Construction of a predictive model for pea protein solubility evaluation and screening of processing-specific varieties using machine learning after interpretable optimization

This study aimed to examine the influence of physicochemical and nutritional indicators on the solubility of various pea varieties. It compared conventional chemometric techniques with four machine learning algorithms to identify the optimal regression model for predicting pea protein solubility. GridSearchCV optimized hyperparameters, while shapley additive explanations (SHAP) were employed to identify the key indicators affecting the model's predictions. Based on k-means cluster analysis (KCA), the analytic hierarchy process (AHP) was used to score specialized pea varieties for solubility in product processing. After hyperparameter optimization and cross-validation, the categorical boosting (CatBoost) model emerged as the best predictor of pea solubility, achieving a coefficient of determination (R 2 ) of 0.92 on the test sets. The key factors influencing pea protein solubility included the nitrogen solubility index (NSI), seed height, allergic protein content, proteins of 20 kDa, moisture content, and proteins of 14 kDa. Additionally, “ Ba Wan 1 ” was identified as the highest-rated pea varieties. These findings provide a theoretical foundation for evaluating the suitability of unknown pea varieties and support the sustainable development of the agricultural processing industries. ● Identified key indicators affecting the solubility of pea protein. ● Using SHAP to explain Catboost models. ● The R P 2 of the Catboost model after hyperparameter optimization is 0.92. ● Using AHP combined with K-means for weight analysis.