An Explainable Artificial Intelligence Framework for Predicting Anti-Diabetic Phytochemicals from Indian Medicinal Plants

Abstract

The systematic identification of bioactive phytochemicals from medicinal plants is hindered by the chemical variation among natural compounds and the low interpretability of traditional computational screening methods. Hence, in this study, we proposed a transparent, explainable artificial intelligence (XAI) framework to predict and rank potential antidiabetic compounds from Indian ethnomedicine. Using physicochemical and structural molecular descriptors, a curated dataset of 386 phytochemicals was created and analysed through classifications with Logistic Regression, Support Vector Machine, Random Forest, and Extreme Gradient Boosting (XGBoost). Model performance was evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and MCC for classification tasks, while Shapley Additive exPlanations (SHAP) were used to explain molecular-level results. The best predictive performance was achieved by XGBoost, with a ROC-AUC of 0.947, showing better generalisation than other classifiers. In conclusion, the SHAP analysis revealed that topological polar surface area, lipophilicity (LogP), potential hydrogen-bonding groups, molecular weight and structural features played key roles as determinants influencing antidiabetic efficacy and alluded to nonlinear structure–activity relationships in medicinal plant phytochemical studies. Moreover, the XAI-driven approach also identified non-dietary pharmacologically relevant phytochemicals quercetin, berberine, kaempferol, gymnemic acid and curcumin as further prospective computational candidates. This framework combines prediction accuracy with structural interpretability to determine candidate structures and can be used as a final step to elevate the phytochemical exploration space by helping choose structure-based natural-product candidates for subsequent experimentally validated target-specific probes.