TITLE:
AI-Enhanced Gut Brain Axis Profiling for Major Depressive Disorder: Integrating Synthetic Multi-Omics, Deep Learning, and Interpretable Precision Therapeutics
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Gut Brain Axis, Major Depressive Disorder, Microbiome, Metabolomics, Deep Learning, SHAP, Precision Nutrition, Explainable AI
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.1,
January
16,
2026
ABSTRACT: Major depressive disorder (MDD) has been repeatedly linked to disruptions of the gut brain axis (GBA), yet practical decision systems that convert multi-omics patterns into patient-specific guidance remain limited. We present an end-to-end, explainable pipeline that learns putative GBA signatures of MDD from synthetic data and translates model attributions into hypothesis-driven nutritional and pharmacological suggestions. We simulated a cohort of N = 1,500 individuals (30% MDD) comprising 200 microbial taxa, 150 metabolites, and 7 clinical features. A regularized dense neural network with class weighting and early stopping was trained and compared with a Random Forest baseline; interpretability was provided by SHAP at global and local levels. On a 20% stratified hold-out test set the deep model achieved Accuracy = 0.98 and AUC = 0.998, with a confusion matrix of [[207, 3], [2, 88]]. Feature attributions concentrated on a compact subset of metabolites and taxa consistent with the planted effects in the simulator; RF importances corroborated these signals (e.g., Metabolite_120, 109, 40, 90; Species_198, 197). We further demonstrate a templated mapping from patient-level SHAP profiles to non-clinical recommendations dietary patterns, prebiotic/probiotic directions, and pathway hypotheses involving, for example, kynurenine metabolism, short-chain fatty acids, and bile acids intended to support clinician-led hypothesis generation rather than direct treatment. Because all data are simulated, performance estimates are optimistic and biological interpretations are illustrative. Nonetheless, the approach shows how multi-omics learning coupled with transparent explanations can organize heterogeneous GBA signals into actionable research hypotheses for precision psychiatry. Code and figures are fully reproducible from a single Colab notebook. Future work will validate the pipeline on real, harmonized cohorts, incorporate compositional microbiome statistics and calibration analyses, and assess generalization across sites and subgroups under clinical oversight.