A Bayesian approach to inferring the phylogenetic structure of communities from metagenomic data

Abstract

Metagenomics provides a powerful new tool set for investigating evolutionary interactions with the environment. However, an absence of model-based statistical methods means that researchers are often not able to make full use of this complex information. We present a Bayesian method for inferring the phylogenetic relationship among related organisms found within metagenomic samples. Our approach exploits variation in the frequency of taxa among samples to simultaneously infer each lineage haplotype, the phylogenetic tree connecting them, and their frequency within each sample. Applications of the algorithm to simulated data show that our method can recover a substantial fraction of the phylogenetic structure even in the presence of high rates of migration among sample sites. We provide examples of the method applied to data from green sulfur bacteria recovered from an Antarctic lake, plastids from mixed Plasmodium falciparum infections, and virulent Neisseria meningitidis samples.

Keywords: Bayesian phylogenetics; metagenomics; microevolution.

Figure 1

A diagram of the lineage model. On the left, a coalescent process leads to a complete genealogy, with the tips marked by pool as colors. The right diagrams the lineage model approximation, showing deep branching events together with cones shading the SNP variation indistinguishable from noise.

Figure 2

Comparison between simulated tree and pool proportions (left) and inferred trees and pool proportions (right). In the inferred model, the dark blue tree shows the maximum posterior probability tree with the light blue trees representing samples from the MCMC. Pie charts on the right show inferred pool proportions, with shades indicating the posterior percentile from dark (5%) to light (95%).

Figure 3

(Left) Percentage of SNP similarity between simulated and inferred lineages. (Right) Comparison between pool proportions for simulated (light gray) and inferred (dark gray) values for each simulated lineage. Blue circles show combined proportions for simulated lineages 2 and 3.

Figure 4

Comparison of simulated and inferred values for lineages (left column) and pool proportions (right column) by number of SNPS (A and B), number of reads (C and D), and error rate (E and F). Insets give mixture value (“Mix”), number of read counts (“Reads”), number of SNPs (“SNPs”), and error rate (“Err”).

Figure 5

(Top) Location on simulated tree of SNPs for six sequence patterns. The branch width is proportional to number of SNPs. (Bottom) Inferred model presentation is the same as in Figure 2.

Figure 6

Inferred lineage model for Chlorobium data from Ace Lake and open ocean samples.

Figure 7

Inferred lineage model for Plasmodium falciparum apicoplast data from 20 clinical samples from northern Ghana.