Deep learning-guided identification of bacteriophage receptor-binding protein candidates for foodborne pathogen detection

Abstract

Foodborne pathogens including Salmonella spp., Escherichia coli and Listeria monocytogenes cause an estimated 600 million illnesses annually. Yet conventional detection methods remain slow, costly, or insufficiently specific for routine food safety surveillance. Phage receptor-binding proteins (RBPs) are attractive recognition elements for biosensors, but their extensive sequence diversity limits reliable computational identification. Here, we present a systematic open-source computational pipeline for identifying and structurally characterising high-confidence RBP candidates from phage genomes targeting these three priority pathogens. The pipeline integrates four stages: deep learning-based RBP prediction, protein structure prediction, structural homology validation, and exploratory molecular docking. Applied to a quality-controlled dataset of 247 complete phage genomes retrieved from the National Center for Biotechnology Information Nucleotide database (31,752 total protein sequences), PhageRBPdetect, built on the ESM-2 protein language model, identified 653 high-confidence RBP candidates. Foldseek structural homology validation against PDB100 confirmed 13 candidates with a structural match probability of 1.0 to known phage adsorption proteins, spanning four structural archetypes. ESMFold-predicted structures showed strong confidence, with a mean model confidence score of 0.89 and a 90.2% prediction success rate. Exploratory rigid-body docking identified YDV08491.1, an E. coli-targeting candidate, as having the most energetically favourable predicted interaction, with a predicted binding energy of -132.3 kcal/mol against OmpF, supporting experimental prioritisation. These candidates structural diversity and predicted host specificity support their future development as phage-based biosensors and biocontrol tools, and this reproducible, accessible pipeline offers a transferable strategy for prioritising RBP candidates in downstream functional studies, pending experimental validation. Author SummaryFoodborne bacterial infections cause an estimated 600 million illnesses every year worldwide, with Salmonella, Escherichia coli, and Listeria monocytogenes among the most significant contributors. Detecting these pathogens quickly and specifically in food supply chains remains a major challenge for global food safety. Bacteriophages viruses that infect bacteria use surface proteins called receptor-binding proteins (RBPs) to recognize their bacterial hosts with remarkable precision, making RBPs promising building blocks for pathogen-specific biosensors. However, the huge sequence diversity of RBPs across phage genomes has made it difficult to reliably identify good candidates computationally. We built an open-source pipeline that combines deep learning, protein structure prediction, and molecular docking to screen phage genomes for high-confidence RBP candidates. Applied to 247 phage genomes targeting the three pathogens above, our pipeline flagged 653 candidate RBPs, of which 13 showed strong structural similarity to known phage adsorption proteins. One candidate, targeting E. coli, showed a particularly strong predicted binding interaction with a bacterial surface protein, making it a priority target for experimental testing. This pipeline gives researchers a reproducible, publicly available strategy for prioritizing phage proteins toward the development of biosensors and biocontrol tools for food safety.