Path-Induced Regularized Graph Contrastive Learning for Drug-Target Interaction Prediction

Le résumé fourni par la source

Existing graph neural network–based methods have achieved notable progress in drug–target interaction (DTI) prediction; however, most of them primarily rely on local neighborhood information and show limited capability in modeling higher-order structural relationships with explicit semantic meaning in biological networks. To address this limitation, we propose a Path-Induced Regularized Graph Contrastive Learning (PIRGCL) framework for DTI prediction on heterogeneous drug–protein graphs. Within a graph contrastive learning paradigm, PIRGCL introduces path-induced regularization as a structural constraint, which enhances the consistency of node representations while explicitly guiding the model to capture higher-order association patterns between drugs and targets. Experimental results on multiple DTI datasets with varying scales and sparsity levels demonstrate that the proposed method achieves consistently competitive performance in terms of AUC and AUPR. For example, on the IDTI dataset, PIRGCL achieves an AUC of 0.958 and an AUPR of 0.961, outperforming several representative baseline methods for DTI prediction. Notably, PIRGCL maintains strong ranking ability and positive-sample discrimination under highly imbalanced settings, indicating good generalization and robustness across heterogeneous biological categories. Parameter sensitivity analysis further validates the stability of the model, while molecular docking case studies provide structural evidence supporting the biological plausibility of the predicted drug–target interactions. The docking results show that the predicted drug–target pairs form reasonable binding conformations within protein binding pockets and exhibit typical hydrogen-bond and hydrophobic interaction patterns.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.