Fig 1.
Subsets of 5000 representative sequences for operational taxonomical units (OTU) of stool samples from the Human Microbiome Project have been drawn. Trees are obtained in three ways: alignment-free method (left), golden standard based on taxomic identity (middle), and alignment-based method using traditional methods (right). The trees are then compared to one another using Treedist, with the distance representing how similar they are to each other (lower number denotes greater similarity). Ten replicates of each comparison have been performed and the results are averaged.
Fig 2.
Tree distances of alignment-free methods and alignment-based methods relative to the gold standard taxonomic trees.
Ten replicate subsets of sequences from Greengenes have been obtained and phylogenies inferred using alignment-free methods (ACS, CVTree, and Kr) and alignment-based methods (PyNAST or MUSCLE-based alignment, FastTree or RAxML inference). TREEDIST distances between the phylogenies inferred by each method as well as the taxonomic gold standard have been computed. Smaller distances indicate better resemblance of the taxonomy in the corresponding inferred phylogenies. Sequences from HMP derived stool samples have been used to compare all the methods. Distances across the replicates are reported. P = PyNAST, M = MUSCLE, F = FastTree, R = RAxML.
Fig 3.
Tree distances of consensus tree of alignment-free methods, consensus of MUSCLE-based alignments, consensus of PyNAST-based alignments, consensus of FastTree inference, consensus of RAxML inference, consensus of all alignment-based methods relative to gold standard taxonomic tree.
Using the same subsets as in Fig 2, a consensus tree based on the three alignment-free methods has been built. Similarly, consensus trees based on different combinations of alignment-based methods are built. TREEDIST distances across the replicates are reported; smaller distances indicate better resemblance of the consensus tree to the gold standard taxonomic tree.
Fig 4.
Principal coordinates analysis of the weighted unifrac distances computed with (A) FASTTREE, (B) ACS, (C) Kr, (D) CV, (E) alignment free consensus phylogenies and grouped by sample location, and grouped by antibiotic treatment type C-control, P-penicillin, V-vancomycin, T-tetracyclin, VP-vancomycin and penicillin after centering according to sample location with (F) FastTree, (G) ACS, (H) Kr, (I) CV, and (J) alignment free consensus phylogenies.
Table 1.
Significance and effect size estimates for PERMANOVA testing (10,000 permutations) of the association of the microbiome and experimental variables.
Fig 5.
Comparison of effect size in PERMANOVA analysis with alignment-free consensus tree and with an alignment-based method.
We have simulated data with various effect sizes by resampling permuted communities from control fecal specimens of the STAT dataset. ω2 has been computed in comparison of the weighted Unifrac distances based on both trees. The plots show the log ratio of the p-value vs. the mean of the estimated effect sizes. In (A) the entire range of effect sizes is considered and we note that at high effect sizes there is agreement in inference based on the two trees. In (B) small effect sizes are examined closely. Here the significant positive intercept of the regression indicates that alignment-free consensus phylogeny results in lower p-value than phylogeny inferred with FastTree based on PyNAST alignment.