Comparing Analytical Methods for the Gut Microbiome and Aging: Gut Microbial Communities and Body Weight in the Osteoporotic Fractures in Men (MrOS) Study

Abstract

Determining the role of gut microbial communities in aging-related phenotypes, including weight loss, is an emerging gerontology research priority. Gut microbiome datasets comprise relative abundances of microbial taxa that necessarily sum to 1; analysis ignoring this feature may produce misleading results. Using data from the Osteoporotic Fractures in Men (MrOS) study (n = 530; mean [SD] age = 84.3 [4.1] years), we assessed 163 genera from stool samples and body weight. We compared conventional analysis, which does not address the sum-to-1 constraint, to compositional analysis, which does. Specifically, we compared elastic net regression (for variable selection) and conventional Bayesian linear regression (BLR) and network analysis to compositional BLR and network analysis; adjusting for past weight, height, and other covariates. Conventional BLR identified Roseburia and Dialister (higher weight) and Coprococcus-1 (lower weight) after multiple comparisons adjustment (p < .0125); plus Sutterella and Ruminococcus-1 (p < .05). No conventional network module was associated with weight. Using compositional BLR, Coprococcus-2 and Acidaminococcus were most strongly associated with higher adjusted weight; Coprococcus-1 and Ruminococcus-1 were most strongly associated with lower adjusted weight (p < .05), but nonsignificant after multiple comparisons adjustment. Two compositional network modules with respective hub taxa Blautia and Faecalibacterium were associated with adjusted weight (p < .01). Findings depended on analytical workflow. Compositional analysis is advocated to appropriately handle the sum-to-1 constraint.

Keywords: Bayesian regression; Compositional analysis; Frailty; Network analysis.

Figure 1.

Beta-diversity and weight principal coordinate analysis. Beta-Diversity measured using Weighted UniFrac (top) and Aitchison distance (bottom). p value computed using MiRKAT adjusted for covariates.

Figure 2.

Association of individual taxa with covariate-adjusted weight. Volcano plots from elastic net regression with conventional Bayesian linear regression of relative abundance adjusted for covariates (top) and compositional Bayesian linear regression of isometric log-ratios of relative abundance adjusted for covariates (bottom). Labelled taxa are the top 10 most significant taxa from each respective analysis. Taxa above the upper horizontal line have p < .05/4 = .0125; taxa above the lower horizontal line have p < .05.

Figure 3.

Heat maps of mean relative abundance of 50 unique taxa comprising top 30 taxa from individual taxa analysis by weight quartile. Arithmetic means clustered using Euclidean distance (top) and compositional means clustered using Aitchison distance (bottom).

Figure 4.

Taxa modules with respective hub taxa derived from weighted gene coexpression network analysis. Modules derived using Pearson correlations (top) and SparCC (bottom). See Supplementary Tables 4 and 6 for taxa and eigentaxon loadings. Edge thickness denotes strength of correlation. Edges with correlation > |0.2| shown.