Developing machine learning-driven QSAR models for predicting bitter activity and bitterness thresholds of oligopeptides.

To improve bitter peptide (BP) prediction, two complementary computational models were developed. The XGBoost-BP classifier achieved strong performance in identifying bitter activity (AUC = 0.992 in cross-validation and 0.930 in the test set) and revealed key physicochemical determinants through SHapley Additive exPlanations (SHAP) analysis. For quantitative evaluation, the BOSS-BPT model accurately predicted bitterness thresholds of di-, tri-, and tetrapeptides (R2 = 0.951-0.964), and identified key variables via selection-frequency and coefficient analysis. Both models identified eight previously uncharacterized BPs, which were further validated through molecular docking and sensory evaluation. These models provide efficient and reliable tools for high-throughput BP screening and threshold assessment in food and pharmaceutical systems.