A Method for Determining Taxonomical Contributions to Group Differences in Microbiomic Investigations

Abstract

Here we show how one can decompose the contribution of different levels of taxonomic classification in terms of their impact on differences in the microbiota when comparing two groups. First we demonstrate a difficulty in trying to estimate taxonomic effects at multiple levels simultaneously and demonstrate an approach to determining which taxa have differences in means that are identified. We then develop a model based on an approach that is popular in the RNA-Seq analysis literature and apply it to the problem of determining which taxa differ between two patient groups. This model provides a more powerful method than simpler alternatives. A Bayesian computational strategy is used to obtain exact inference. Simulation studies indicate that the procedure works as intended, and an application to the study of COPD demonstrates the method's practical utility. Software is provided for implementing the method.

Keywords: Bayesian modeling; COPD; microbiota; overdispersed counts.

FIG. 1.

Posterior medians of the ratio of the mean in the healthy group to the mean in the COPD group for the phylum Actinobacteria (lighter values indicate higher levels in the healthy group). Some family names have been truncated. COPD, chronic obstructive pulmonary disease.

FIG. 2.

Posterior medians of the ratio of the mean in the healthy group to the mean in the COPD group for the phylum Bacteroidetes (lighter values indicate higher levels in the healthy group). Some family names have been truncated.

FIG. 3.

Posterior medians of the ratio of the mean in the healthy group to the mean in the COPD group for four phyla with only a single genus (lighter values indicate higher levels in the healthy group).

FIG. 4.

Posterior medians of the ratio of the mean in the healthy group to the mean in the COPD group for the phylum Firmicutes (lighter values indicate higher levels in the healthy group).

FIG. 5.

Posterior medians of the ratio of the mean in the healthy group to the mean in the COPD group for the phylum Fusobacteria (lighter values indicate higher levels in the healthy group).

FIG. 6.

Posterior medians of the ratio of the mean in the healthy group to the mean in the COPD group for the phylum Proteobacteria (lighter values indicate higher levels in the healthy group). The class names have been truncated (removing “proteobacteria”).

FIG. 7.

Simulation results for six scenarios. Each plot shows the frequency with which the posterior probability of a difference between two groups exceeds 95% (the horizontal line is at 0.05). Taxa are ordered from left to right by taxonomical courseness (i.e., phyla are on the left and genera on the right).