DeepPepPI: A deep cross-dependent framework with information sharing mechanism for predicting plant peptide-protein interactions

Abstract

Peptide-protein interactions (PepPIs) play a crucial role in various fundamental biological activities in plants. As traditional experimental technologies are usually time-consuming and labor-intensive, it is indispensable to develop computational methods for PepPI identification, which has potential applications across diverse fields including food safety and peptide drug discovery. However, existing methods have received relatively little attention in adequately representing the interaction information between peptides and proteins, consequently limiting the prediction performance. To this end, we present DeepPepPI, a novel deep cross-dependent framework for accurate prediction of plant PepPIs. Concretely, a data-driven context-embedded representation (DCR) module is developed as the peptide feature extractor that can capture rich contextual semantic information, even from short sequences. To fully characterize inherent properties of proteins, we propose a bi-level self-correlation search (BSS) module, which integrates the primary sequence and secondary structure into a unified space to learn their potential relationships. In addition, a cross-dependent feature integration (CFI) module with information sharing mechanism is introduced, aimed at providing a comprehensive feature representation to portray the intricate interaction patterns between peptides and proteins from a global perspective. Extensive experimental results conducted on the benchmark datasets show that DeepPepPI achieves superior prediction performance compared to other state-of-the-art methods.