A multi-task learning framework with a frozen protein language model for unified prediction of food-derived peptide functions
Food Bioscience · 2026-10 · 110396
DOI: 10.1016/j.fbio.2026.110396 ↗Kurz gefasst
Fachartikel in der Zeitschrift Food Bioscience, 2026-10. Verfasst von Minh Quoc Ha und 5 weiteren Autor:innen. Die Arbeit nennt 69 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Food-derived bioactive peptides often need to meet efficacy, sensory, and safety requirements simultaneously. However, most computational predictors target individual activities, while existing multi-task models offer limited insight into joint learning. We developed the Multi-Task Learning Peptide Predictor (MTL-PepPred), a unified framework predicting 21 peptide functions. MTL-PepPred integrates a frozen protein language model, learnable embeddings, a Transformer, a convolutional neural network, and task-uncertainty weighting. Freezing the protein language model reduced trainable parameters by 86% (755.7 to 104.7 million), with an accuracy loss of less than one percentage point relative to full fine-tuning. We also curated a balanced signal-peptide dataset of 6,826 sequences, including 3,413 positives, extending the UniDL4BioPep benchmark to 21 tasks. Across five independent runs, MTL-PepPred achieved 87.4 ± 0.2% accuracy, 92.8 ± 0.3% area under the receiver operating characteristic curve (AUC), 92.2 ± 0.4% area under the precision-recall curve, and 73.7 ± 0.5% Matthews correlation coefficient. It matched or outperformed the published task-specific predictor for 13 of the 18 tasks with an available task-specific predictor, and both multi-task predictors on the two antimalarial tasks. Signal-peptide AUC was 99.9% on its own test split but 76.8% on an independent benchmark of experimentally confirmed non-secretory proteins, suggesting a dataset-specific screen rather than general signal-peptide recognition. By predicting 21 functions in one pass, MTL-PepPred enables prioritisation of multifunctional peptides before experimental validation, reducing stored parameters by 95% relative to 21 single-task models. A 22nd activity was incorporated by training a new classification head without affecting predictions for the original 21 tasks.
Zitieren
APA
Ha, M. Q., Son, H. X., Kam, R., Seale, B., Bang-Berthelsen, C. H., & Le, T. T. (2026). A multi-task learning framework with a frozen protein language model for unified prediction of food-derived peptide functions. Food Bioscience, 110396. https://doi.org/10.1016/j.fbio.2026.110396
Vancouver
Ha MQ, Son HX, Kam R, Seale B, Bang-Berthelsen CH, Le TT. A multi-task learning framework with a frozen protein language model for unified prediction of food-derived peptide functions. Food Bioscience. 2026:110396. doi:10.1016/j.fbio.2026.110396
BibTeX
@article{Ha2026,
author = {Ha, Minh Quoc and Son, Ha Xuan and Kam, Rothman and Seale, Brent and Bang-Berthelsen, Claus Heiner and Le, Thao T.},
title = {A multi-task learning framework with a frozen protein language model for unified prediction of food-derived peptide functions},
journal = {Food Bioscience},
year = {2026},
pages = {110396},
doi = {10.1016/j.fbio.2026.110396}
}
EUVIMED