A multimodal DTA prediction method based on triple-view contrastive learning

Abstract Drug-target affinity (DTA) prediction is a crucial step in drug discovery, facilitating the acceleration of lead compound screening and drug repurposing, significantly reducing costs and shortening new drug development timelines. Although deep learning methods have made significant advancements, current approaches often rely on single-view or unimodal features and fail to effectively explore the potential relationships between drugs and targets, as well as between targets and targets. As a result, these methods struggle to capture the complex multi-view features of drug-target interactions, leading to limited representation power and suboptimal prediction accuracy. To address this, we propose MTCL-DTA, a multimodal DTA prediction method based on triple-view contrastive learning. This method constructs three distinct molecular representation perspectives. First, molecular views are constructed by using the drug SMILES string and molecular graph and the target sequence and target molecular graph to capture the internal structural features of the molecules. Second, structural and semantic views are constructed based on a drug-target bipartite network to model the inter-molecular interaction features. Finally, multi-view contrastive learning is employed to optimize the view features, thereby achieving more accurate representations of drugs and target proteins and enhancing the prediction accuracy of DTA. Experiments on benchmark datasets demonstrate that MTCL-DTA achieved prediction accuracies with mean squared errors (MSE) of 0.172 and 0.126 on the Davis and KIBA datasets, respectively, improving by 10.88 and 10%, respectively, over the current state-of-the-art models. Meanwhile, the $$r_m^2$$ r m 2 values reach 0.795 and 0.809, with 13.73 and 2.92% improvements, respectively. Furthermore, case studies demonstrate that the model successfully predicted novel binding relationships between Doxorubicin, Cetuximab, and unknown proteins, further validating its practical application in real-world scenarios.