MAFSyn: Drug synergy prediction via hierarchical attentive fusion and biological context integration

Combination therapy is a promising strategy for cancer treatment, yet experimental screening remains costly and time-consuming. Current computational methods for drug synergy prediction rely on flat feature concatenation, overlooking the hierarchical nature of molecular structures and the functional biological context of gene interactions. To overcome these limitations, this study proposes MAFSyn, a deep learning framework designed to learn hierarchical drug representations and biologically informed cell line embeddings for accurate and generalizable synergy prediction.

MAFSyn constructs drug representations via a two-stage attentive fusion strategy. First, graph-based global topological scaffolds and fingerprint-based local structural patterns are fused to capture global chemical contexts. Second, SMILES-based substructure sequences are integrated via a multi-head self-attention mechanism to refine the embedding with local functional semantics. For cell lines, multi-omics profiles are propagated over a protein-protein interaction (PPI) network to generate interaction-aware embeddings that incorporate topological dependencies among genes.

Comprehensive experiments on the O'Neil benchmark dataset demonstrate that MAFSyn achieves superior performance compared to state-of-the-art methods. For regression tasks, MAFSyn improves MSE by 5.5% and RMSE by 1.3% in leave-one-drug-out generalization, exhibiting superior generalization capability to unseen drugs. Ablation studies confirm the critical contribution of each proposed module, and case studies further validate the model's practical potential in identifying novel synergistic combinations. Results from external dataset experiments also indicate that our model has generalization capability across different data sources.

By hierarchically integrating structural and functional information and incorporating biological network context, MAFSyn significantly improves prediction accuracy and generalization. The proposed framework offers a robust and reliable computational tool for prioritizing potential synergistic drug combinations in cancer therapy.