Introduction
Protein–peptide interactions (PepPIs) are essential to a wide range of biological processes, including signal transduction, gene regulation, cellular homeostasis, and metabolic modulation. Recent advances in protein language models and deep learning have promoted the development of sequence-based PepPIs prediction methods. However, most existing approaches mainly focus on pair-level interaction discrimination and remain limited in modeling pair-specific residue-level binding patterns. In particular, the binding residues of a protein or peptide are not fixed sequence attributes, but depend strongly on the specific interaction partner and the corresponding molecular context. To address these challenges, we propose ResGrid-PepPI, an accurate and interpretable framework for peptide–protein interaction prediction and residue-level sequence labeling. In this framework, protein and peptide sequences are encoded using ProtT5-derived semantic representations, and dual-tower sequence representation learning is combined with dual-branch residual fusion to capture both global sequence matching information and pair-specific interaction signals. For residue-level prediction, binding site identification is formulated as a pair-specific sequence labeling task, in which binding or non-binding labels are assigned to each valid residue in the protein and peptide sequences. Cross-chain context modeling, directional pair-grid aggregation, and side-specific residue labeling heads are further introduced to characterize fine-grained residue-pair interaction patterns. Experimental evaluations demonstrate that ResGrid-PepPI achieves superior predictive performance and robust generalization on benchmark datasets, independent test datasets, and CD-HIT cold-start datasets. In addition, virtual alanine scanning and binding affinity prediction experiments verify the model's potential in identifying functional residues and characterizing binding strength differences across peptide–protein complexes. These findings demonstrate that ResGrid-PepPI provides a reliable computational framework for PepPIs prediction, residue-level interpretation, and downstream peptide drug discovery.
