DRUGSYNC: Prediction of Synergistic Drug Combinations Using GCN Graph Convolutional Network based and Pre-learned Drug-induced Gene Expression Profiles

Drug-drug synergy (DDS) is crucial for identifying novel and effective drug combinations. With the accumulation of experimental data (DrugCombDB) and the uprising of deep learning approaches, many efforts have been made to computationally predict effective drug pairs. However, most models had limited success, with the highest ROC AUC of 0.65 for predicting pairs of unseen drugs. Drug-induced gene expressions profiles (DGEP) from datasets like LINCS L1000 offer a robust proxy to describe drug effects. However, overlap between DrugCombDB and experimental DGEP data is sparse only ∼2%. To address these limitations, we propose the DRUGSYNC ( D rug R esponse and U tilization for G raph-based SYN ergy C lassification) prediction framework that leverages pre-learned drug-target interaction (DTI) and pre-learned DGEP as features. Graph Convolutional Network (GCN) was used to represent the inherent network of genetic interactions. Our pre-learned DGEP achieved 0.877 Pearson correlation coefficients (PCC). Furthermore, our framework excels by integrating pre-learned DGEP and DTI of paired drugs for DDS prediction, achieving superior performance over benchmark technologies. Notably, it demonstrates exceptional accuracy in predicting interactions between unseen drugs, with an ROC AUC score of 0.854. The analysis further emphasized the importance of accurate DGEP predictions. This method can be extended to any cell lines.