A structure-sequence synergistic learning framework for prioritizing functional peptides from food protein hydrolysates.

Abstract

Food-derived peptide libraries generated by protein hydrolysis are heterogeneous, making screening costly and difficult to scale across assay endpoints. Existing computational approaches rely on sequence information and often fail to capture conformational determinants or maintain robust performance under limited and partially disjoint data settings. We present a structure-sequence synergistic learning framework for prioritizing functional peptide candidates from food protein hydrolysates. The framework combines a pre-trained protein language model for contextual sequence encoding with a geometry-aware graph neural network for three-dimensional structural representation, and fuses both modalities for task-specific prioritization. Using an LC-MS/MS-derived dataset spanning five screening tasks, 6224 labeled peptide-task instances were constructed through docking-informed surrogate supervision. The proposed model achieved an average balanced accuracy of 0.870, exceeding sequence-only and structure-only baselines with average values of 0.838 and 0.799, respectively. The gain was more evident in data-limited tasks, supporting the complementary value of geometric features beyond linear sequence representations. Sample-level rescue showed that structural information corrected more sequence-only errors than it introduced across all tasks. In external peptide-disjoint evaluations on 1383 unseen sequences from three independently collected datasets, the framework retained measurable ranking ability, with PR-AUC values of 0.717, 0.641, and 0.675. Conformational sensitivity analysis supported stable candidate ranking across low-energy conformers, and preliminary assays provided candidate-level validation for WWF and YYW. Overall, this study provides a structure-sequence computational framework for narrowing LC-MS/MS-derived peptide search spaces before synthesis and activity assays.