• RSM gave the results with 90.1% digestibility, 6.6 bitterness intensity and 1.00 EAAI • Machine learning (ML) model captured non-linear relationships better than RSM • Bitterness intensity dropped 95% by ML versus RSM with improvement of digestibility • ML surpassed RSM in balancing proteins qualities by reducing bitter peptides The growing demand for plant-based proteins in children’s nutrition faces significant challenges due to low digestibility, undesirable taste profiles, and imbalanced amino acid composition. Traditional optimization methods like response surface methodology (RSM) often fail to navigate this complex, non-linear landscape, leaving a gap for more powerful predictive tools. In this study, RSM-Central Composite Design (RSM-CCD) was systematically compared with a neural network-driven multi-objective optimization strategy to optimize formulation and enzymatic hydrolysis parameters. The RSM-CCD approach yeilded a substrate with 90.08% digestibility and a1.00 essential amino acid index (EAAI), but a high bitterness intensity (BI) of 6.6 . In contrast, the neural network model, trained over 600 epochs with the optimal solution emerging at epoch 540, achieved superior overall performance, reaching 90.64% digestibility, an EAAI of 1.18, and a remarkably low BI of 0.29. This demonstrates the neural network's ability to capture complex interactions missed by RSM. To validate these outcomes, peptidomics analysis of the hydrolysates confirmed that the neural network-optimized process fundamentally altered the hydrolysis pattern, effectively eliminating the generation of key bitter peptides. Collectively, this study not only delivers a highly digestible and palatable plant-based protein substrate suitable for pediatric nutrition, but also establishes machine learning as a transformative approach for addressing longstanding multi-attribute optimization challenges in food product development.
Liu et al. (2026) studied this question.