Plant Diversity & Gut Microbiome Analysis

What this research found

The popular advice to eat more than 30 different plants a week rests largely on survey data from the American Gut Project (AGP). A reanalysis of 2,676 AGP participants using compositional statistics found that plant-type diversity does shift gut microbial community structure, but by a negligible amount, and that an eight-genus log-ratio which separated high- from low-plant eaters strongly in-sample collapsed to zero effect when tested on held-out countries. The conclusion was to redesign a follow-up feeding trial rather than fund one on the observed effect size.

• The community-wide effect is statistically detectable but negligible in size: plant exposure explained 0.116% of variance in a partial redundancy analysis on Aitchison distances (p = 0.004), and 0.111% once BMI was added to the model (p = 0.013).
• A supervised log-ratio balance built from eight genera separated the two groups strongly within the sample, with an adjusted coefficient of +0.450, t = 6.89, p = 8.3e-12, Cohen's d = 0.602 and partial R-squared of 3.41%. The effect held after BMI adjustment (+0.416, p = 3.35e-09).
• That balance does not transport across geography. Holding out the USA, United Kingdom, and Other cohorts in turn, training coefficients averaged +0.63 with all p below 1e-8, while held-out coefficients collapsed to a mean of -0.0004 and 0 of 3 folds reached significance.
• Predictive accuracy came from the diet questionnaire rather than the microbiome. A covariates-only classifier already reached out-of-sample AUC 0.82, adding the balance changed AUC by -0.02, and a balance-plus-demographics model with the dietary items removed managed only 0.63.
• Taxon selection was cohort-specific, with mean Jaccard overlap of 0.35 against the globally selected balance (0.14 for the USA, 0.23 for the UK, 0.67 for Other), although the order Clostridiales recurred in the numerator of every fold.
• A geographically stratified crossover feeding trial powered on the transportability-adjusted effect (d = 0.301) would need 54 subjects assuming a within-subject correlation of 0.7, against 350 participants for the equivalent parallel design.

How it was done

The AGP fecal sequence-variant table and mapping file were pulled from Figshare and reduced to one sample per participant, giving 2,676 people: 1,973 reporting 10 or fewer plant types per week and 703 reporting more than 30. Community structure was tested by partial redundancy analysis on centred-log-ratio-transformed counts, with 999 permutations of the exposure label restricted to within country strata so that geographic cohort structure could not manufacture the result. A forward-selection routine then built an interpretable genus-level log-ratio balance from 555 lineage groups, which was checked by five-fold cross-validation and by refitting with each country held out. Power calculations for parallel and crossover feeding designs closed out a 22-page manuscript with 9 figures and 46 references.

Data sources

• American Gut Project fecal sOTU table, deblur 125nt with blooms removed (Figshare article 6137192, 9,511 fecal samples)
• American Gut Project mapping file (Figshare article 6137315, 17,854 rows by 452 columns)
• Analysis cohort of 2,676 participants and 21,640 sequence variants aggregated to 555 genus-level groups

Figures from this analysis