ProDTI: Prior-Guided Positive-Unlabeled Learning for Robust Drug-Target Interaction Prediction.

ProDTI achieves up to 15% higher accuracy than existing methods in predicting drug-target interactions using a prior-guided PU learning framework.

• Why it matters: Accurate DTI prediction is essential for efficient drug discovery, but the lack of verified negative interactions causes bias and instability in deep learning models.
• What they did: The authors developed ProDTI, a hybrid prior that combines biochemical similarities and semantic representations, integrating it into a PU learning framework with mechanisms for stabilization and boundary sharpening.
• The result: ProDTI consistently outperforms state-of-the-art models across benchmarks, especially in cold-start scenarios, enabling more reliable computational drug discovery.