Machine learning-enabled discrimination and key feature identification of functional whey protein ingredients based on multidimensional compositional, physicochemical, and structural properties

Commercial functional whey protein ingredients are highly heterogeneous, making reliable classification challenging. In this study, 16 commercial whey protein ingredients, including microparticulated (MWP), thermostable (TWP), and gel-forming whey proteins (GWP), were characterized using 41 compositional, physicochemical, functional and structural variables. Eleven machine-learning algorithms were evaluated using repeated nested group-based cross-validation. XGBoost showed the strongest overall performance, achieving a mean accuracy of ~0.972 and a macro-F1 score of ~0.970, while permutation testing supporting above-chance discrimination (p = 0.001). Feature-importance and SHAP analyses consistently identified β-lactoglobulin, α-lactalbumin and secondary-structure-related variables as important contributors to classification. XGBoost retained high classification performance when only four features were retained within each outer training fold, suggesting that fewer measurements could suffice for preliminary classification. Although larger sample sets and further external validation are required before industrial application, these findings demonstrate the feasibility of identifying multivariate signatures associated with predefined functional whey protein classes.