Uncertainty Quantification for Drug -Target Interaction and Binding Region Prediction using Bayesian Deep Learning

Abstract

Reliable prediction of drug–target interactions (DTIs) under distribution shift remains a key challenge in computational drug discovery. In this study, we propose a binding region–aware DTI prediction framework that jointly outputs interaction scores, binding region probabilities, and predictive uncertainty. By integrating sequence-based binding region learning with pretrained molecular representations and Monte Carlo dropout–based uncertainty quantification, the proposed model enables confidence-aware predictions even when structural information is limited. Across multiple DTI benchmarks, the model achieves performance comparable to or better than existing methods and demonstrates robust generalization under out-of-distribution settings. Uncertainty-aware analysis further shows that the proposed framework effectively distinguishes high-confidence predictions from unreliable ones. Overall, these results indicate that jointly incorporating binding region information and uncertainty quantification is a practical strategy for improving the reliability and interpretability of DTI prediction.