Graph–Vector Hybrid Model for Predicting Synergistic Drug-Drug Interactions Toward Clinical Decision Support Integration

Drug-drug interactions are often associated with negative clinical outcomes and reduced therapeutic efficacy. But they can also be deliberately used to produce synergistic effects in polypharmacy, to enhance treatment outcomes in therapeutic regimens. To predict synergistic and antagonistic interactions, this study proposes a computational model that integrates molecular descriptors capturing physicochemical properties, structural similarity from Morgan fingerprints, solubility predictions obtained via a deep learning (DL) model, and drug-likeness filtering based on Lipinski’s Rule of Five. Initially a standardized molecular preprocessing pipeline is applied resulting in high-fidelity Simplified Molecular Input Line Entry System representations and related drug properties being exploited. Feature engineering is applied to compute molecular descriptors and the solubility is then predicted using a Graph Convolutional Network (GCN). Morgan fingerprints are produced to measure structural similarity with the Tanimoto coefficient. The paper proposes a hybrid model combining a GCN with a Feedforward Neural Network (FFNN). Using deep structural embeddings in addition to conventional descriptors, this model captures complex molecular substructure patterns, increasing prediction accuracy compared to baseline models. The proposed GCN–FFNN ensemble achieved an AUC of 0.83, significantly outperforming the Random Forest (RF) baseline (AUC of 0.62) and showed consistent performance across stratified validation folds in classifying both synergistic and antagonistic interactions. This is further supported by precision (0.81), recall (0.79), and F1-score (0.80) scores. The predictive framework can be integrated into intelligent Clinical Decision Support Systems giving clinicians real-time, data-driven insights to plan more efficient treatment plans while avoiding antagonistic interactions.