620 likes | 773 Views
Computational Metagenomics : Algorithms for Understanding the " Unculturable " Microbial Majority . Sourav Chatterji UC Davis Genome Center schatterji@ucdavis.edu. Background. The Microbial World. Exploring the Microbial World. Culturing Majority of microbes currently unculturable .
E N D
Computational Metagenomics: Algorithms for Understanding the "Unculturable" Microbial Majority SouravChatterji UC Davis Genome Center schatterji@ucdavis.edu
Exploring the Microbial World • Culturing • Majority of microbes currently unculturable. • No ecological context. • Molecular Surveys (e.g. 16S rRNA) • “who is out there?” • “what are they doing?”
Interpreting Metagenomic Data • Nature of Metagenomic Data • Mosaic • Intraspeciespolymorphism • Fragmentary • New Sequencing Technologies • Enormous amount of data • Short Reads
Overview of Talk • Metagenomic Binning • PhyloMetagenomics • The Big Picture/ Future Work
Overview of Talk • Metagenomic Binning • Background • CompostBin • PhyloMetagenomics • The Big Picture/ Future Work
Metagenomic Binning Classification of sequences by taxa
Current Binning Methods • Assembly • Align with Reference Genome • Database Search [MEGAN, BLAST] • Phylogenetic Analysis • DNA Composition [TETRA,Phylopythia]
Current Binning Methods • Need closely related reference genomes. • Poor performance on short fragments. • Sanger sequence reads 500-1000 bp long. • Current assembly methods unreliable • Complex Communities Hard to Bin.
Genome Signatures • Does genomic sequence from an organism have a unique “signature” that distinguishes it from genomic sequence of other organisms? • Yes [Karlin et al. 1990s] • What is the minimum length sequence that is required to distinguish genomic sequence of one organism from the genomic sequence of another organism?
DNA-composition metrics The K-mer Frequency Metric CompostBin uses hexamers
DNA-composition metrics • Working with K-mers for Binning. • Curse of Dimensionality : O(4K) independent dimensions. • Statistical noise increases with decreasing fragment lengths. • Project data into a lower dimensional space to decrease noise. • Principal Component Analysis.
PCA separates species Gluconobacter oxydans[65% GC] and Rhodospirillum rubrum[61% GC]
Effect of Skewed Relative Abundance Abundance 1:1 Abundance 20:1 B. anthracis and L. monogocytes
A Weighting Scheme For each read, find overlap with other sequences
A Weighting Scheme 4 5 5 3 Calculate the redundancy of each position. Weight is inverse of average redundancy.
N å w X i i = = μ i 1 w N N å = - - T M w (X μ ) (X μ ) w i i w i w = i 1 Weighted PCA • Calculate weighted mean µw : • Calculates weighted co-variance matrix Mw • Principal Components are eigenvectors of Mw. • Use first three PCs for further analysis.
Weighted PCA separates species PCA Weighted PCA B. anthracis and L. monogocytes : 20:1
Semi-Supervised Classification • 31 Marker Genes [courtesy Martin Wu] • Omni-present • Relatively Immune to Lateral Gene Transfer • Reads containing these marker genes can be classified with high reliability.
Semi-supervised Classification Use a semi-supervised version of the normalized cut algorithm
The Semi-supervised Normalized Cut Algorithm • Calculate the K-nearest neighbor graph from the point set. • Update graph with marker information. • If two nodes are from the same species, add an edge between them. • If two nodes are from different species, remove any edge between them. • Bisect the graph using the normalized-cut algorithm.
Apply algorithm recursively Generalization to multiple bins Gluconobacter oxydans [0.61], Granulobacter bethesdensis[0.59] and Nitrobacter hamburgensis [0.62]
Generalization to multiple bins Gluconobacter oxydans [0.61], Granulobacter bethesdensis[0.59] and Nitrobacter hamburgensis [0.62]
Testing • Simulate Metagenomic Sequencing • Variables • Number of species • Relative abundance • GC content • Phylogenetic Diversity • Test on a “real” dataset where answer is well-established.
Conclusions • Satisfactory performance • No Training on Existing Genomes • Sanger Reads • Low number of Species
Overview of Talk • Metagenomic Binning • Phylo-Metagenomics • Background • Incorporating Alignment Accuracy • The Big Picture/ Future Work
Phylogenetic Trees Charles Darwin, First Notebook on Transmutation of Species (1837)
Population Structure of Communities Garcia Martin et al., Nat. Biotechnology (2006)
Gene Family Characterization Yooseph et al., PLoS Biology (2007)
Manual Masking • Require skilled and tedious manual intervention • Subjective and non-reproducible • Impractical for high throughput data • Frequently ignored. “Garbage-in-and-garbage-out”
Probabilistic Masking using pair-HMMs • Probabilistic formulation of alignment problem. • Can answer additional questions • Alignment Reliability • Sub-optimal Alignments Durbin et al., Cambridge University Press (1998)
à Pr[x y , x, y] i j à = Pr[x y ] i j Pr[x, y] Probabilistic Masking • What is the probability residuesxiand yjare homologous? • Posterior Probability the residues xiand yjare homologous • Can be calculated efficiently for all pairs (and gaps) in quadratic time.
å à d Pr[r r ] ij i j i, j å d ij i, j Scoring Multiple Alignments • Calculate the “posterior probability matrix” and distances dijbetween every pair of sequences. • Weighted “sum of pairs” score for column r:
Testing The Balibase 3.0 Benchmark Database
Testing • Realign sequences using MSA programs like Clustalw. • Sensitivity: for all correctly aligned columns, the fraction that has been masked as good • Specificity: for all incorrectly aligned columns, the fraction that has been masked as bad
Performance Sensitivity Specificity Prob Mask 97% 93% Gblocks 53% 94%
Effect on Phylogenetic Inference Gblocks simulated data-set, PhyML likelihood tree
Consistency between Alignment Programs • Yeast Genome Data Set • 7 yeast species, 1502 “orthologs” in each. • Wong et al. , Science (2008). • Aligned using 7 programs • Different programs often give inconsistent answers. • Garbage in, Garbage Out? • Partial Data, confusing global alignment programs. • No Masking
Consistency between Alignment Programs Masking remove ~33% of inconsistencies
Consistency between Alignment Programs Masking remove ~75% of inconsistencies with high support
The Final Result A Phylogenetic Database/Pipeline (with Martin Wu)
Overview of Talk • Metagenomic Binning • Phylo-Metagenomics • The Big Picture/ Future Work
Population Structure Venter et al. , Science (2004)