SuperGATSyn: a super-node graph attention network for drug synergy prediction

Abstract

Identifying synergistic drug combinations is crucial for cancer treatment but poses significant challenges due to combinatorial complexity and the time-consuming wet-lab validation. This study proposes SuperGATSyn, a super-node graph attention framework for drug synergy prediction. The model introduces a cell-feature super-node into each drug molecular graph, enabling cell-line features to directly participate in molecular message passing, rather than being incorporated only at the final prediction stage. A shared graph attention encoder symmetrically processes drug pairs to obtain their super-node representations, which are then fused through an element-wise max operation and concatenated with cell-line features for final prediction. Parameter sharing and element-wise max fusion operations collectively ensure the invariance of drug pairs to the input order. The super-node representation directly serves as the graph embedding, replacing conventional graph readout operations. It can identify key molecular fragments through attention weight analysis, providing interpretable insights into synergy mechanisms. Experiments on the O'Neil dataset show that SuperGATSyn achieves an ROC AUC of 0.94, a PR AUC of 0.94, and an ACC of 0.87 in five-fold cross-validation. On the AstraZeneca dataset, the model achieves an ACC of 0.65 and a TPR of 0.78 on unseen drug combinations. These results suggest that SuperGATSyn provides an effective framework for co-modeling drug structures and cell-line features in drug synergy prediction.