SCL-DTI: Semi-Supervised Contrastive Learning for Drug-Target Interactions Prediction

The discovery and development of new drugs are critical yet challenging tasks in the pharmaceutical industry, often hindered by the time-consuming and costly nature of traditional drug development processes. Predicting drug-target interactions (DTIs) is a pivotal step in accelerating these processes, reducing costs, and increasing the success rate of identifying effective and safe drugs. This paper introduces a novel method, SCLDTI, which leverages semi-supervised contrastive learning to predict new DTI links, thereby enhancing the efficiency of drug discovery. Our approach addresses the limitations of existing DTI prediction methods, which include reliance on extensive labeled data and the inability to effectively utilize unlabeled data. By employing a data augmentation strategy based on random edge dropout and a multi-task training strategy that optimizes both the traditional prediction task and the semi-supervised learning task, SCLDTI significantly improves the prediction of drug-target interactions. We compare our method with existing approaches, including ligand-based models, target-based models, and other deep learning methods, demonstrating superior performance in DTI prediction. Our contributions not only propose a new method for DTI prediction but also open avenues for applying semi-supervised contrastive learning in drug discovery and development.