Abstract
Motivation
Accurate Drug–Target Interaction (DTI) prediction is critical for drug discovery. While deep learning advances this field, the fundamental lack of verified negative interactions inevitably causes biased risk estimation. Although formulating DTI prediction as a Positive-Unlabeled (PU) learning problem is theoretically sound, directly applying PU learning to deep models suffers from early-stage optimization instability and decision boundary ambiguity.Results
We propose ProDTI, a prior-guided PU learning framework integrating domain knowledge into the PU objective. We construct a hybrid prior combining explicit biochemical similarities with implicit semantic representations from large-scale pre-trained models. This prior drives two synergistic mechanisms: soft-label distillation to stabilize the optimization trajectory, and hard-anchor regularization to explicitly penalize reliable negatives and sharpen decision boundaries. Evaluated on five benchmarks, ProDTI consistently outperforms state-of-the-art baselines, exhibiting superior robustness in challenging cold-start scenarios and providing a reliable tool for computational drug discovery.Availability and Implementation
Source code and data are available at https://github.com/WangYu-AI4SCI/ProDTI.