Abstract
Plant peptide-protein interactions (PepPI) play a crucial role in plant growth, development, immune regulation, and environmental adaptation. However, existing computational methods still face several challenges in PepPI prediction. First, most methods fail to adequately integrate multimodal information such as sequence, structure, and disorder properties, leading to inadequate characterization of complex interaction patterns. Second, existing models have difficulty in capturing cross-dependent features between peptides and proteins, limiting the prediction performance. Finally, the lack of a unified framework capable of integrated modeling from both global and local perspectives leads to a still imperfect understanding and prediction of PepPI. To this end, we propose a multiple characterization framework based on cross-modal feature fusion-MultiRepPI-for efficient prediction of plant PepPI. In this framework, we innovatively introduce several key modules aimed at comprehensively improving the representation of peptide and protein features. First, a cross-modal encoding module (CME) is designed by fusing convolutional neural networks, recurrent neural networks, and feature enhancement mechanisms, which is capable of extracting multi-scale deep features from peptide and protein sequences, and thus better capturing their interactions at different levels. Secondly, bi-directional attention and PepPI gating mechanisms were used to design the cross-modal attention module (CMA), which deeply mines the key interaction patterns between peptides and proteins, and enhances the ability of the model to focus on important binding sites. Finally, the disordered feature extraction (DFE) module was designed to specifically identify disordered regions of plant proteins and extract dynamic features to further enhance the accuracy of plant PepPI prediction. A systematic experimental evaluation of MultiRepPI was conducted on a benchmark dataset. The experimental results show that MultiRepPI provides a significant improvement in both prediction performance and binding residue recognition compared to existing state-of-the-art methods. This framework provides a reliable tool for efficient prediction of plant PepPIs, while laying a solid foundation for plant biology research and peptide drug development.