PepPFN: protein-peptide binding residues prediction via pre-trained module-based Fourier Network

Molecular activities in vivo, such as protein-peptide binding residues, contribute to understanding disease mechanisms and discovering targeted drugs, but are often limited by the cost of wet experiments. Therefore, this paper proposes a Pre-trained module-based Fourier Network inspired by Fourier Neural Operators for predicting protein-peptide binding residues (PepPFN). First, the model captures highlatent protein representation and implicit information through Bert and ESM, respectively. Subsequently, the contrastive learning module is used to optimize the representation of binding residues to mitigate the effects of data diversity. Furthermore, the order and frequency information of the residual is obtained by strengthening the model with Fourier network. Comprehensive analysis of multiple tasks and protein visualization results demonstrates that PepPFN outperforms state-of-the-art methods on multiple performance metrics. Datasets and the source code is available at https://github.com/xueleecs/PepPFN.git.