Assessment for antibiotic resistance in Helicobacter pylori: A practical and interpretable machine learning model based on genome-wide genetic variation.

Helicobacter pylori (H. pylori) antibiotic resistance poses a global health threat. Accurate identification of antibiotic resistant strains is essential for the control of infection. In the present study, our goal is to leverage the whole-genome data of H. pylori to develop practical and interpretable machine learning (ML) models for comprehensive antibiotic resistance assessment. A total of 296 H. pylori isolates with genome-wide data were downloaded from the Bacterial and Viral Bioinformatics Resource Center (BV-BRC) and the National Center for Biotechnology Information (NCBI) databases. By training ML models on feature sets of single nucleotide polymorphisms from SNP calling (SNPs-1), antibiotic-resistance SNP annotated by the Comprehensive Antibiotic Resistance Database (SNPs-2), gene presence or absence (GPA), we generated predictive models for four antibiotics and multidrug-resistance (MDR). Among them, the models that combined SNPs-1, SNPs-2, and GPA data demonstrated the best performance, with the eXtreme Gradient Boosting (XGBoost) consistently outperforming others. And then we utilized the SHapley Additive exPlanations (SHAP) method to interpret the ML models. Furthermore, a free web application for the MDR model was deployed to the GitHub repository (https://H.pylori/MDR/App/). Our study demonstrated the promise of employing whole-genome data in conjunction with ML algorithms to forecast H. pylori antibiotic resistance. In the future, the application of this approach for predicting H. pylori antibiotic resistance would hold the potential to mitigate the empiric administration.