Multiscale Drug–Cell Interaction Modeling via Hybrid Graph Architectures and Knowledge-Aware Fusion for Drug Synergy Prediction

Accurate prediction of drug combination synergy remains challenging due to the vast combinatorial space and strong cellular context dependency of drug responses. To address these limitations, we present MILSyn, a multimodal interaction learning framework designed to capture multiscale drug–cell interactions through three integrated components: (1) relational modeling of heterogeneous graph structures using a hybrid RGCN-GATv2 architecture, (2) cross-modal fusion of biochemical features spanning SMILES, substructures, 3D conformations, and multiomic profiles, and (3) knowledge-aware reasoning via LLM-derived semantic embeddings. Evaluated on DrugComb and four benchmark data sets, MILSyn consistently outperforms state-of-the-art models, demonstrating robust generalization in cold-start scenarios and notable performance gains under few-shot fine-tuning for rare tumors. Attention-based analyses provide exploratory summaries of gene-, pathway-, and substructure-level features associated with the model’s predictions, generating hypotheses for further investigation. By combining scalability with interpretability, MILSyn provides a practical framework for prioritizing candidate drug combinations in precision oncology.