AI Maps ME/CFS Biology Linking Gut Microbes, Immunity and Metabolites to Symptoms

ME/CFS affects over 10 million people worldwide and is characterized by debilitating fatigue, post-exertional malaise, cognitive impairment, pain, and gastrointestinal symptoms. Its multifactorial etiology, symptom heterogeneity, and nonlinear disease progression have made it extremely difficult to study—and nearly impossible to address with single-biomarker or single-omics approaches. This paper introduces BioMapAI, a systems-level AI framework designed specifically to handle this complexity.

The team enrolled 249 participants over 3–4 years: 153 ME/CFS patients (75 short-term disease <4 years; 78 long-term >10 years) and 96 age- and gender-matched healthy controls. Across 515 time points, they collected gut microbiome data via whole-genome shotgun metagenomics (1,293 species; 9,993 KEGG genes), plasma metabolomics (958 metabolites by untargeted LC-MS/MS), immune profiling by flow cytometry (443 immune cell and cytokine features), standard blood laboratory tests (48 features), and 12 validated clinical symptom scores spanning fatigue, pain, cognitive efficiency, sleep, orthostatic intolerance, and gastrointestinal function.

BioMapAI is a fully connected deep neural network with two shared hidden layers for general pattern learning and a parallel task-specific hidden layer for each of the 12 clinical outputs. SHAP explainability was integrated to quantify feature importance at both the symptom and disease levels. The model achieved 91% AUC in 10-fold cross-validation for ME/CFS versus control classification—outperforming elastic net, SVM, gradient boosting, and a standard DNN—and generalized to held-out and independent external cohorts. Individual omics-specific models (DeepMECFS-Immune, DeepMECFS-Metabolome, etc.) and an integrated 154-feature model (DeepMECFS-Omics) are publicly available.

The SHAP-derived connectivity map revealed mechanistically specific alterations in ME/CFS. Microbial depletion of short-chain fatty acids (notably butyrate) and branched-chain amino acids was linked to hyperactivation of mucosal-associated invariant T (MAIT) cells and γδT cells, which overproduce IFN-γ and granzyme A—an inflammatory signature correlated with worse perceived health and reduced social functioning. Separately, reduced microbial tryptophan and benzoate metabolism was associated with diminished connections to plasma lipids and bile acids, tracking with fatigue, emotional dysregulation, and sleep disturbances. These maps were adjusted for age, gender, and other clinical confounders, enabling cleaner interpretation of disease-specific versus health-associated interactions.

This work is notable for several reasons beyond ME/CFS: it provides a generalizable AI framework for integrating multi-omics data with heterogeneous clinical phenotypes across complex chronic diseases. Given the recognized overlap between ME/CFS and long COVID in etiology and symptomatology, these findings may also illuminate mechanisms underlying post-viral syndromes more broadly. Caveats include the predominantly female cohort (68%), the relatively modest sample size for deep learning, and the cross-sectional nature of the omics-symptom associations, which limits causal inference.