Hi — I ran a SHAP feature attribution analysis on an E. coli MIC prediction model (XGBoost, 6,082 NARMS Sensititre isolates, grouped CV) and compared the top predictive genes against AMRrules v1.0.0.
53 gene-drug associations with strong empirical support (SHAP >= 0.1, n >= 50) aren't in the current E. coli ruleset. Most are well-known mechanisms — sharing in case the quantitative evidence is useful for your curation process.
CSV with SHAP values, prevalence, and sample sizes:
https://github.com/kglazier/amrcast/blob/master/experiments/amrrules_candidate_genes.csv
Happy to reformat if useful.
Project: https://github.com/kglazier/amrcast
Hi — I ran a SHAP feature attribution analysis on an E. coli MIC prediction model (XGBoost, 6,082 NARMS Sensititre isolates, grouped CV) and compared the top predictive genes against AMRrules v1.0.0.
53 gene-drug associations with strong empirical support (SHAP >= 0.1, n >= 50) aren't in the current E. coli ruleset. Most are well-known mechanisms — sharing in case the quantitative evidence is useful for your curation process.
CSV with SHAP values, prevalence, and sample sizes:
https://github.com/kglazier/amrcast/blob/master/experiments/amrrules_candidate_genes.csv
Happy to reformat if useful.
Project: https://github.com/kglazier/amrcast