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

Area of Science:

• Food Science
• Computational Chemistry
• Biotechnology

Background:

• Bitter peptides (BP) pose challenges in food and pharmaceutical development.
• Accurate prediction of BP activity and thresholds is crucial for product quality and consumer acceptance.
• Existing methods for BP identification and quantification are often time-consuming and resource-intensive.

Purpose of the Study:

• To develop and validate computational models for predicting bitter peptide (BP) activity and bitterness thresholds.
• To identify key physicochemical determinants contributing to bitterness.
• To facilitate high-throughput screening and assessment of BPs.

Main Methods:

• Development of two complementary computational models: XGBoost-BP for activity prediction and BOSS-BPT for threshold prediction.
• Utilized SHapley Additive exPlanations (SHAP) for feature importance in XGBoost-BP.
• Employed selection-frequency and coefficient analysis for BOSS-BPT.
• Validated predicted BPs using molecular docking and sensory evaluation.

Main Results:

• XGBoost-BP achieved high performance in predicting bitter activity (AUC=0.992 in cross-validation, 0.930 in test set).
• BOSS-BPT accurately predicted bitterness thresholds for various peptide lengths (R²=0.951-0.964).
• Both models identified eight novel bitter peptides, confirmed by molecular docking and sensory evaluation.

Conclusions:

• The developed computational models offer efficient and reliable tools for bitter peptide (BP) prediction.
• These models enable high-throughput screening and threshold assessment, significantly advancing food and pharmaceutical research.
• The findings provide insights into the physicochemical basis of bitterness, aiding in the design of palatable products.