Gut microbiome functional pathways outperform taxonomic profiles in predicting immune checkpoint inhibitor response in non-small cell lung cancer: an interpretable machine learning approach with SHAP

Introduction: Lung cancer remains the leading cause of cancer mortality worldwide, with non-small cell lung cancer (NSCLC) accounting for the majority of cases. Although immune checkpoint inhibitors (ICIs) have transformed the therapeutic landscape of NSCLC, clinical responses remain highly variable. Emerging evidence implicates the gut microbiome in modulating the outcomes of ICI treatment; however, most studies to date have focused on taxonomic composition rather than microbial functional capacity. This study aimed to systematically compare the predictive value of taxonomic versus functional gut microbiome features across multiple ICI-related outcomes. Methods: Pretreatment fecal samples from 77 Japanese patients with NSCLC receiving ICIs were profiled using 16S rRNA sequencing. Six feature sets, comprising three taxonomic (family, genus, and species) and three functional (KEGG Orthology, Enzyme Commission, and MetaCyc pathways), were assessed using permutational multivariate analysis of variance for their association with clinical outcomes, including treatment response, irAEs, progression-free survival, and overall survival. Machine-learning models were subsequently developed based on MetaCyc pathway features to predict treatment response, with nested internal and external validation to ensure robustness and SHapley Additive exPlanations (SHAP) analysis for model interpretability. Results: ), and PWY-5088 (L-glutamate degradation VIII to propanoate), achieved robust predictive performance, substantially outperforming any single feature. SHAP analysis confirmed that the primary drivers of responder classification were pathways involved in nitrogen metabolism and short-chain fatty acid biosynthesis. Conclusions: In this study, gut microbial functional profiles consistently outperformed taxonomic features in predicting ICI response in patients with NSCLC. These findings suggest that metabolic pathway-based signatures may capture functional microbiome-host interactions more effectively and hold greater promise as translatable, safer targets for precision intervention, particularly through metabolite-oriented strategies.