PerturbSynX: Deep learning framework for predicting drug combination synergy scores using drug induced gene perturbation data.

Drug synergy prediction plays a vital role in the discovery of effective cancer combination therapies by identifying drug pairs that work better together than individually. However, exhaustive experimental testing of all possible combinations remains resource intensive and impractical. To address this, we propose a deep learning based multitask learning framework that integrates multi modal biological data to simultaneously predict drug pair synergy scores and individual drug response outcomes. The model incorporates molecular descriptors and drug induced gene expression signatures to represent drugs, while untreated cancer cell lines are encoded through their gene expression profiles. A hybrid architecture based on bidirectional long short term memory (BiLSTM) layers and attention mechanisms is employed to capture complex interactions between drug features and cell line characteristics. The fused representation is passed through mutual attention layers to model intricate dependencies, and the resulting output is used to jointly predict the synergy score and the individual responses of each drug. The proposed model achieves an RMSE of 5.483, PCC of 0.880, and R2 of 0.757, demonstrating substantial improvements over existing approaches. These results underscore the model's strong predictive capability, particularly with the use of a mixed activation function and the Adamax optimizer at a learning rate of 0.001. To assess robustness, we conducted sensitivity analysis on drug input order, revealing high consistency but minor asymmetries attributable to the leave-pair-out training strategy. Overall, the proposed model outperforms baseline methods and offers a reliable, biologically grounded computational tool for accelerating drug synergy discovery in cancer research.