Accurate prediction of drug-target interactions (DTI) is critical for accelerating drug discovery. In this study, we propose deep contrastive learning (DeepCL), a novel deep contrastive learning framework that significantly improves DTI prediction accuracy, particularly in low-coverage scenarios. Unlike traditional methods, DeepCL introduces a generalized sigmoid activation function to resolve numerical underflow issues and employs a margin-based contrastive loss to enforce better separation between interacting and non-interacting pairs. Extensive experiments on three benchmark datasets (Davis, BindingDB, and BIOSNAP) demonstrate that DeepCL outperforms state-of-the-art methods in both standard and zero-shot settings, achieving superior AUPR and AUROC scores. Methodologically, DeepCL constructs a dual-pathway architecture: it leverages the ESM-2 protein language models to capture rich contextual protein features and Morgan fingerprints for precise molecular structural representation. These heterogeneous features are aligned in a shared latent space via modality-specific projectors. DeepCL contributes a robust, scalable, and numerically stable solution for DTI prediction.
Li et al. (Thu,) studied this question.