Multi-domain based heterogeneous network for drug-target interaction prediction

Recent studies have emphasized the importance of computational approaches in predicting drug-target interactions (DTIs) for drug discovery and chemogenomics studies. Precise prediction of DTIs plays a crucial role in exploring a vast space of drug compounds. However, existing in silico approaches suffer from the following limitations. Firstly, most molecular representation learning methods neglect the sub-structural characteristics of drug-target pairs (DTPs), resulting in challenging interpretations of the predictions. In addition, many models focus on limited in-domain datasets, exhibiting unsatisfactory results when applied to predict new DTIs. To mitigate these defects, we present MHNF-DTI here, a Multi-domain based Heterogeneous Network Framework designed for the integrated prediction and interpretation of DTIs with high interpretability and generalization ability. Importantly, the novel framework utilizes a transformer encoder that integrates multilayer graph attention networks, effectively capturing the sub-structural properties of drug compounds and target sequences, make it able to adapt to the shared structures of different DTPs while enhancing the molecular representation capabilities. Additionally, to improve the generalization ability of the model and mitigate the potential hidden ligand bias pitfalls, a new multi-domain label reversal dataset is constructed for training. Experimental results demonstrated that the proposed MHNF-DTI improved DTI prediction performance compared to the existing state-of-the-art baselines.