AMRpredictor: Visualizing Machine Learning-Based AMR Predictions

Overview

AMRpredictor is an interactive web application designed to visualize the performance and interpretability of machine learning (ML) models developed for predicting antimicrobial resistance (AMR) phenotypes in ESKAPEE pathogens. The app allows users to explore per-antibiotic model metrics and inspect the genomic features most influential for predictions, as determined by SHAP (SHapley Additive Explanations) values.

What You Can Do Here

• Predict AMR phenotypes: Upload your bacterial genome (FASTA) and get real-time resistance predictions for multiple antibiotics across 6 ESKAPEE pathogens.
• Understand predictions with SHAP: Visualize which genomic features (k-mers, genes, promoters) contributed most to each prediction using SHAP explanations.
• Explore model performance: View accuracy, precision, recall, F1 scores, and confusion matrices for each antibiotic-genus combination.
• Analyze feature importance: Browse and search the top genomic features (AMR genes, promoters, rRNA) driving predictions across antibiotics.
• Download datasets and models: Access curated genome assemblies, antibiograms, trained ML models, and SHAP outputs for your own research.
• View dataset statistics: Explore the distribution of strains across genera and data sources with interactive charts.

Behind the Models

The underlying models were trained on a curated dataset of 18,916 genome assemblies paired with manually checked antibiograms, covering 40 antibiotics. K-mer frequency encoding (sizes 3, 4, 5) was applied separately to AMR and core protein sequences, promoter DNA, and rRNA regions. Random Forest and Extreme Gradient Boosting (XGBoost) classifiers were trained for each antibiotic and evaluated using cross-validation.

Explore More

• GitHub repository (scripts, pipelines, documentation): https://github.com/dianalabgr/amrprediction
• Zenodo data repository (models, SHAP outputs, datasets): https://zenodo.org/records/16213507
• NCBI SRA BioProject (metagenomic data): PRJNA1290626

Disclaimer

This web application is intended for visualization and educational purposes only. It does not provide clinical AMR predictions for diagnostic or therapeutic use.