Download PPTX - Tandy Warnow

Ensembles of HMMs and their use
in biomolecular sequence analysis
Nam-phuong Nguyen
Carl R. Woese Institute for Genomic Biology
University of Illinois at Urbana-Champaign
Human Microbiome
• 10 times more bacteria cells
than human cells
• Important role in regulating
health
• Disruption associated with risk
factors for diseases
Metagenomics
• Analyzing DNA sequences
from environmental sample
• Sequencing technology
produces short fragments of
DNA
• Typical datasets contain
millions of reads
Phylogenetic pipeline
Hu
Ch
Go
Or
=
=
=
=
AGGCTATCACCGACTCCA
TAGCTATCACGACCGC
TAGCTGACCGC
TCACGACCGACA
Hu
Ch
Go
Or
=
=
=
=
-AGGCTATCACGACCTCCA
TAG-CTATCACGACCGC-TAG-CT-----GACCGC--------TCACGACCGACA
Using the MSA and tree to identify reads
Hu
Ch
Go
Or
Qu
=
=
=
=
=
-AGGCTATCACGACCTCCA
TAG-CTATCACGACCGC-TAG-CT-----GACCGC--------TCACGACCGACA
TCACCCC
-------TCACC-CC----
Q
Represent MSA using a
profile Hidden Markov Model (HMM)
Phylogenetic Placement
• Align each query sequence to backbone
alignment:
• HMMALIGN (Eddy, Bioinformatics 1998)
• PaPaRa (Berger and Stamatakis, Bioinformatics 2011)
• Place each query sequence into backbone tree,
using extended alignment:
• pplacer (Matsen et al., BMC Bioinformatics 2010)
• EPA (Berger et al., Systematic Biology 2011)
Phylogenetic Placement
• Align each query sequence to backbone
alignment:
• HMMALIGN (Eddy, Bioinformatics 1998)
• PaPaRa (Berger and Stamatakis, Bioinformatics 2011)
• Place each query sequence into backbone tree,
using extended alignment:
• pplacer (Matsen et al., BMC Bioinformatics 2010)
• EPA (Berger et al., Systematic Biology 2011)
HMMER and PaPaRa results
Backbone size:
500
5000 fragments
20 replicates
0.0
Increasing rate evolution
Standard approach (single HMM)
HMM 1
Large evolutionary diameter
New approach
HMM 1
Smaller
evolutionary
diameter
HMM 1
HMM 2
HMM 2
HMM 3
HMM 4
Ensemble of HHMs (eHMMs)
HMM 1
HMM 3
HMM 2
HMM 4
SEPP (10% rule) Simulated Results
Backbone size:
500
5000 fragments
20 replicates
0.0
Increasing rate evolution
Summary so far
• Use DNA sequences to build an MSA
and tree
• Use an existing MSA and tree to identify
a sequence
• eHMMs for aligning a sequence to an
existing MSA
Metagenomic taxon identification
Objective: classify short reads in a metagenomic sample
Abundance profiling
Objective: distribution of the species (or genera, or
families, etc.) within the sample
For example, the distribution of a sample at the species
level might be:
Species A: 10%
Species B: 25%
Species C: 55%
Species D: 1%
Species E: 9%
Genome-based profiling
A
A
B
Population of 2 bacteria, A and
B. B has twice as large
genome as A.
True profile: 67% A, 33% B
Profile estimated from reads: 50% A,
50%B
Single copy marker-based
profiling
A
A
Population of 2 bacteria, A and
B. B has twice as large
genome as A.
Each have a single copy of gene C
B
True profile: 67% A, 33% B
Profile estimated from reads: 67% A,
33%B
TIPP: Taxonomic Identification and
Phylogenetic Profiling
Fragmentary unknown reads
for a gene
Known full length sequences for a gene,
and an alignment and a tree
ACCG
CGAG
CGG
GGCT
…
…
…
…
ACCT
AGG...GCAT
(species1)
TAGC...CCA
(species2)
TAGA...CTT
(species3)
AGC...ACA
(species4)
ACT..TAGAA
(species5)
TIPP: Taxonomic Identification and
Phylogenetic Profiling
• Nguyen et al., Bioinformatics, 2014
Reads
Assign to
marker genes
Marker
genes
Classify
reads
Compute
profile
Abundance profiling
• Objective: Distribution of the species (or genera, or families,
etc.) within the sample.
• Leading techniques:
– PhymmBL (Brady & Salzberg, Nature Methods 2009)
– NBC (Rosen, Reichenberger, and Rosenfeld, Bioinformatics 2011)
– MetaPhyler (Liu et al., BMC Genomics 2011), from the Pop lab at the
University of Maryland
– MetaPhlAn (Segata et al., Nature Methods 2012), from the Huttenhower
Lab at Harvard
– mOTU (Bork et al., Nature Methods 2013)
• MetaPhyler, MetaPhlAn, and mOTU are marker-based
techniques (but use different marker genes).
“Hard” genome datasets (known
genomes and high indel error)
Note: NBC, MetaPhlAn, and Metaphyler cannot classify any sequences from at
least of the high indel long sequence datasets. mOTU terminates with an error
message on all the high indel datasets.
“Novel” genome datasets
Note: mOTU terminates with an error message on the
long fragment datasets and high indel datasets.
TIPP compared to other
profiling methods
• TIPP is highly accurate, even in the presence of
novel genomes and high sequencing error
• All other methods are less robust
• Accurate profiles can be estimated using only a
portion of the reads
Ensemble of HMMs
• Represent MSA using many HMMs
• Modifications enable
– Fast and accurate alignment of fragmentary and ultra-large
datasets (Nguyen et al., Genome Biology 2015 )
– Improved protein homology detection (in preparation)
• Currently in use for
–
–
–
–
–
Vertebrate nuclear receptor evolution (in preparation)
1KP Plant phylogenomics study (in preparation)
Identification of cardioviruses in rats (in preparation)
Identification of microbial sample (in preparation)
and many others…
Ensemble of HMMs
• Represent MSA using many HMMs
• Modifications enable
– Fast and accurate alignment of fragmentary and ultra-large
datasets (Nguyen et al., Genome Biology 2015 )
– Improved protein homology detection (in preparation)
• Currently in use for
–
–
–
–
–
Vertebrate nuclear receptor evolution (in preparation)
1KP Plant phylogenomics study (in preparation)
Identification of cardioviruses in rats (in preparation)
Identification of microbial sample (in preparation)
and many others…
Real biological data is messy
Full-length P450 gene
~500 amino acid
residues
Total sequences before
filtering ~225K
Challenge: How do we
align large datasets
with fragmentary
sequences?
HMMs for MSA


Given seed alignment and a collection of sequences for
the protein family:

Represent seed alignment using a profile HMM

Align each additional sequence to the HMM

Use transitivity to obtain MSA
Drawbacks:

Requires seed alignment

Poor accuracy on evolutionarily divergent datasets
Old approach using single HMM
HMM 1
SEPP/TIPP approach
HMM 1
HMM 3
HMM 2
HMM 4
How small of a subset size do we go to?
HMM 1
HMM 3
HMM 2
HMM 4
Keep all HMMs
HMM 1
Keep all HMMs
HMM 1
HMM 2
HMM 3
Nested Hierarchical Ensemble of HMMs
HMM 1
HMM 2
HMM 4 m
HMM 7
HMM 3
HMM 5
HMM 6
UPP: Ultra-large alignment using
phylogeny aware profiles
UPP: Ultra-large alignment using
phylogeny aware profiles
UPP: Ultra-large alignment using
phylogeny aware profiles
UPP: Ultra-large alignment using
phylogeny aware profiles
UPP: Ultra-large alignment using
phylogeny aware profiles
UPP: Ultra-large alignment using
phylogeny aware profiles
Experimental Design

Examined both simulated and biological DNA, RNA, and AA
datasets

Generated fragmentary datasets from the full-length datasets

Explored impact of algorithmic design

Compared Clustal-Omega, Mafft, Muscle, PASTA, and UPP

ML trees estimated on alignments

Scored alignment and tree error


5/14/14
Alignment error measured as average of SPFN and
SPFP
Tree error measured in FN rate or Delta FN rate
UPP Algorithmic Parameters
• Decompose or not? Use an ensemble of
HMMs or just a single HMM?
• Use a small (100 sequence) or large (1000
sequence) backbone?
RNASim Alignment Error
Full-length datasets
5/14/14
Alignment error on fragmentary 16S.T
Fragmentary datasets
5/14/14
Running time on simulated RNA datasets
UPP has close to linear runtime scaling
UPP compared to other
alignment methods
• PASTA and UPP result in accurate alignments
and trees on full-length sequences (PASTA
slightly more accurate trees)
• UPP is more robust on fragmentary data
• Using combination of UPP+PASTA can give best
overall result
Summary
• Ensemble of HMMs
• TIPP for identification and profiling
• UPP for ultra-large alignment
Acknowledgements
• Illinois
• Tandy Warnow
• Rebecca Stumpf
• Bryan White
• Mike Nute
• Brenda Wilson
• UCSD
• Siavash Mirarab
• UMD
• Mihai Pop
• Bo Liu
• U of Copenhagen
• Alonzo Alfaro-Núñez
• Tom Hansen
• Anders Hansen
• Funding
• NSF 09-35347
• NSF 08-20709
• NSF 0733029
• University of Alberta
Questions?
• Available at
https://github.com/smirarab/sepp