Bacterial associations in the healthy human gut microbiome across populations

Abstract

In a microbial community, associations between constituent members play an important role in determining the overall structure and function of the community. The human gut microbiome is believed to play an integral role in host health and disease. To understand the nature of bacterial associations at the species level in healthy human gut microbiomes, we analyzed previously published collections of whole-genome shotgun sequence data, totaling over 1.6 Tbp, generated from 606 fecal samples obtained from four different healthy human populations. Using a Random Forest Classifier, we identified 202 signature bacterial species that were prevalent in these populations and whose relative abundances could be used to accurately distinguish between the populations. Bacterial association networks were constructed with these signature species using an approach based on the graphical lasso. Network analysis revealed conserved bacterial associations across populations and a dominance of positive associations over negative associations, with this dominance being driven by associations between species that are closely related either taxonomically or functionally. Bacterial species that form network modules, and species that constitute hubs and bottlenecks, were also identified. Functional analysis using protein families suggests that much of the taxonomic variation across human populations does not foment substantial functional or structural differences.

Linking Inflammatory Bowel Disease Symptoms to Changes in the Gut Microbiome Structure and Function

“…Bacterial genome relative abundances in each sample were estimated using a probabilistic framework based on a mixture model. The framework utilized an Expectation-Maximization (EM) algorithm to perform soft assignment of the reads to the reference genomes and was found to be highly accurate ( Xia et al, 2011 ; Loftus et al, 2021 ). We have previously demonstrated that samples with less than 250,000 mapped reads display diminished accuracy for taxonomic profiling, consequently all samples that contained less than 250,000 mapped reads threshold were not used for downstream analysis ( Loftus et al, 2021 ).…”

“…The framework utilized an Expectation-Maximization (EM) algorithm to perform soft assignment of the reads to the reference genomes and was found to be highly accurate ( Xia et al, 2011 ; Loftus et al, 2021 ). We have previously demonstrated that samples with less than 250,000 mapped reads display diminished accuracy for taxonomic profiling, consequently all samples that contained less than 250,000 mapped reads threshold were not used for downstream analysis ( Loftus et al, 2021 ). When calculating relative abundances, any value below 1 × 10 –5 was considered statistical noise (and set to 0).…”

“…To reduce the dimensionality of our data, we utilized only bacterial species that were present in at least 90% of samples within each diagnosis group (IBD, non-IBD, Healthy-1, and Healthy-2) for our downstream analysis ( Loftus et al, 2021 ). The union of the bacterial species present at a prevalence greater than or equal to 90% in each diagnosis group was then used for the classification of the signature species.…”

“…The signature species were used to create a sample-taxa matrix of CLR-transformed relative abundances in each sample. The GGM framework, as previously described, was used to generate the bacterial association networks using the above sample-taxa matrices for each cohort ( Loftus et al, 2021 ). In brief, the HUGE package in R was used to compute a sparse precision matrix.…”

“…Even after CLR-transformation of the sequencing data, pairwise correlation methods are unable to account for conditional independence between bacterial species causing these methods to produce inaccurate bacterial association networks ( Kurtz et al, 2015 ). In this paper, we used a Gaussian Graphical Model (GGM) framework in conjunction with a graphical lasso (glasso) to construct bacterial association networks from the CLR-transformed relative abundance data ( Friedman et al, 2008 ; Loftus et al, 2021 ). Utilizing the GGM framework on the CLR-transformed data, enables us to approximate the covariance structure of the absolute abundances as well as account for conditional independence between the constituent species ( Aitchison, 1982 ; Wermuth and Lauritzen, 1990 ).…”