GATRSyn: Advancing Anticancer Drug Synergy Prediction Through Graph Attention Networks and Transformer-based Feature Re-embedding.

The integration of deep learning with prior knowledge and multi-omics data enables precise synergy prediction of anticancer drug combinations. However, substantial computational and storage requirements challenge deep learning in managing extensive datasets and multi-scenarios. This study proposes a generalized two-stage framework to address this issue. It decomposes the drug synergy prediction into two key modules: offline feature extraction and multi-scenario adaptive prediction. By integrating drug-cell-associated protein-protein interaction networks with pharmacogenomics data, we first employ an edge-augmented GAT to construct powerful offline feature extractors for drugs and cell lines concurrently, which identify their shared patterns. Next, Transformer-based re-embedding is designed to fine-tune offline features online, which captures cross-modal interactions between specific drugs and cell lines. It can serve as a portable module to seamlessly integrate into various predictive modules. Under this framework, we develop GATRSyn, a dedicated two-stage model for multi-scenario drug synergy prediction. Compared to state-of-the-art deep learning models across multiple high-throughput screening datasets, GATRSyn achieves optimal or near-optimal results across eight distinct prediction scenarios. GATRSyn also effectively screens clinically validated or literature-reported synergistic drug combinations lacking measured scores. Moreover, synergy analysis of eight representative samples and enrichment analysis of pivotal proteins of cell lines and drugs highlights GATRSyn’s ability to model and elucidate targeted therapeutic efficacy of drug combinations against intracellular signaling nodes. Our framework enhances large-scale, multi-scenario drug synergy prediction and provides a comprehensive reference for individualized anticancer treatment.