Multi-Functional Peptide Discover with Amino Acid-level Fusion of Sequence Information and Structure Feature

Multifunctional peptide identification is a challenging task due to the complex sequence-structure-function relationships. Existing methods often rely on single-modal features, limiting their ability to comprehensively model the intricate interplay between sequence and structural information. In this study, we propose MFPep-AAF, a novel amino acid (AA)-level multimodal fusion framework for multi-functional peptide identification by integrating sequential information and structural feature. MFPep-AAF harnesses a cross-modal attention mechanism to dynamically fuse AA-level semantics from a fine-tuned protein language model and AA-level structural constraints from a graph attention network. This fine-grained fusion strategy enables the model to effectively capture both local residue interactions and global sequence-structure relationships for functional prediction. Experimental results on benchmark datasets demonstrate that MFPep-AAF achieves state-of-the-art performance in terms of absolute true metric. These results underscore the advantages of integrating multimodal features, providing a robust and reliable framework for multifunctional peptide prediction.