Predicting protein adsorption across sources

Proteins need to reach newly formed bubble surfaces quickly when making foamed foods. In this project, I bring together 47 adsorption curves from 12 protein sources and 14 studies, harmonize the measurements, and describe each curve with a compact kinetic model. I then compare a composition-based kinetic predictor with TabPFN, a machine-learning model, by predicting protein sources excluded entirely from training. Both approaches predict native seed-storage proteins from composition with similar accuracy. The useful finding is where prediction stops transferring: processing, colloidal state, and different biological protein classes weaken the relationship. The dataset and scripts offer a starting point for screening plant proteins, while predicting foam stability still requires additional measurements.