Download Cancer Invasion Signaling Network

Methods
Inference of Signaling Network
Network inference was performed as in Vaske et al, 2008. The parameters for E-gene
expression distributions were as in the previous publication: means were set to -1.75 and
1.75 for the positive and negative distributions, and all s.d. were 0.5.
In addition to the E-genes selected previously for network inference, E-gene expression
spots were chosen from the expression data new to this study. A spot was selected if the
means of expression under a knockdown varied significantly more than the expression
means of each replicate number. For each spot on the microarray, the F-statistic F =
MSTRNAi / MSTReplicate, where MSTRNAi is the mean sum of squares among the RNAi
treatment grouping and MSTReplicate is the mean sum of squares among the replicate
number treatment grouping. E-gene spots significant at the 0.05 level from an Fdistribution with df=(13,2) were included for network inference.
We assessed the bootstrap confidence of inferred network features. For each of 1000
repetitions, we generated a sampled expression matrix of the same size as the original by
selecting rows from the original expression matrix with replacement and inferred a
signaling network. For any predicted interaction from the original data, the bootstrap
confidence was calculated as the fraction of predictions from sampled matrices that had
the same interaction mode as the original predicted interaction.
Frontier Expansion
The signaling network inferred above was used to score every gene from the microarray
for attachment to the network. As in Vaske et al. 2008, the LAR score was calculated as
the ratio of maximum likelihood of connection to the network over the likelihood of
disconnection from the network.
For Gene Ontology (GO) [cite GO] enrichment analysis of the frontier E-genes, a
frontier set was constructed for each S-gene consisting of any attached E-gene that had a
positive LAR for inhibitory or activating attachment. Identifiers were mapped to Entrez
Gene for both frontier sets and GO sets. Enrichment scores were calculated for every
frontier set-GO set overlap using the hypergeometric distribution, and the GO sets were
sorted by their maximum enrichment p-value with any frontier set. The list of GO sets
was filtered by 1) removing sets with fewer than 10 or more than 500 members and 2)
removing any set whose intersection with a GO set with better enrichment p-value is 10%
of the size of either set. Using the p-values for enrichment between all frontier sets and
this filtered list of GO sets, q-values were calculated using the QVALUE package in R
with default parameters.
Results
Gene knockdown implicates 13 genes as required for invasion in
colon cancer
Cancer Invasion Signaling Network
We predicted a signaling network between the invasiveness genes using FG-NEM (see
Methods) on 1114 E-genes. The resulting is fully connected, and spans all three tiers of
S-gene discovery, using results from three different microarray platforms.
We assessed confidence in network features by bootstrap. Choose a few examples of links
where the confidence of a particular branch point is lower than the over confidence that
there is a branch somewhere (i.e. CCR9).
The signaling network has three entry points: SCN5A, STK24, and KRT20/RPL32. The
network can be broken into four topological and functional domains, shown in Figure
(F_NETWORK). Domain I consists largely of integral membrane proteins and proteases,
and is upstream of all other network domains. Domain III exhibits calcium dependent
signaling and protease behavior. Domain IV is downstream of domains I and II, has an
independent signaling input from STK24, and contains genes related to secretion.
NARG1 is regulated by NMDA receptors, which admit both NA+ and CA2+ ions into
cell. In other cell types, the NMDA receptor functionally control lymphocyte activation
and are thought to be critical in synaptic plasticity
CCR9 elevates cytosolic calcium when activated by CCL9/TECK.
CAPN12 is a protease which is activated by calcium. Other members of the family are
known to modulate integrin-cytoskeletal interactions
Appears to be regulatory, due to a signal from upstream kinases. May be co-translational
translocation, due to EIF2AK2.
CHMP4C is part of the ESCRT III complex, which remodels the cell membrane and
degrades membrane proteins. The EST AA099748 would be from a rare isoform that may
allow membrane remodeling in other contexts.
SCRN3 is part of a gene family with members that regulate secretion from mast cells, and
potentially secretes from lysosomes, which may be connected to CHMP4C.
SEC24D is involved in vesicle trafficking, and export from the endoplasmic reticulum.
Signaling network predicts new gene and pathway regulation
throughout invasiveness pathway
SCN5A is a top-level invasiveness regulator
STK24 is a top-level regulator of secretory output
Figure Legends
Figure NETWORK
Predicted Network Interactions. A. Inferred S-gene network and Frontier. Nodes
represent S-genes (ovals), E-genes (gray boxes), and Gene Ontology categories (white
boxes). Arrows indicate activation, and tees indicate repression. Mixed arrow/tee line
endings indicate GO set enrichment among both activated and inhibited E-genes. B.
Expression values of selected E-genes. Each row shows the log-ratio expression of a
single E-gene under various shRNA knockdowns to a GFP shRNA knockdown control.
C. S-gene interaction confidence. Each pixel in the heatmap corresponds to an S-gene
interaction’s bootstrap confidence. For each interaction, the parent S-gene is labeled to
the rich, and the child S-gene is labeled to the bottom. Note that though NEM include all
transitive interaction, they are not displayed in (B) for simplicity. Therefore, a row
shows bootstrap confidence of an S-gene being upstream of other genes, and a column
shows bootstrap confidence of a gene being downstream of other genes.