Unraveling the physicochemical differences among Osborne protein classes via bioinformatics and AI

Osborne fractionation remains a cornerstone in food science for categorizing seed storage proteins (SSPs), yet molecular distinctions among the classes remain unclear. This study employs a computational framework integrating structural modeling, AI (artificial intelligence)-driven classification, and molecular dynamics (MD) simulations to elucidate these underlying physicochemical differences. Using a dataset of 1039 SSPs from 215 species, sequence and structural-based features were extracted and compared to identify class-specific characteristics, such as low hydrophobic patch area of albumins. Machine learning (ML) classifiers, including binary support vector machines and graph convolutional networks were trained on these features, achieving validation and test accuracies ranging from 96.0 % to 100.0 %. Model interpretations using SHapley Additive exPlanations and saliency mapping revealed key distinguishing features between albumin/prolamin and globulin/glutelin, respectively. For the albumin and prolamin classes, physicochemical feature comparisons and ML classifiers identified factors underlying their solubility differences, such as the low abundance of charged residues in prolamins. On the other hand, although certain features, such as mean surface electric potential, distinguished globulins from glutelins, no clear association was found between these features and experimental solubility trend. Notably, saliency analysis of globulins and glutelins highlighted loop and helical regions outside the conserved β-barrel motifs, where compositional differences in glutamic acid, glycine, serine, and glutamine residues were observed. MD simulations explored solvent-specific conformational changes in representative SSPs, with all-atomic simulations performed on single monomers and coarse-grained simulations conducted with multiple monomers. For 2S soy albumin and 19 kDa maize prolamin, distinct hydrogen bonding patterns was observed during their adaptations to 70 % ethanol environment, and the expected aggregation tendency was reproduced in the multiple-monomer simulation. Taken together with the highlighted features in ML classification, these results suggest that the experimental solubility of albumins and prolamins can be explained at the monomeric level. However, for pea legumin A (globulin) and rice glutelin A1, no clear differences in structural and aggregation dynamics were observed, and monomeric properties alone failed to account for their distinct solubility. These findings suggest that glutelin insolubility is likely dictated by inter-protein disulfide networks rather than intrinsic monomeric characteristics, aligning with previous experimental observations.