Twin cross contrastive learning with multi-modality fusion for drug-target affinity prediction

Accurate prediction of drug-target binding affinity (DTA) can provide valuable insights for accelerating drug discovery and repositioning. While deep learning has demonstrated remarkable progress in facilitating DTA prediction, most existing methods fail to adequately learn informative representations for drugs and targets from multifaceted perspectives. The extraction of molecular intrinsic properties and topological information from interaction networks, while enabling their synergistic integration, remains a challenge in this task. To this end, Twin Cross Contrastive Learning with multi-modality fusion framework (TCCL) is proposed for DTA prediction. Firstly, a dual-stream encoder is employed to capture multi-modal features encompassing semantic and structural information at the molecule scale, enabling comprehensive characterization of drugs and targets. Additionally, the relationships between drugs and targets are captured at the network scale by leveraging the topological structure information of the drug-target bipartite graph. Furthermore, a twin cross contrastive learning architecture is designed to separately contrast the semantic and structural representations learned from the molecule-scale with the network-scale representations, aiming to enhance the discriminative ability of TCCL. Extensive experiments show that TCCL outperforms state-of-the-art methods on two benchmark datasets, which demonstrates its effectiveness DTA prediction. Moreover, case study implemented in real-world scenarios suggests that TCCL has notable generalization ability, which can serve as an effective and powerful tool for DTA prediction.