Data-driven classification of Escherichia coli using protein language model ascertains O-serotype determining genes

Serotyping of Escherichia coli is a critical tool for clinical diagnostics, epidemiology, and vaccine development. Traditional serotyping methods rely on antigen–antibody interactions or reference-based genomic sequence similarity tools that depend on direct matches to predefined serotype gene databases, making their performance highly sensitive to database completeness and limiting their ability to detect or generalize to novel, divergent, or recombinant variants. In this study, we introduce a data-driven approach using a Protein Language Model (PLM) to classify E. coli O serotypes. We applied the ESM-2 model to encode protein sequences into vector representations and trained a machine learning classifier to predict serotypes from genomic data. A dataset of 11,272 E. coli genomes was analyzed, revealing nine ML-prioritized markers—wcaM, wcaL, wcaK, wzzE, wzxC, wecC, glmM, garR, and hisD—that significantly contribute to O serotype classification. The classification model based on these genes outperformed traditional bioinformatics tools, achieving an accuracy of 93% using a Random Forest classifier. Notably, the proposed model exhibits high recall for low-frequency and underrepresented serotypes, leading to improved balanced performance across classes and contributing to the overall increase in classification accuracy. Our findings suggest that PLM-based analysis enhances serotype prediction, providing a scalable and efficient framework for high-throughput epidemiological surveillance and exploratory bacterial classification, complementing current diagnostic methods.