Enhancing drug-target interaction prediction with graph representation learning and knowledge-based regularization

Accurately predicting drug-target interactions (DTIs) is crucial for accelerating drug discovery and repurposing. Despite recent advances in deep learning-based methods, challenges remain in effectively capturing the complex relationships between drugs and targets while incorporating prior biological knowledge.

We introduce a novel framework that combines graph neural networks with knowledge integration for DTI prediction. Our approach learns representations from molecular structures and protein sequences through a customized graph-based message passing scheme. We integrate domain knowledge from biomedical ontologies and databases using a knowledge-based regularization strategy to infuse biological context into the learned representations.

We evaluated our model on multiple benchmark datasets, achieving an average AUC of 0.98 and an average AUPR of 0.89, surpassing existing state-of-the-art methods by a considerable margin. Visualization of learned attention weights identified salient molecular substructures and protein motifs driving the predicted interactions, demonstrating model interpretability.

We validated the practical utility by predicting novel DTIs for FDA-approved drugs and experimentally confirming a high proportion of predictions. Our framework offers a powerful and interpretable solution for DTI prediction with the potential to substantially accelerate the identification of new drug candidates and therapeutic targets.