Abstract:
Predicting drug combinations is crucial for effective cancer treatment. Combining multiple drugs can reduce resistance and improve outcomes. However, identifying these combinations experimentally is costly and time-consuming. This study introduces SynerDL, a deep-learning model for predicting synergistic drug combinations. We developed SynerDB, an enhanced dataset with more drug-protein and protein-protein interactions than previous datasets like DrugCombDB and the O’Neil Dataset. We used protein-protein interaction (PPI) graphs to generate protein embeddings and a transformer-based multi-head attention approach captured drug features from drug SMILES. SynerDL outperformed existing models, such as KGNN, GCN, GraRep, and DeepSynergy. It reduced both false positives and false negatives. SynerDL predicted several new synergistic drug combinations for breast cancer. Twenty of these combinations have been validated in the literature. On the SynerDB dataset, SynerDL achieved an accuracy of 0.90, precision of 0.85, recall of 0.94, F1-score of 0.89, and AUC-ROC of 0.96. It is the most reliable model for predicting drug combination synergy. Incorporating more protein interaction data further improved the model’s accuracy.
