TransGAT-DTA: A multi-task framework for drug-target affinity prediction and conditional molecule generation

Discovering novel molecules that effectively interact with target proteins remains a time-consuming and costly challenge in drug development. Current machine learning approaches are typically limited to single-task settings and cannot simultaneously address drug-target affinity prediction and targeted molecule generation. To overcome this limitation, we propose TransGAT-DTA, a multi-task learning framework that integrates a shared graph-Transformer architecture for molecular feature encoding with gated CNNs for hierarchical protein sequence representation. Cross-modal feature integration is achieved via a learnable alignment module, and a dynamic gradient coordination mechanism balances the regression objective for affinity prediction with the variational lower-bound optimization for molecule generation. Additionally, conditional control attention guides the model to generate molecules consistent with predicted target affinities. Experimental evaluation on benchmark datasets including BindingDB, Davis, and KIBA demonstrates that TransGAT-DTA reduces mean squared error in affinity prediction relative to single-task models, while simultaneously enabling the generation of high-quality, target-specific molecules. The end-to-end design mitigates error accumulation inherent in conventional cascaded approaches and provides a robust framework for bidirectional target-molecule optimization. These results establish a foundation for future applications in multi-target drug discovery.