Abstract
To shorten the complex and time-consuming process of the identification method of the traditional food angiotensin-I-converting enzyme (ACE-I) inhibitory peptides, we propose AHTPeptideFusion based on a segmented fusion with the protein language model and deep learning. The statistical analysis found that hydrophobic amino acids, N-terminal valine is a dominant amino acid in the activity of ACE-I inhibitory peptides. In 12 machine learning (ML) algorithms, the transformer outperformed the other 11 models, with the best performance in predicting short and medium peptides. In the external dataset, AHTPeptideFusion fused by transformer and random forest (RF) showed excellent performance (accuracy > 0.9) in predicting ACE-I inhibitory peptides with lengths ranging from 2 to 15 amino acid residues and different activity distributions, and the reliability and accuracy of AHTPeptideFusion was demonstrated by synthetic peptide and ACE-I inhibition experiments. In addition, hydrogen bonding and electrostatic interaction between 4 synthetic peptides and active residues of ACE-I were found by molecular docking. To further explore the ACE-I inhibitory peptides from animal-derived foods, we established an automated pipeline consisting of the trinity of proteomics, virtual enzymatic digestion and AHTPeptideFusion, and tapped the ACE-I inhibitory peptide released from royal jelly after digestion in the gastrointestinal tract. In conclusion, this computational pipeline will become a powerful screening tool for active peptides from animal-derived foods, which can help food scientists accelerate the mining and design of active peptides from animal-derived foods. Overall, AHTPeptideFusion will be a powerful ACE-I inhibitor peptide prediction tool, it can help food scientists accelerate the mining and design of ACE-I inhibitory peptides.