brainbow and Cre) into an already complex background is cumbersome. We therefore attempted to combine constructs encoding XFPs and Cre in a single cassette. In this transgene, called ‘Autobow’, we substituted a self-excising Cre recombinase38 for the stop sequences in Brainbow 3.1 (Fig. 3a). The Thy1 regulatory elements lead to expression of Cre in differentiated neurons; Cre then simultaneously activates an XFP and excises itself. We generated three founder mice using this construct. Two were analyzed as adults, and both of these exhibited 542 | VOL.10 NO.6 | JUNE 2013 | nature methods b % of total neurons © 2013 Nature America, Inc. All rights reserved. Figure 2 | Improved visualization of neurons in Brainbow 3 mice. (a,b) Cerebellum from Brainbow 1.0 (a) and Brainbow 3.1 (line 3) L7-Cre (b) mice. Insets show high-magnification views of boxed regions. The farnesylated XFPs clearly label the fine processes and dendritic spines (b, inset arrows), which are missing in the cytoplasmic labeling (a, inset). (c,d) Retina from a Brainbow 3.0 (line D) Islet-Cre mouse expressing EGFP (blue), mOrange2 (green) and mKate2 (red). Intrinsic XFP fluorescence (c) and a nearby section immunostained with chicken anti-GFP, rabbit anti-mOrange2 and guinea pig anti-mKate2 (d) are shown. (e–h) Immunostained cerebellum section from Brainbow 3.2 (line 7) parvalbumin-Cre mouse. Separate channels of region boxed in e are shown in f–h. (i) Fraction of all labeled neurons that express EGFP, mOrange2 (mO2) or mKate2 (mK2) (n = 2,037 neurons from 15 regions of two brains). Scale bars: 40 µm (a,b), 20 µm (c,d), 25 µm (e–g), 10 µm (insets). 50 n = 2,037 40 30 20 10 0 EGFP+ mO2+ mK2+ combinatorial expression of XFPs in multiple neurons (Fig. 3b and Supplementary Fig. 2e). We were concerned that precocious Cre activation in the germ line might lead to loss of the cassette. We therefore established a line from the third founder and examined mice in the second and sixth generations. Color range was limited in this line, perhaps because only a few copies had been integrated into the genome, but the variety of colors and level of expression were similar in both generations (Fig. 3c,d). Thus, Autobow transgenes can be stably maintained. Brainbow using Flp recombinase and FRT sites A second recombination system, orthogonal to Cre-loxP, could be used to independently control expression of distinct XFPs in, for example, excitatory and inhibitory neurons. In this way, the color of a neuron could denote cell identity, a feature lacking in currently available Brainbow lines15,17–20. We therefore tested a second recombination system, in which Flp recombinase acts on Flp recombinase target (FRT) sites. The Flp-FRT system has been used in conjunction with Cre-loxP in mice16 and in Brainbow-like transgenes in Drosophila24,25. We tested previously described mutant FRT sites39,40 to find incompatible sets (Supplementary Fig. 6) and used these sets to construct ‘Flpbow’ lines (Fig. 4a). In one cassette, we fused the XFPs to an epitope tag41, which allowed for discrimination of cells labeled by Cre- and Flp-driven cassettes (Supplementary Fig. 7). RESOURCE a 2 1 3 pA P Cerulean pA 2 PhiYFP pA N mKate2 pA 1 P Cerulean pA 2 PhiYFP pA N mKate2 pA 2 P 2 3 P 2 N Thy1 P 2 N Cre PhiYFP mKate2 pA N mKate2 pA pA d npg © 2013 Nature America, Inc. All rights reserved. Cerulean PhiYFP mKate2 b c Cre When Flpbow mice were mated to Flp-expressing mice42,43, we observed multicolor labeling (Fig. 4b,c). Although the few lines tested to date exhibit narrow expression patterns, these results demonstrate that Flp- and Cre-based Brainbow systems can be used in combination. Brainbow adeno-associated viral vectors In parallel to developing Brainbow transgenic lines, we generated adeno-associated viral (AAV) vectors to provide spatial and temporal control over expression and to make the method applicable to other species. Because the Brainbow 3.1 cassette described above is >6 kilobase pairs (kb), but the capacity of AAV vectors is <5 kb, we re-engineered the cassette. On the basis of results from initial tests (Supplementary Fig. 8), we devised a scheme in which loxP sites with left or right element mutations44 were used for unidirectional Cre-dependent inversion (Fig. 5a,b). Farnesylated XFPs were positioned in reverse orientation to prevent Figure 3 | Autobow. (a) Schematic of the Autobow construct and the steps that lead to expression of the three XFP combinations (outcomes 1–3). P, loxP site; 2, lox2272; N, loxN; pA, polyadenylation sequence. (b) Labeling of hippocampal neurons by Autobow in founder 3. The 20 large neurons (diameter <5 µm) in this section are labeled in 20 distinct colors. Antibody-amplified Cerulean, PhiYFP and mKate2 are in blue, green and red, respectively. (c,d) Cortical neurons of an Autobow mouse line in the second (c) and sixth (d) generations. Antibody-amplified Cerulean, PhiYFP and mKate2 are as in b. Scale bars, 50 µm. Cre-independent expression, and WPRE elements were added to increase expression. In this design, recombination can lead to three outcomes from two XFPs: XFP1, XFP2 or neither. We generated two AAVs with two XFPs each, such that co-infection would lead to a minimum of eight hues (3 × 3 – 1; Fig. 5c). Because AAV can infect cells at high multiplicity, the number of possible colors is 8. An additional feature is that excision of the nonexpressed XFP in a second step (Fig. 5a) enhances and equalizes expression of the remaining XFP (Fig. 5d,e). We infected cortex, cerebellum and retina of Cre transgenic mice with these vectors. When examined 3–5 weeks later, neurons were labeled in multiple colors (Fig. 5f–j and Supplementary Video 3). Near injection sites, high levels of infectivity led to coexpression of all XFPs in single cells so that neurons appeared gray or white. The variety of colors increased with distance from these sites and then decreased again in sparsely injected regions, presumably because each labeled neuron received only one virion (Supplementary Fig. 9). Methods to optimize Brainbow imaging Obtaining high-quality images from of tissues expressing Brainbow transgenes is challenging. Because colors are derived by mixing images of multiple fluorophores over a wide range of concentrations, factors that differentially affect the labels degrade the final image. In addition, it is usually necessary to image a tissue volume rather than a single section, so methods for taking image stacks must be optimized. Here we summarize guidelines for imaging Brainbow tissue. Sample preparation. To minimize background, section thickness should be less than the working distance of the objective, generally <100 µm for high–numerical aperture (high NA) lenses. It is also important to match the refractive indices of the immersion medium and the sample because chromatic aberrations caused by mismatches between these values lead to spatial offsets between color channels (Supplementary Fig. 10). Commercial antifade nature methods | VOL.10 NO.6 | JUNE 2013 | 543 RESOURCE npg © 2013 Nature America, Inc. All rights reserved. Figure 4 | Flpbow. (a) Schematics of Flpbow transgenes. In Flpbow 1, the loxP sites of Brainbow 1.0 were replaced by the incompatible FRT sites Frt5T2 and Frt545. In Flpbow 3, the loxP sites of Brainbow 3.2 were replaced by incompatible FRT sites F3, Frt5T2 and Frt545. XFPs were fused to SUMO-Star (‘S-’). f indicates farnesylation; pA, polyadenylation sequence; W, woodchuck hepatitis virus post-transcriptional regulatory element. (b) Neurons in caudate putamen of a Flpbow 1 Wnt-Flp double transgenic. Neurons are labeled by at least nine colors (red, orange, yellow, yellow-green, green, cyan, blue, purple and pink). Native fluorescence of ECFP, tdTomato and antibody-amplified PhiYFP are in blue, red and green, respectively. (c) Mossy fibers in a Flpbow3 Wnt-Flp double transgenic. Fluorescence of antibody-amplified EGFP, mOrange2 and mKate2 are in blue, red and green, respectively. Scale bars: 50 µm (b), 20 µm (c). a Flpbow 1 Thy1 5T2 545 tdTomato PhiYFP pA 545 ECFP pA Flpbow 3 Thy1 F3 5T2 545 ØNFPnls b mountants such as Vectashield (Vector Labs) or ProLong Gold (Invitrogen) that have refractive indices of ~1.47 are optimal for objectives that use glycerin as the immersion medium. Polyvinyl alcohol mountants (such as Mowiol 4-88; Sigma-Aldrich) provide a better match for oil-immersion (refractive index of ~1.52) objectives. Confocal laser scanning microscopy. Epifluorescence microscopy can be used for imaging thin sections (<10 µm) or monolayer cultures, but confocal microscopes are preferable for thick specimens because they decrease contamination by light from outside the plane of focus. Newly developed two-photon multi-XFP imaging techniques are also useful45–47. Apochromatic or fluorite microscope objectives that are corrected for three or more colors are strongly recommended. Most lens manufacturers specify preferred oils and coverslips. Using the wrong oil or coverslip degrades the sharpness of focus and increases chromatic aberration. Fluorophores with overlap in the excitation or emission spectra should be imaged sequentially rather than simultaneously to minimize fluorescence cross-talk and thereby optimize color separation. Laser power should be set as low as possible for several reasons. First, all planes are bleached as each image plane is scanned, so generation of stacks leads to gradual bleaching and decreased signal through the stack. Second, because each fluorophore bleaches at a different rate, colors may shift during imaging. Third, linear signaling requires that fluorophores emit photons at submaximal rates; at higher excitation intensities, only the out-of-focus signal is increased48. Fourth, if one fluorophore species is saturated but another is not, small changes in laser power will affect the fluorophore intensities differently, leading to color change. With high-NA objectives, a laser power of just a few milliwatts is saturating; this is generally a small percentage of the total power the laser can provide. With laser power adjusted to a low level, the photomultipliertube voltages and digital gains must be set to relatively high values. In some confocal microscopes it is possible to compensate for signal loss from deep layers through automatic adjustment of laser 544 | VOL.10 NO.6 | JUNE 2013 | nature methods pA 5T2 pA F3 S-mO2f W pA 5T2 S-EGFPf W pA 545 S-mK2f W pA c power, photomultiplier-tube voltage or digital gain as a function of depth. Imaging parameters can be adjusted to obtain images with similar signal ranges throughout the stack. Image processing. Brainbow images must be postprocessed to maximize color information, but care is needed to avoid introducing artifacts. Often one begins by reducing noise. Because confocal laser scanning of multicolor stacks is generally done at speeds of ~1 µs per pixel or less to save time, the small number of photons collected for each pixel gives rise to sufficient shot noise to cause perceptible local color differences. This problem can be minimized by slower scanning or averaging of multiple scans, but when this is infeasible, simple filtering and deconvolution methods are helpful (Fig. 6a–c). For example, median or Gaussian filters with radii of 0.5–2 pixels reduce color noise, but at the expense of resolution. Deconvolution algorithms (Online Methods) are more challenging to use than simple filters but can remove color noise without compromising spatial resolution (Supplementary Fig. 11). Subsequent processing steps can expand the detectable color range and correct for color shifts (Fig. 6d,e). To obtain easily perceived color differences, pixel intensity values for each channel in each image are normalized to the same minimum and maximum intensity values for that color in the whole image stack. This linearly stretches all channels and images to the full dynamic range. Color shifts also arise because illumination strength is generally uneven across the imaging field and differs among lasers. This effect can be attenuated by intensity or shading correction for each channel46,49. The resulting composite RGB images provide maximum color separation for viewing by eye (Fig. 6f). DISCUSSION The goal of the work reported here was to design, generate and characterize improved reagents for multicolor Brainbow imaging of neurons in mice. First, we generated new transgenic lines that overcome some limitations of the Brainbow 1 and 2 lines that are currently available15. The improvements were the substitution of XFPs (especially red and orange fluorescent proteins) AAV no. 2 5′ITR EF1 W pA 3′ITR W pA 3′ITR 1 1 2 mChef mTFPf mTFPf mChef 3 4 mTFPf mTFPf npg EF1 mChef 5′ITR mTFPf AAV no. 1 TagBFPf a Figure 5 | Brainbow AAV. (a) AAV Brainbow constructs and recombination scheme. Farnesylated TagBFP and EYFP or mCherry and mTFP were placed in reverse orientation between mutant loxP sites. EF1α, regulatory elements from elongation factor 1α gene; W, woodchuck hepatitis virus post-transcriptional regulatory element; pA, polyadenylation sequence; mChef, mCherryf; triangles, loxP mutants and their recombination products (dark and light green sectors indicate wild-type and mutant portions of loxP sites, respectively); 1–6, recombination events. (b) Outcomes of recombination events numbered in a. Unfilled arrows, direction of intervening cDNA. An X signifies that a fully mutant (light green) loxP site cannot serve as substrate for Cre. (c) Eight color outcomes resulting from pairs of Brainbow AAVs following recombination as shown in a. This is a minimum value because it does not account for differences in relative intensity of the four XFPs. (d,e) Test of color balance. The mTFPf-mChef AAV vector was injected at low titer into the cortex of a Thy1-Cre mouse, and neurons of each color were counted. An image from the cortex (d) and the fraction of neurons expressing mTFPf (58%) and mChef (42%; n = 1,523 neurons in four sections of three mice; s.d. in red) are shown. (f–h) Cortex of parvalbumin (PV)-Cre mice injected with the two Brainbow AAVs. High-magnification views of adjacent planes from the yellow (g) and white (h) boxed regions in f are shown. (i) Retina from AAV-injected mouse expressing Cre in retinal ganglion cells. Arrows, dendrites; arrowheads, axons. (j) Cerebellum from AAV-injected PV-Cre mouse. Fan-shaped dendrites of Purkinje cells (white arrows) and interneurons (yellow arrows) are indicated. In f–j, antibody-amplified mTFP and EYFP are in green, TagBFP is in blue and mCherry is in red. Scale bars: 100 µm (d), 20 µm (f), 10 µm (g,h), 50 µm (i,j). EYFPf mChef 5 b 6 c AAV no. 1 AAV no. 2 Output 6 f 1 2 3 4 5 6 d e 70 % of total neurons © 2013 Nature America, Inc. All rights reserved. RESOURCE g 60 h 50 40 30 20 10 0 + mTFPf mChef+ j i that are more photostable and less prone to aggregation than those used initially; use of XFPs with minimal sequence homology so they could be immunostained separately; farnesylation of the XFPs for even staining of somata and the finest processes; insertion of a stop cassette to increase color variety by eliminating broad expression a b c f e d of a default XFP; inclusion of a nonfluorescent marker in the default position to facilitate screening of multiple lines; and insertion of a WPRE to boost expression (Figs. 1 and 2). These lines, which incorporate regulatory elements from the Thy1 gene, enable marking of many but not all neuronal types. To date, elements tested other than those from the Thy1 gene do not support the high expression levels needed to image Brainbow 1 and 2 material. The ability to immunostain provided by Brainbow 3 cassettes may allow weaker promoters to be used. Second, we designed two additional transgenes and performed initial tests to demonstrate that they can be used effectively in vivo. Figure 6 | Processing a Brainbow image. (a) Original image from parvalbuminCre mouse cortex injected with mixed Brainbow AAVs. (b) Region boxed in a. (c) Deconvolution, to decrease noise without sacrificing spatial resolution. (d) Intensity normalization, to expand perceptible color range. (e) Colorshift correction (Supplementary Fig. 10). (f) Fully processed image. Yellow arrows indicate the sequence of image processing. White arrowheads indicate corresponding objects in the original and color shift–corrected images. Scale bars, 10 µm (a,f), 3 µm (b–e). nature methods | VOL.10 NO.6 | JUNE 2013 | 545 npg © 2013 Nature America, Inc. All rights reserved. RESOURCE One construct, called Autobow, incorporates a self-excising Cre recombinase. Autobow lacks the temporal and spatial control afforded by the use of specific Cre lines or ligand-activated Cre (CreER). However, because it does not require the generation of double transgenics, it may be useful for rapid screening of neuronal morphology in mutant mice or mice submitted to various experimental interventions (such as drug treatments). The other novel transgene, Flpbow, replaces loxP sites with FRT sites so that recombination can be controlled by Flp recombinase rather than Cre recombinase. Flpbow 3 also incorporates an epitope tag so that XFPs in Flpbow can be distinguished immunohistochemically from XFPs in Brainbow. By using Cre and Flp transgenic lines with distinct, defined specificities, it should be possible to map separate sets of neurons in a single animal. Finally, we generated Brainbow AAV vectors. These, along with recently described Brainbow herpes viral vectors50, may be more useful than Brainbow transgenic mice in some situations. Similar to Autobow, they avoid the need for double-transgenic animals. Because the time of infection can be varied, these vectors provide an alternative to CreER for temporal control. Moreover, localized delivery of AAVs enables the tracing of connections from known sites to multiple targets and discrimination of longdistance inputs from local connections. The three most broadly useful Brainbow 3 lines (Brainbow 3.0 line D, Brainbow 3.1 line 3, Brainbow 3.1 line 18 and Brainbow 3.2 line 7) have been provided to Jackson Laboratories (http://www. jax.org/; stock numbers 21225–21227) for distribution. The two AAVs shown in Figure 5 can be obtained from the University of Pennsylvania Vector Core (http://www.med.upenn.edu/gtp/ vectorcore/). Plasmids used to generate Brainbow 3.0, 3.1, and 3.2, Autobow and Flpbow 1.1 and 3.1 mice are available through Addgene (http://www.addgene.org/). Methods Methods and any associated references are available in the online version of the paper. Accession codes. Addgene plasmids: 45176, 45177 (Brainbow 3.0), 45178 (Brainbow 3.1), 45179 (Brainbow 3.2), 45182, 45187 (Autobow), 45180 (Flpbow 1.1), 45181 (Flpbow 3.1). Note: Supplementary information is available in the online version of the paper. Acknowledgments This work was supported by grants from the US National Institutes of Health (5U24NS063931) and the Gatsby Charitable Foundation and by Collaborative Innovation Award no. 43667 from the Howard Hughes Medical Institute. We thank S. Haddad for assistance with mouse colony maintenance; X. Duan, L. Bogart and J. Lefebvre for testing Brainbow mice and AAVs; R.W. Draft for valuable discussions and advice; R.Y. Tsien (University of California, San Diego) for mOrange2 and TagRFPt; and D.M. Chudakov (Institute of Bioorganic Chemistry of the Russian Academy of Sciences) for TagBFP, PhiYFP, mKate2 and eqFP650. AUTHOR CONTRIBUTIONS D.C., K.B.C. and T.L. performed experiments. D.C., J.W.L. and J.R.S. designed experiments, interpreted results and wrote the manuscript. COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests. Reprints and permissions information is available online at http://www.nature. com/reprints/index.html. 546 | VOL.10 NO.6 | JUNE 2013 | nature methods 1. Tsien, R.Y. The green fluorescent protein. Annu. Rev. Biochem. 67, 509–544 (1998). 2. Heim, R. & Tsien, R.Y. Engineering green fluorescent protein for improved brightness, longer wavelengths and fluorescence resonance energy transfer. Curr. Biol. 6, 178–182 (1996). 3. Rizzo, M.A., Springer, G.H., Granada, B. & Piston, D.W. An improved cyan fluorescent protein variant useful for FRET. Nat. Biotechnol. 22, 445–449 (2004). 4. Shagin, D.A. et al. GFP-like proteins as ubiquitous metazoan superfamily: evolution of functional features and structural complexity. Mol. Biol. Evol. 21, 841–850 (2004). 5. Ai, H.W., Henderson, J.N., Remington, S.J. & Campbell, R.E. Directed evolution of a monomeric, bright and photostable version of Clavularia cyan fluorescent protein: structural characterization and applications in fluorescence imaging. Biochem. J. 400, 531–540 (2006). 6. Subach, O.M. et al. Conversion of red fluorescent protein into a bright blue probe. Chem. Biol. 15, 1116–1124 (2008). 7. Ai, H.W., Shaner, N.C., Cheng, Z., Tsien, R.Y. & Campbell, R.E. Exploration of new chromophore structures leads to the identification of improved blue fluorescent proteins. Biochemistry 46, 5904–5910 (2007). 8. Karasawa, S., Araki, T., Nagai, T., Mizuno, H. & Miyawaki, A. Cyanemitting and orange-emitting fluorescent proteins as a donor/acceptor pair for fluorescence resonance energy transfer. Biochem. J. 381, 307–312 (2004). 9. Shaner, N.C. et al. Improving the photostability of bright monomeric orange and red fluorescent proteins. Nat. Methods 5, 545–551 (2008). 10. Shaner, N.C. et al. Improved monomeric red, orange and yellow fluorescent proteins derived from Discosoma sp. red fluorescent protein. Nat. Biotechnol. 22, 1567–1572 (2004). 11. Shcherbo, D. et al. Far-red fluorescent tags for protein imaging in living tissues. Biochem. J. 418, 567–574 (2009). 12. Shcherbo, D. et al. Near-infrared fluorescent proteins. Nat. Methods 7, 827–829 (2010). 13. Shaner, N.C., Steinbach, P.A. & Tsien, R.Y. A guide to choosing fluorescent proteins. Nat. Methods 2, 905–909 (2005). 14. Feng, G. et al. Imaging neuronal subsets in transgenic mice expressing multiple spectral variants of GFP. Neuron 28, 41–51 (2000). 15. Livet, J. et al. Transgenic strategies for combinatorial expression of fluorescent proteins in the nervous system. Nature 450, 56–62 (2007). 16. Branda, C.S. & Dymecki, S.M. Talking about a revolution: the impact of site-specific recombinases on genetic analyses in mice. Dev. Cell 6, 7–28 (2004). 17. Snippert, H.J. et al. Intestinal crypt homeostasis results from neutral competition between symmetrically dividing Lgr5 stem cells. Cell 143, 134–144 (2010). 18. Red-Horse, K., Ueno, H., Weissman, I.L. & Krasnow, M.A. Coronary arteries form by developmental reprogramming of venous cells. Nature 464, 549–553 (2010). 19. Rinkevich, Y., Lindau, P., Ueno, H., Longaker, M.T. & Weissman, I.L. Germ-layer and lineage-restricted stem/progenitors regenerate the mouse digit tip. Nature 476, 409–413 (2011). 20. Schepers, A.G. et al. Lineage tracing reveals Lgr5+ stem cell activity in mouse intestinal adenomas. Science 337, 730–735 (2012). 21. Tabansky, I. et al. Developmental bias in cleavage-stage mouse blastomeres. Curr. Biol. 23, 21–31 (2013). 22. Gupta, V. & Poss, K.D. Clonally dominant cardiomyocytes direct heart morphogenesis. Nature 484, 479–484 (2012). 23. Pan, Y.A., Livet, J., Sanes, J.R., Lichtman, J.W. & Schier, A.F. Multicolor Brainbow imaging in zebrafish. Cold Spring Harb. Protoc. 2011, pdb.prot5546 (2011). 24. Hampel, S. et al. Drosophila Brainbow: a recombinase-based fluorescence labeling technique to subdivide neural expression patterns. Nat. Methods 8, 253–259 (2011). 25. Hadjieconomou, D. et al. Flybow: genetic multicolor cell labeling for neural circuit analysis in Drosophila melanogaster. Nat. Methods 8, 260–266 (2011). 26. Lang, C., Guo, X., Kerschensteiner, M. & Bareyre, F.M. Single collateral reconstructions reveal distinct phases of corticospinal remodeling after spinal cord injury. PLoS ONE 7, e30461 (2012). 27. Badaloni, A. et al. Transgenic mice expressing a dual, CRE-inducible reporter for the analysis of axon guidance and synaptogenesis. Genesis 45, 405–412 (2007). 28. Lichtman, J.W., Livet, J. & Sanes, J.R. A technicolour approach to the connectome. Nat. Rev. Neurosci. 9, 417–422 (2008). 29. Lakso, M. et al. Targeted oncogene activation by site-specific recombination in transgenic mice. Proc. Natl. Acad. Sci. USA 89, 6232–6236 (1992). 30. Paterna, J.C., Moccetti, T., Mura, A., Feldon, J. & Büeler, H. Influence of promoter and WHV post-transcriptional regulatory element on AAVmediated transgene expression in the rat brain. Gene Ther. 7, 1304–1311 (2000). 31. Madisen, L. et al. A robust and high-throughput Cre reporting and characterization system for the whole mouse brain. Nat. Neurosci. 13, 133–140 (2010). 32. Hippenmeyer, S. et al. A developmental switch in the response of DRG neurons to ETS transcription factor signaling. PLoS Biol. 3, e159 (2005). 33. Srinivas, S. et al. Cre reporter strains produced by targeted insertion of EYFP and ECFP into the ROSA26 locus. BMC Dev. Biol. 1, 4 (2001). 34. Guo, C., Yang, W. & Lobe, C.G. A Cre recombinase transgene with mosaic, widespread tamoxifen-inducible action. Genesis 32, 8–18 (2002). 35. Zhang, X.M. et al. Highly restricted expression of Cre recombinase in cerebellar Purkinje cells. Genesis 40, 45–51 (2004). 36. Rossi, J. et al. Melanocortin-4 receptors expressed by cholinergic neurons regulate energy balance and glucose homeostasis. Cell Metab. 13, 195–204 (2011). 37. Campsall, K.D., Mazerolle, C.J., De Repentingy, Y., Kothary, R. & Wallace, V.A. Characterization of transgene expression and Cre recombinase activity in a panel of Thy-1 promoter-Cre transgenic mice. Dev. Dyn. 224, 135–143 (2002). 38. Bunting, M., Bernstein, K.E., Greer, J.M., Capecchi, M.R. & Thomas, K.R. Targeting genes for self-excision in the germ line. Genes Dev. 13, 1524–1528 (1999). 39. McLeod, M., Craft, S. & Broach, J.R. Identification of the crossover site during FLP-mediated recombination in the Saccharomyces cerevisiae plasmid 2 µm circle. Mol. Cell Biol. 6, 3357–3367 (1986). 40. Schlake, T. & Bode, J. Use of mutated FLP recognition target (FRT) sites for the exchange of expression cassettes at defined chromosomal loci. Biochemistry 33, 12746–12751 (1994). 41. Peroutka, R.J., Elshourbagy, N., Piech, T. & Butt, T.R. Enhanced protein expression in mammalian cells using engineered SUMO fusions: secreted phospholipase A2. Protein Sci. 17, 1586–1595 (2008). 42. Farley, F.W., Soriano, P., Steffen, L.S. & Dymecki, S.M. Widespread recombinase expression using FLPeR (flipper) mice. Genesis 28, 106–110 (2000). 43. Awatramani, R., Soriano, P., Rodriguez, C., Mai, J.J. & Dymecki, S.M. Cryptic boundaries in roof plate and choroid plexus identified by intersectional gene activation. Nat. Genet. 35, 70–75 (2003). 44. Araki, K., Okada, Y., Araki, M. & Yamamura, K. Comparative analysis of right element mutant loxP sites on recombination efficiency in embryonic stem cells. BMC Biotechnol. 10, 29 (2010). 45. Entenberg, D. et al. Setup and use of a two-laser multiphoton microscope for multichannel intravital fluorescence imaging. Nat. Protoc. 6, 1500–1520 (2011). 46. Mahou, P. et al. Multicolor two-photon tissue imaging by wavelength mixing. Nat. Methods 9, 815–818 (2012). 47. Wang, K. et al. Three-color femtosecond source for simultaneous excitation of three fluorescent proteins in two-photon fluorescence microscopy. Biomed. Opt. Express 3, 1972–1977 (2012). 48. Conchello, J.A. & Lichtman, J.W. Optical sectioning microscopy. Nat. Methods 2, 920–931 (2005). 49. Ducros, M. et al. Efficient large core fiber-based detection for multichannel two-photon fluorescence microscopy and spectral unmixing. J. Neurosci. Methods 198, 172–180 (2011). 50. Card, J.P. et al. A dual infection pseudorabies virus conditional reporter approach to identify projections to collateralized neurons in complex neural circuits. PLoS ONE 6, e21141 (2011). npg © 2013 Nature America, Inc. All rights reserved. RESOURCE nature methods | VOL.10 NO.6 | JUNE 2013 | 547 © 2013 Nature America, Inc. All rights reserved. npg ONLINE METHODS Brainbow constructs. cDNA encoding the following fluorescent proteins were used: EGFP1, EYFP2, ECFP2, mCerulean3, PhiYFP4 (Evrogen), mTFP5 (Allele Biotechnology), TagBFP6 (Evrogen), EBFP2 (ref. 7) (Addgene), Kusabira-Orange8, TagRFPt9, mOrange2 (ref. 9), tdTomato10, mCherry10, mKate2 (ref. 11) (Evrogen) and eqFP650 (ref. 12) (Evrogen). A HRAS farnesylation sequence51 was used to tether XFPs to the cell membrane. A nuclear localization signal (NLS, APKKKRKV) was added to the N-terminal end of PhiYFP(Y65A). WPRE sequence was used to stabilize mRNA and enhance nuclear mRNA export30,31. Polyadenylation signals were from the SV40 T antigen for mouse transgenes and from the human growth hormone gene for AAV. Brainbow constructs were assembled by standard cloning methods. A cloning scaffold containing concatenated loxP mutant sequences and unique restriction digestion sites was synthesized (DNA2.0, Inc.) to facilitate cloning. Brainbow modules were cloned into the pCMV- N1 mammalian expression vector (Clontech) for transient mammalian cell expression. Brainbow mouse constructs were cloned into a unique XhoI site in a genomic fragment of Thy1.2 containing neuronspecific regulatory elements14. Brainbow AAV constructs were cloned into vectors provided by the University of Pennsylvania Virus Core. Constructs were tested by expression in HEK293 cells (ATCC) before generation of mice or AAV. Mice. Transgenic mice were generated by pronuclear injection at the Harvard Genome Modification Core. Mice were maintained on C57B6 or CD-1 backgrounds. Brainbow mice were crossed to mice that expressed Cre or Flp recombinases, including PV-Cre32, Islet-Cre33, CAGGS-CreER34, L7-Cre35, ChAT-Cre36, Thy1-Cre37, Wnt-Flp43 and Actin-Flp42. Both male and female mice were used. All experiments conformed to NIH guidelines and were carried out in accordance with protocols approved by the Harvard University Standing Committee on the Use of Animals in Research and Teaching. AAV injection. Two Brainbow AAVs were mixed to equal titer (7.5 × 1012 genome copies per ml) before injection. For retina injection, adult mice were anesthetized with ketamine-xylazine by intraperitoneal injection. A small hole was made in the temporal eye by puncturing the sclera below the cornea with a 30 1/2–G needle. With a Hamilton syringe with a 33-G blunt-ended needle, 0.5–1 µl of AAV virus was injected intravitreally. After injections, animals were treated with Antisedan (Zoetis) and monitored for full recovery. For cortex injection, adult mice were anesthetized with isoflurane via continuous delivery through a nose cone and fixed to a stereotaxic apparatus. Surgery took place under sterile conditions with the animal lying on a heating pad. One microliter of 1:5 saline-diluted AAV mix (1.5 × 1012 genome copies per ml) was injected over 10 min. The head wound was sutured at the end of the experiment. One injection of the nonsteroidal antiinflammatory agent meloxicam was given at the end of the surgery, and mice were kept on a heating pad with accessible moistened food pellets and/or HydroGel until they had fully recovered. The mice were given another dose of meloxicam 1 d later and examined 4–6 weeks after infection. Antibodies. Expression vectors were constructed to produce His tag fusions of XFPs in Escherichia coli. Proteins were produced nature methods in bacteria, purified using His Fusion Protein Purification Kits (Thermo Scientific), concentrated to >3 mg/ml, and used as immunogens to produce rat anti-mTFP, chicken anti-EGFP, rat anti-PhiYFP, rabbit anti-mCherry and guinea pig anti-mKate2. (Covance). Chicken anti-EGFP IgY was purified from chicken egg yolks using Pierce Chicken IgY Purification Kit (Thermo Scientific). Other sera were used without purification. Other antibodies used were: rabbit anti-GFP (ab6556, Abcam), rabbit anti-PhiYFP (AB604, Evrogen), chicken anti-SUMOstar (AB7002, LifeSensors), DyLight 405–conjugated goat anti-rat (Jackson ImmunoResearch) and Alexa fluorescent dye–conjugated goat anti-rat 488, anti-chicken 488 and 594, anti-rabbit 514 and 546, and anti–guinea pig 647 (Life Technologies). Histology. Mice were anesthetized with sodium pentobarbital before intracardiac perfusion with 2%–4% paraformaldehyde in PBS. Brains were sectioned at 100 µm using a Leica vt1000s vibratome. Muscle and retina were sectioned at 20 µm in a Leica CM1850 cryostat or processed as whole mounts. For immunostaining, tissues were permeabilized by 0.5% Triton X-100 with 0.02% sodium azide in StartingBloc (Thermo Scientific) at room temperature for 2 h and then incubated with combinations of anti-XFPs (see above) for 24–48 h at 4 °C. After extensive washing in PBST (0.01 M PBS with 0.1% Triton X-100), all secondary antibodies (1:500) were added for 12 h at 4 °C. Finally, sections and tissues were mounted in Vectashield mounting medium (Vector Labs) and stored at –20 °C until they were imaged. Antibody combinations used in figures are as follows. In Figures 1e–h; 2b,d–g; and 4c and Supplementary Figures 2a–d, 4a and 5, primary antibodies are chicken anti-GFP (1:2,000), rabbit anti-mCherry (1:1,000, for mOrange2) and guinea pig anti-mKate2 (1:500). Secondary antibodies are Alexa dye– conjugated goat anti-chicken 488, anti-rabbit 546, and anti–guinea pig 647. In Figure 3b–d and Supplementary Figure 2e, primary antibodies are chicken anti-GFP (for ECFP), rabbit anti-PhiYFP (1:1,000) and guinea pig anti-mKate2. Secondary antibodies are Alexa dye–conjugated goat anti-chicken 488, anti-rabbit 546 and anti-guinea pig 647. In Figure 4b, rabbit anti-PhiYFP and Alexa dye–conjugated goat anti-rabbit 514 were used. In Figure 5d, primary antibodies are rat anti-mTFP and rabbit anti-mCherry. Secondary antibodies are Alexa dye–conjugated goat anti-rat 488 and anti-rabbit 546. In Figure 5f–j and Supplementary Figure 9, primary antibodies are guinea pig anti-mKate2 (for TagBFP), rat anti-mTFP (1:1,000), chicken anti-GFP (for EYFP) and rabbit anti-mCherry. Secondary antibodies are Alexa dye–conjugated goat anti-rat 488, anti-chicken 488, anti-rabbit5 46 and antiguinea pig 647. In Supplementary Figure 1, primary antibodies are rat anti-PhiYFP (1:1,000) and rabbit anti-mCherry (1:1,000, for mOrange2). Secondary antibodies are Alexa dye–conjugated goat anti-rat 488 and anti-rabbit 546. In Supplementary Figure 7, primary antibodies are rabbit anti-EGFP (1:1,000, for ECFP) and chicken anti-SUMOstar (1:1,000). Secondary antibodies are Alexa dye–conjugated goat anti-rabbi 514 and anti-chicken 594. Imaging. Fixed brain and muscle samples were imaged using a Zeiss LSM710 confocal microscope. Best separation of multiple fluorophores was obtained by using a 405-nm photodiode laser for TagBFP and DyLight 405, a 440-nm photodiode laser for mTFP, a 488-nm Argon line for EGFP and Alexa 488, a 514-nm doi:10.1038/nmeth.2450 collected at 545–600 nm in channel 2. In the antibody-amplified samples, conjugated Alexa dyes normally produced much stronger fluorescence signal than XFPs. The Zeiss microscope we used was optimized for imaging the Alexa 488/546/647 combination. The fixed dichroic mirror was DM488/561/633. Alexa 488 and Alexa 647 were excited by 488-nm and 633-nm lasers simultaneously, and fluorescence was collected at 495–590 nm in channel 1 and 638–780 nm in channel 2, respectively. In the subsequent scan, a 561-nm laser was used to excite mOrange2, and fluorescence was collected at 566–626 nm in channel 2. 51. Hancock, J.F., Cadwallader, K., Paterson, H. & Marshall, C.J.A. CAAX or a CAAL motif and a second signal are sufficient for plasma membrane targeting of ras proteins. EMBO J. 10, 4033–4039 (1991). npg © 2013 Nature America, Inc. All rights reserved. Argon line for EYFP and Alexa 514, a 561-nm photodiode for mOrange2 and Alexa 546, a 594-nm photodiode for mCherry, mKate2 and Alexa 594 or a 633-nm photodiode for Alexa 647. Images were obtained with 16× (0.8 NA), and 63× (1.45 NA) oil objectives. Confocal image stacks for all channels were acquired sequentially, and maximally or 3D-view projected using ImageJ (NIH). Intensity levels were uniformly adjusted in ImageJ. Optimal imaging for Brainbow 3 tissue used a Zeiss LSM710 with fixed dichroic mirror combinations of DM455+514/594 to reduce lag time between the two sequential scans. EGFP and mKate2 were excited by 458-nm and 594-nm lasers simultaneously, and fluorescence was collected at 465–580 nm in channel 1 and 605–780 nm in channel 2, respectively. In a subsequent scan, a 514-nm laser was used to excite mOrange2, and fluorescence was doi:10.1038/nmeth.2450 nature methods NEW TOOLS FOR THE BRAINBOW TOOLBOX Dawen Cai, Kimberly B. Cohen, Tuanlian Luo, Jeff W. Lichtman and Joshua R. Sanes Supplementary Figures and Tables Contents: • • Supplementary Figures 1 – 11 Supplementary Tables 1 – 2 Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 1. Use of the non-fluorescent default protein to analyze expression potential of a Brainbow3.1 mouse line. Sagittal section of brain from Brainbow 3.1 (line 3), stained with rat anti-PhiYFP. (a) Low power image. C, cerebellum; H, Hippocampus. (b, c) Higher magnification view of cerebellar (C) and hippocampal (H) regions boxed in a. Bars are 1mm in a, 200μm in b, 20μm in c. Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 2. Neuronal labeling in Brainbow3;cre double transgenic and Autobow mice. (a) Whole mount of retina from Islet1-Cre;Brainbow 3.1 (line 10) showing retinal ganglion cells. (b,c) Cerebellum from Brainbow 3.1 (line 3); parvalbumin-cre mouse. Low power image in b shows Purkinje (yellow arrows) and basket cells (cyan arrow). Axon terminals of basket cells arborize on initial segments of Purkinje cell axons. Higher power view in c shows axons from multiple basket cells, labeled in distinct colors, wrapping a Purkinje cell body and axonal initial segment (yellow arrow). (d) Brainstem from PV-Cre;Brainbow 3.1 (line 4) showing axonal terminals of inhibitory neurons. In a-d, antibody amplified EGFP, mOrange2 and mKate2 are in blue, green and red, respectively. (e) Labeling of hippocampal neurons by Autobow (line 2). Antibody amplified Cerulean, PhiYFP and mKate2 are in blue, green and red, respectively. Bars are 30μm in a, 20μm in b, 10μm in c and d, 100μm in e. Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 3. Farnesylated XFPs label neurons more evenly than cytoplasmic XFPs. Retinal bipolar cells are labeled by (a) cytoplasmic GFP and (b) farnesylated membrane GFP. To image processes of neurons labeled with cytoplasmic GFP, it is necessary to saturate the somata. Bars are 5µm. Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 4. Immunostaining enhances Brainbow signals without compromising color diversity. Motor axons innervating extraocular muscles of a Brainbow 3.0 mouse (line D). a shows native XFP fluorescence. b shows a nearby section immunostained with non-crossreactive chicken-anti-GFP, rabbit-anti-mOrange2 and guinea pig-anti-mKate2. Bars are 50μm. Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 5. XFP expression balance in Brainbow 3. Immunostained cerebellum section from Brainbow 3.1 (line 3);PV-cre mouse. Merged image is shown in a and separate channels in b (EGFP), c (mOrange2) and d (mKate2). (e) Percent of Purkinje cells expressing each XFP (208 neurons from 12 regions). Bars are 20μm. Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 6. Tests for incompatible mutant Frt sites. Canonical Frt site and multiple Frt variants have been reported to allow Flp-dependent recombination between the same variants but not between each other17, 18. To verify incompatibility, three constructs were generated and tested: (a) Construct testing incompatibility between Frt5T2 and Frt545 sites. When co-transfected with this construct and FLP recombinase, HEK293 cells showed high level of Kusabira-Orange but no detectable mKate2. This indicated that Frt5T2 and Frt545 are incompatible. (b) Adding FrtF3 site and TagBFP to the construct in a allowed a three-way test of FrtF3, Frt5T2 and Frt545. Cotransfection of this construct with Flp recombinase showed TagBFP expression only, indicating that the FrtF3 site is incompatible with both Frt5T2 and Frt545. (c) Adding the canonical Frt site and TagBFP to the construct in a allowed a three-way test of Frt, Frt5T2 and Frt545. Co-transfection of this construct with Flp recombinase showed high level of TagBFP expression as well as detectable levels of Kusabira-Orange and mkate2. This indicates that canonical Frt site undergoes Flp dependent excision with Frt5T2 and Frt545 sites. Bars are 100μm. Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 7. SUMOstar as epitope tag in transgenic mouse. Section from the hippocampus from a transgenic mouse that expressed SUMOstar-ECFP fusion protein from the Thy1 promoter. The section was stained with rabbit anti-GFP and chicken anti-SUMOstar, plus appropriate secondary antibodies. (a) Native fluorescence of ECFP. (b) Rabbit anti-GFP (c) Chicken anti-SUMOstar. Bar is 50μm. Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 8. Tests of Brainbow AAV constructs. In order to accommodate the 5kB packaging size of AAV yet achieve high expression level, several promoter elements, polyadenylation sequences and stop mechanisms were tested in constructs based on Brainbow1. (a) This series tested the performance of a 1kB hybrid promoter CB6 (CMV enhancer, chicken β-actin promoter and hybrid intron, Penn Vector Core) as well as a small stop cassette relying on an open reading frame shift mechanism (“ORF shift” round corner empty box). The ORF shift encodes a Kozak sequence and translational start (red solid arrow) to initiate an “out of frame” translation of the following mTFP. Upon Cre recombination, this stop cassette is removed and mTFP, PhiYFP or TagRFPt is expressed. AAV made from this construct displayed cre-independent expression of XFPs in wildtype mouse brains. Moreover, expression from the CB6 promoter was weaker than that from a EF1α promoter, tested in parallel. (b) This series used a pair LoxP left and right element mutants (LE-RE) to construct a Cre dependent “unidirectional spin” cassette. A Brainbow1 construct expressing Kusabira-Orange, EGFP and mCherry was placed in inverted orientation between the LE-RE loxP sites. AAV generated from this construct exhibited no Creindependent expression in mouse brain. However, the proportions of the three XFPs expressed following infection of Cre transgenic mice was greatly influenced by the level of Cre expression. When Cre levels were low, Kusabira-Orange was expressed in most cells. In contrast, when Cre levels were high, very little expression of Kusabira-Orange was detected, with most cells expressing mCherry and a minority expressing EGFP. This imbalance led us to use the LE-RE elements exclusively (Fig. 2). (c) This series incorporated a STOP cassette generated from a tandem Myc epitope tag and a bovine growth hormone polyadenylation sequence, totaling 0.6kB. To accommodate the packaging size of AAV, only two XFPs were incorporated into each construct. Infection with pairs of AAVs led to multicolor labeling (see d). As in constructs shown in b, however, relative expression of the two XFPs in each vector varied among Cre lines. Based on these tests, we designed new AAVs (Fig. 5a) incorporating EF1α promoter, LE-RE mutant sites, farnesylated XFPs, a single WPRE sequence, and a 0.5kB human growth hormone polyadenylation sequence. (d) Brain from a Thy1-cre mice infected with the two AAV shown in c. Native fluorescence of TagBFP in blue, Kusabira-Orange and EGFP in green, mCherry in red. Bar is 50μm. Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 9. Labeling color pallet changes at different distances to Brainbow AAV injection site. (a) Sagittal section of PV-Cre mouse brain shows fluorescent labeling by Brainbow AAV infection. White arrow indicates injection site. (b) Magnified view of blue boxed regions in a. (c, d) Magnified views of regions boxed in b. Bars are 300μm in a, 50μm in b and 20μm in c, and d. Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 10. Chromatic aberration in Brainbow images. (a,b) X-Y plane (a) and cross sections (b) show axial and lateral chromatic aberration, as well as laser misalignment cause color offset in X-Y-Z directions. (c,d) Color-shift correction of images in a,b align the channels and restore correct colors. Arrowheads indicate corresponding objects in the original and corrected images. Bars are 3µm. Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 11. Noise reduction methods comparison. (a) Original image. (b) Enlarged view of region boxed in a. (c) Guassian filter with radius of 1.0 pixel. (d) Median filter with radius of 1.0 pixel. (e) Deconvolution by Huygens software. Bars are 10µm in a, 2µm in b-e. Nature Methods: doi:10.1038/nmeth.2450 Supplementary Table 1. Fluorescent proteins screened in transgenic mice. EBFP2 TagBFP ECFP Cerulean mTFP1 EGFP EYFP PhiYFP * 383 399 433 433 462 488 516 525 448 456 475 475 492 507 529 537 Brightness (EGFP=1) 0.55 1.00 0.33 0.82 1.63 1.00 1.48 1.58 Kusabira Orange 548 559 0.94 Bright; Aggregation Fugia concinna Rb : KuO mOrange2 tdTomato tagRFP-T mCherry mKate2 eqFP650 549 554 555 587 588 592 565 581 584 619 635 650 1.06 2.88 1.00 0.48 0.76 0.48 Bright; Aggregation Bright Bright Dim; Aggregation Bright Dim; Far-red Discosoma sp. Discosoma sp. Entacmaea quadricolor Discosoma sp. Entacmaea quadricolor Entacmaea quadricolor Rb : DsRed Rb : DsRed GP : mKate2 Rb : DsRed GP : mKate2 GP : mKate2 Fluorescent Protein Ex (nm) Em (nm) in vivo property species Antibody Dim; Scattering Bright; Scattering Dim; High Expression Bright; Photobleaching Bright Bright; High Expression Bright; High Expression Bright; PFA Quenching Aequorea victoria Entacmaea quadricolor Aequorea victoria Aequorea victoria Clavularia sp. Aequorea victoria Aequorea victoria Phialidium sp. Ck : GFP GP : mKate2 Ck : GFP Ck : GFP Rt : TFP Ck : GFP Ck : GFP Rt : PhiYFP Columns are as follows: (1) Proteins tested; see refs. 2-12. (2) Ex, excitation peak. (3) Em, emission peak. (4) Brightness calculated as the product of extinction coefficient and quantum yield, with EGFP=1. (5) Global result from observation of brain tissue in transgenic mice. Main drawbacks encountered were low fluorescence in vivo despite high intrinsic brightness, presumable because of low expression; high scattering of deep blue proteins; rapid photobleaching; aggregation in neurons; and sensitivity to paraformaldehyde fixation. (6) Species from which XFP was derived. (7) Antibodies used to amplify signal: Ck:GFP, chicken-anti-GFP; GP:mKate2, Guinea pig-anti-mKate2; Rt:TFP, rat-anti-mTFP1; Rt:PhiYFP, rat-anti-PhiYFP; Rb:KuO, rabbit-anti-Kusabira Orange; Rb:DsRed, rabbit-anti-DsRed. * The nonfluorescent mutant PhiYFP(Y65A) was used as stainable “STOP”. Nature Methods: doi:10.1038/nmeth.2450 Supplementary Table 2. Brainbow 3 transgenic mouse lines. Cell type labeled (Brightness*) Expression # line subset size 1st/2nd/3rd/4th Spinal cord / Peripheral Central Nervous System A medium not tested RGC(+), CTL(++), HPL(++), CBR(+) B large not labeled whole brain (+) Brainbow-3.0 MN (+++), SIN (++), Of/Gf/Kf D large RGC (++), RBP (++), CTL(+++), HPL(+++) DRG (++), SSN (++) F medium MN(++), DRG(+++) CTL(++), HPL(++) 2 large not tested whole brain (+) 3 medium MN (++) CTL (++), HPL (+++), CBR (++), CPC (+++) 4 large MN (+), DRG (++) CTL(+), HPL(+++), CBR (++), HB (+) 10 small not labeled RGC (++), RAC (++), CTL (+), HPL (++) 11 small MN (+), DRG (++) RGC (++), RAC (++), CTL(++), HPL(+++) Brainbow-3.1 16 small not labeled CP (++), TH (++) nØ/Of/Gf/Kf 17 very small not labeled CTL(+++), HPL(+++), CP (+++) RGC (++), RAC (++), RBC (++), RHC (++), 18 medium not tested CTL(+++), CBR (+++), MB (+++), HB (+++) 21 small not labeled CTL (+), HPL (+) 23 medium MN (+), DRG (+) HPL (++), MF (+) 2 small NM (+++), DRG (+++) CTL (+++),HPL (+++) RGC (+++), RAC (+++), 4 small not labeled CTL (++++),HPL (++++) Brainbow-3.2 5 large not tested whole brain (+) nØ/Ofw/Gfw/Kfw 6 medium NM (+++), DRG (++) CTL (+++),HPL (+++) RGC(+++), RAC(+++), RBP (+++), 7 medium SIN (++) CTL (++++), HPL (++++), CGC (++++) 1 Large NM (++), DRG (+) CTL (++), HPL (++) Autobow 2 very small not labeled CTL (++), HPL (++) CreINT/Cer/P/K 3 medium NM(+) CTL (+), HPL (+) 1 medium NM (+), DRG (+) CTL (+), HPL (+) 2 large not tested whole brain (+) 3 large DRG (+) RGC (++), whole brain (+) Flpbow-3.0 5 very small not labeled RGC (+) T/P/C RGC (++), CTL (++), HPL (++), 6 medium NM (++), SIN (++) CP (++), MB (++), HB (++) 7 large NM (++) CTL (++), HPL (++), MB (++), HB (++) Flpbow-3.1 nØ/sOf/sGf/sKf 4 large not tested MF (+), HB (+) 5 6 small large not labeled MN (+), RDG (+) RGC (+), CTL (+), HPL (+), MF (+) whole brain (+) A total of 46 Brainbow 3 transgenic mouse lines were generated. 15 lines, not described here died, had, no detectable XFP expression, or did not transmit the Brainbow transgene to their progeny. Lines available from Jackson Laboratory are indicated in bold. f, farnesylated XFP; n, nucleic localization sequence; w, WPRE; s, SUMO tag; CreINT, intron inserted Cre; O, monomeric Orange2; G, EGFP; K, monomeric Kate2; P, PhiYFP; Ø, “dark” PhiYFP; T, tandem-dimer Tomato; Cer, Cerulean; C, ECFP; MN, motor neuron; SIN, spinal cord interneuron; SSN, skin sensory neuron; DRG, dorsal root ganglion; RGC, retina ganglion cell; RAC, retina amacrine cell; RBP, retina bipolar cell; RHC, retina horizontal cell; CBR, cerebellar neuron; CPC, cerebellar Purkinje cell; CGC, cerebellar granule cell; CTL, cortical neuron; HPL, hippocampal neuron; CP, caudate putamen; TH, thalamus; MB, mid brain; HB, hindbrain; MF, Mossy fiber. * Specific antibodies effectively enhanced the brightness of all XFPs in all lines. § Autobow line 1 and 2, founders were screened; line 3, founder as well as the first and the 6th generations were screened. Nature Methods: doi:10.1038/nmeth.2450