AI-assisted screening of soybean varieties: Linking protein composition to tofu quality

Precise control of soy-protein gelation is essential for consistent tofu manufacture, yet current cultivar-selection methods remain slow, empirical, and prone to batch-to-batch variability. This study introduces the first AI-assisted platform specifically designed to translate SDS-PAGE protein fingerprints from a diverse panel of 170 soybean lines into rapid, objective tofu quality predictions. Leveraging computer vision for the digitization of protein subunits, coupled with an optimized support vector machine (SVM) classifier, our workflow bridges a critical gap between biochemical analysis and functional food quality. The model classified seeds into three tofu-quality classes-firm, high-yield, and high-moisture-through hierarchical clustering of multiple quality metrics (firmness, yield, moisture content, protein content, and textural attributes). These labels reflected the predominant characteristics of each class, and the approach achieved 82 % accuracy, outperforming k-nearest neighbor, decision-tree, random-forest, and gradient-boosting algorithms. A feature-weight analysis revealed that 11S A₃ and B subunits are strong determinants of tofu gelation and texture, whereas 7S α' subunits favored water retention and softer textures. Collectively, the workflow compresses months of benchtop tofu trials into minutes of image analysis, offering breeders an objective metric for genotype ranking and processors a chemistry-based specification tool. We consider this framework as a pilot-scale predictive model, establishing a foundation for future large-scale applications. Beyond tofu, the platform provides a transferable blueprint for data-guided design of plant-protein hydrocolloids, enabling rapid, clean-label texture engineering in the expanding plant-based sector.