Universal deep learning strategy leveraging peptide atomic 3D structural information identifies diverse bioactive peptides from wheat germ protein hydrolysates

Existing deep learning strategies for identifying bioactive peptides (BAPs) are often limited by single-task models and shallow sequence-based features, which restricts their generalizability. This study introduced an auto-encoder based deep learning framework (BioPepAE) that leverages peptide atomic 3D information from AlphaFold 3 to enable recognition of multiple types of BAPs, including anti-hypertensive peptides (AHPs), anti-oxidant peptides (AOPs), and anti-aging peptides (AAPs). BioPepAE achieved high accuracy (94.65% for AHPs, 95.47% for AOPs, 92.86% for AAPs) and demonstrated strong generalization on independent tests. BioPepAE identified 8 AHPs, 13 AOPs, and 5 AAPs from wheat germ protein hydrolysates, and subsequent in vitro assays confirmed the bioactivity of randomly selected peptides, thus validating its predictive accuracy and practical utility. Moreover, BioPepAE requires no complex parameter tuning for identifying different BAPs. This study presents a robust and versatile framework for the universal and accurate identification of BAPs by integrating 3D structural information with deep learning.