SMCL-DTA: surface-aware multi-view contrastive learning for drug-target affinity prediction

Accurate prediction of drug-target affinity (DTA) is crucial for accelerating drug discovery and advancing precision therapy. Existing approaches face fundamental limitations, including independent treatment of molecular and protein representations, susceptibility to overfitting with limited data, and inadequate augmentation strategies that may destroy critical molecular substructures. To address these challenges, we propose SMCL-DTA, a novel Surface-Aware Multi-View Contrastive Learning framework for DTA prediction. SMCL-DTA introduces three key innovations: an intelligent masking strategy guided by graph topology to preserve chemically meaningful substructures, a multi-view contrastive learning framework with regression-aware positive pair construction, and the comprehensive integration of molecular and protein surface features to capture biophysically relevant interaction information. Extensive experiments on six benchmark datasets demonstrate that SMCL-DTA achieves highly competitive performance under conventional warm-start evaluations while exhibiting robust generalization capability across multiple scaffold-based cold-start settings, providing an effective framework for realistic drug-target affinity prediction. Ablation studies confirm the effectiveness of each component, highlighting the synergistic benefits of our integrated design. Furthermore, wet-lab experiments involving virtual screening against the SOAT1 target validate the practical utility of SMCL-DTA in identifying potential bioactive compounds. These results collectively highlight the potential of SMCL-DTA as a powerful and generalizable tool for computational drug discovery.