Fig. 3: Overall effect size of diet/lifestyle vs. microbial compositional features.

A Logistic regression models were trained using age + sex (basic), basic + diet/lifestyle features, basic + microbiome features, and basic + diet/lifestyle + microbiome features. Null models included basic + random noise features that matched the range of the original variables. Compared to basic features (AUC = 0.69), diet/lifestyle variables improved cross validated performance significantly (AUC = 0.79) (p = 0.004 rank-sum test) while microbiome features did not (AUC = 0.67, p = 0.25). B Pearson correlation between diet and lifestyle variables significantly related to ASD phenotype within the logistic regression model (vertical) and all other lifestyle variables (horizontal). Columns are annotated by Z-score from a slope test within the combo (basic + diet/lifestyle + microbiome) logistic regression model. C Pearson correlation between axes of variation (Principal component analysis) that are related to ASD phenotype within the combo logistic regression model (vertical) and all lifestyle variables (horizontal). Columns are annotated by Z-score from a slope test within the combo model. D ASV abundances are ranked based on their scores across principal components. A set is either the eight biomarkers associated with ASD or the three associated with TD (Table 1). Axes where biomarkers appear significantly skewed to one end or the other (as compared to randomly distributed) as determined by gene set enrichment analysis are represented.