Language model encoded multi-scale feature fusion and transformation for predicting protein-peptide binding sites

Protein-peptide interactions serve as a pivotal and indispensable role in diverse biological functions and cellular processes. Although recent studies have begun to employ language models for predicting protein-peptide binding sites (PPBS), the majority of previous approaches have persisted in utilizing intricate sequence-based feature engineering or incorporating costly experimental structural information. To overcome these limitations, we develop a novel sequence-based end-to-end PPBS predictor using deep learning, named Language model encoded Multi-scale Feature Fusion and Transformation (LMFFT). The proposed model starts with a single protein language model for comprehensive multi-scale feature extraction, including residue, dipeptide, and fragment-level representations, which are implemented by the dipeptide embedding-based fragment fusion and further enhanced through the dipeptide contextual encoding. Moreover, multi-scale convolutional neural networks are applied to transform multi-scale features by capturing intricate interactions between local and global information. Our LMFFT achieves state-of-the-art performance across three benchmark datasets, outperforming existing sequence-based methods and demonstrating competitive advantages over certain structure-based baselines. This work provides a cost-effective and efficient solution for PPBS prediction, advancing revealing the sequence-function relationship of proteins.