Abstract
Accurate prediction of half-maximal inhibitory concentration (IC(50)) values is critical for accelerating drug discovery, yet traditional quantitative structure-activity relationship (QSAR) models often have limited ability to capture both local structural patterns and global physicochemical properties essential for bioactivity. We developed a hybrid deep learning framework that integrates graph neural networks with explicit molecular descriptors to address this limitation. The model learns from molecular graphs encoding atomic and bond features while incorporating interpretable physicochemical properties and structural fingerprints. Trained and validated on 14,316 compounds across nine diverse biological targets including kinases, nuclear receptors, and proteases, our approach achieved an overall test R(2) of 0.87, consistently outperforming previously reported methods by 6-42% across evaluated targets. The model demonstrated robust generalization with near-identical training and test performance, while maintaining partial interpretability through transparent descriptor contributions and attention mechanisms. By synergistically combining data-driven learning with domain knowledge, this hybrid framework offers improved accuracy and interpretability for structure-activity modeling, facilitating more efficient compound prioritization and optimization in early stage drug discovery programs.