Interpretable Machine Learning of Ion-Associated Variation in Tomato Polyphenol Oxidase Activity across Developmental and Postharvest Stages

Polyphenol oxidase (PPO) activity varies with developmental stage, chemical treatment, and treatment concentration, making its response patterns difficult to characterize using single-factor comparisons. This study developed an interpretable analytical and machine-learning framework to investigate Catecholase and Cresolase activity in tomato across immature, mature, and post-harvest stages. The final dataset comprised 1,476 observations, including 837 Catecholase and 639 Cresolase measurements. All 492 enzyme × stage × treatment × concentration conditions contained three observations, and nine observed zero-activity measurements were retained in the Cresolase dataset. Descriptive analyses and concentration–response modelling revealed nonlinear and non-monotonic patterns that varied across enzymes and experimental conditions. Concentration-adjusted interaction analyses indicated different Stage × Treatment patterns between the two enzymes, although the estimated interaction effect sizes differed substantially. Chemical-descriptor comparisons showed that predictive performance depended on the representation of salt treatments, while anion-identity-only models performed poorly under the evaluated modelling and validation settings. A hierarchical modelling framework comprising common, stage-specific, and profile-specific concentration–response structures was evaluated using grouped five-fold cross-validation. The hierarchical G3 model achieved pooled out-of-fold (R^2) values of 0.9887 for Catecholase and 0.9873 for Cresolase. Relative to the G2 baseline, pooled RMSE decreased by approximately 3.83% and 3.00%, respectively, although the fold-wise consistency of improvement differed between enzymes. These results indicate that tomato PPO activity exhibits enzyme-specific and developmentally variable response patterns and that interpretable modelling can characterize these patterns while distinguishing predictive associations from causal effects. The findings apply to the evaluated treatments and experimental conditions and do not establish isolated causal effects of individual ions or physicochemical descriptors.