Predicting multifunctional peptides based on a multi-scale ResNet model combined with channel attention mechanisms

Peptides are biomolecules composed of multiple amino acid residues connected by peptide bonds, which are widely involved in physiological and biochemical processes in organisms and exhibit diverse functions. In previous studies, the focus was primarily on single-functional peptides. However, research trends indicate that an increasing number of multifunctional peptides are being identified and discovered. To address this challenge, we proposed a deep learning method based on multi-scale ResNet as the backbone combined with a channel attention mechanism (called MSRC) for the identification of multifunctional peptides. Furthermore, the data imbalance problem was solved through the comprehensive use of online data augmentation and confidence-based weighted loss functions. Experimental results demonstrated that the proposed MSRC method achieved an accuracy of 0.688 with an absolute true rate of 0.619. Notably, in predicting minority class peptides such as AEP, AHIVP, and BBP, the MSRC model exhibited heightened sensitivity, showcasing its exceptional capability in addressing issues related to minority classes. By enhancing the precision in identifying and predicting multifunctional peptides, the MSRC method was poised to contribute significantly to advancements in drug discovery, disease treatment, and biotechnology.