Abstract
Plant-based proteins offer sustainable alternatives to animal sources, yet their lower digestibility remains a critical barrier to widespread applications. Current digestibility assessment methods require days of analysis and gram-scale samples, creating significant bottlenecks in protein optimization workflows. This study developed an ensembled deep learning framework that transforms digestibility prediction from a resource-intensive process to a rapid, minimal-sample assessment. By systematically characterizing 23 diverse plant protein isolates across multiple physicochemical dimensions, we trained a feedforward neural network based on augmented data. Our model identified α-helix content, random coil content, and solubility as key digestibility indicators. This insight enabled the construction of a streamlined three-feature model that reduced assessment time by 80% while requiring only one-hundredth of standard sample amounts. When validated against independent published datasets, the model achieved rational prediction accuracy, with an R(2) = 0.91. These findings establish a transformative framework for accelerating plant protein development, enabling rapid screening of novel sources and targeted modification strategies to enhance nutritional bioavailability, ultimately advancing sustainable food system transitions.