The athlete microbiome project: Integrating deep learning to reveal microbial associations of physical fitness

Regular physical training improves human fitness and health through direct effects on muscle and metabolism and indirect effects via alterations in gut microbiome composition. To determine whether athletes harbor a distinct gut microbiome and whether microbiome composition associates with established markers of physical fitness, VO₂max and percent fat-free mass, we conducted a secondary analysis of amplicon sequencing data and metadata from published human microbiota studies across three continents. Participants were categorized as athletes (n = 656) or non-athletes (n = 199). Using multivariate statistics, random forest models, and a multilayer perceptron neural network, we identified structured differences in gut microbiome composition, associated with fitness metrics and athletic status. Random forest regression models explained up to 63% of the variance in percent fat-free mass and 45% in VO₂ max, with taxa such as <i>Faecalibacterium</i>, <i>Megamonas</i>, <i>Bifidobacterium</i>, and <i>Blautia</i> ranking among the most informative predictive features across analyses. Classification models further demonstrated that athletic status could be predicted from microbiome composition: a mixed-effects-informed random forest achieved a balanced accuracy of 71%, while a multilayer perceptron captured coordinated, multivariate microbial patterns and achieved stable performance across stratified five-fold cross-validation and an independent, held-out test set comprising 20% of samples not used during model training (balanced accuracy = 0.91; AUC = 0.97). Together, these findings indicate that athletic status and fitness-related traits are associated with gut microbiome compositional patterns, highlighting candidate taxa for mechanistic validation and demonstrating the utility of integrative machine learning for distinguishing athletes from non-athletes.