DrugInsight: Multi-Source Evidence Fusion for Explainable Drug-Drug Interaction Prediction Using Graph Neural Networks

DrugInsight is an end-to-end, interpretable drug- drug interaction (DDI) prediction system. It evaluates interaction probability and severity while generating structured clinical explanations. The system utilizes RDKit and PyTorch Geometric to process drug SMILES strings into molecular graphs. These graphs are encoded by an AttentiveFP Graph Neural Network (GNN) into fixed-length structural embeddings. The embed- dings are concatenated with a comprehensive 12-dimensional pharmacological feature vector. This vector encompasses shared biological entities from DrugBank and post-market safety signals (PRR) from the TWOSIDES database. A Multi-Layer Perceptron (MLP) binary classifier processes this combined representation. The final interaction probability is computed using a tiered fusion model. This model adaptively weights curated DrugBank rules, ML classifier scores, and TWOSIDES pharmacovigilance signals to generate a unified risk index and severity classification. The system achieves a validation AUC of 0.7065 under a rigorous cold-start drug-level split. This demonstrates robust generaliza- tion to entirely unseen compounds. DrugInsight ensures high transparency through a rule-based explainer detailing metabolic competition, pharmacodynamic overlap, and explicit clinical rec- ommendations. The system is packaged for versatile deployment via a Streamlit web interface, a FastAPI REST endpoint, and a command-line interface (CLI).