Uncertainty-quantified deep learning enables reliable protein-drug interaction prediction

Accurate prediction of protein-drug interac- tions(PDIs) is crucial for drug discovery, yet traditional models lack uncertainty quantification essential for high-stakes phar- maceutical decisions. We develop a Bayesian Neural Network framework with a deep architecture(512 − 256 − 128 neurons) optimized using AdamW with adaptive scheduling. Training on 2000 protein-drug pairs with 12 physicochemical features, our model achieved 98% accuracy, 0.9987 ROC-AUC, and 0.9972 PR-AUC. Critically, it demonstrated robust uncertainty quantification(mean uncertainty: 0.0136) with 100% accuracy on high-confidence predictions. This framework enables reliable compound prioritization for experimental validation by quanti- fying prediction confidence, advancing trustworthy AI in drug- discovery.