Haoshuang Wu, Yu Deng, Zhen Li, Mingming Qi, Yanshi Wei, Linlin Zhuo, Quan Zou, Wenqian Zhang
Naturally occurring in animals, plants, and other organisms, peptides exhibit strong antibacterial activity and hold promise as alternatives to conventional antibiotics. However, evaluating peptide properties such as toxicity, antibacterial activity, and anticancer potential remains a critical step in the drug development. Current peptide property prediction methods often rely on single-scale sequence or structural information, failing to capture the complex multi-scale data necessary for accurate predictions. To address this challenge, we propose UniPept, a novel model for peptide property prediction that integrates atomic-level and residue-level features to generate robust peptide representations. We first extract atomic two-dimensional(2D) descriptors and atom-atom pairwise geometry from the peptide sequence to derive atomic-level geometric features. Residue-level features are obtained using a protein language model, while a residue dictionary provides amino acid index profiles and residue-residue geometric relationships to construct residue-level geometric representations. Finally, atomic and residue features are integrated via a gated attention mechanism. This multi-scale fusion strategy enhances peptide representations. Extensive experiments demonstrate that the proposed method accurately predicts peptide properties and that UniPept outperforms competing approaches in interaction and binding affinity prediction tasks. These findings underscore the potential of UniPept in advancing peptide drug development and providing insights into biological and clinical applications.