DOI RECORD
Explainable LightGBM-Based Drug Repositioning Framework for Cardiovascular Disease Prediction Using DrugBank Knowledge and Feature Fusion
Abstract
Cardiovascular disease (CVD) remains one of the leading causes of mortality worldwide, necessitating efficient strategies for identifying new therapeutic interventions. This study proposes an explainable machine learning framework for cardiovascular drug repurposing by integrating DrugBank XML data with drug, disease, and drug-disease association datasets. A comprehensive preprocessing pipeline comprising XML parsing, disease mapping, categorical encoding, TF–IDF text representation, feature engineering, and heterogeneous feature fusion is employed to generate robust predictive features. An optimized LightGBM classifier is utilized to prioritize potential cardiovascular drug candidates from the fused feature space, while SHAP-based explainable artificial intelligence provides both global and local interpretation of model predictions. Experimental evaluation demonstrates that the proposed framework achieves an accuracy of 99.33%, F1-score of 99.38%, and ROC-AUC of 99.17% on the DrugBank dataset. Furthermore, 10-fold cross-validation confirms the robustness and generalization capability of the model, yielding 99.55% accuracy, 99.56% F1-score, and 99.24% ROC-AUC, while consistently outperforming conventional machine learning baselines.