Tiny microbes, enormous impacts: what matters in gut microbiome studies?

Abstract

Many factors affect the microbiomes of humans, mice, and other mammals, but substantial challenges remain in determining which of these factors are of practical importance. Considering the relative effect sizes of both biological and technical covariates can help improve study design and the quality of biological conclusions. Care must be taken to avoid technical bias that can lead to incorrect biological conclusions. The presentation of quantitative effect sizes in addition to P values will improve our ability to perform meta-analysis and to evaluate potentially relevant biological effects. A better consideration of effect size and statistical power will lead to more robust biological conclusions in microbiome studies.

Fig. 1

PCoA differences in PCR primers can outweigh differences among individuals within one body site, but not the differences between different body sites. In the Human Microbiome Project (HMP) dataset, when V1-3 and V3-5 primers are combined across body sites, a the effect of PCR primers is small compared to b the effect of body site. However, if we analyze individual body sites such as c the mouth or d the mouth subsites, the effect of primer is much greater than the difference between different individuals (or even of different locations within the mouth) at that specific body site. GI gastrointestinal

Fig. 2

PCoA patterns of technical and biological variation. Two groups (black, gray) with significantly different distances (P < 0.05) and varying effect size. a A large separation in PCoA space and large effect size. Separation in PCoA space (shown here in the first two dimensions) may be caused by technical differences in the same sample set, such as different primer regions or sequence lengths. b Clear separation in PCoA space, similar to patterns seen with large biological effects. In cross-sectional studies, age comparisons between young children and adults or comparisons between Western and nonWestern adults might follow this pattern. c Moderate biological effect. d Small biological effect. Sometimes effects can be confounded. In e the technical effect and in f the biological effect are conflated because the samples were not randomized. In g and h, there is a technical and a biological effect, but the samples were randomized among conditions, so the relative size of these effects can be measured

Fig. 3

Relative effect sizes of biological covariates on the human microbiome. Principal coordinates projection of unweighted UniFrac distance, using data from Yatsunenko et al. [45], shows a age (blue gradient; missing samples in red) separating the data along the first axis and b country (USA, orange; Malawi, green; Venezuela, purple) separating the data along the second principal coordinates axis. c Body mass index in adults has a much more subtle effect, and does not separate along any of the first three principal coordinate axes (normal, red; overweight, green; obese, blue; missing samples, gray)