A sequence-informed hybrid framework for predicting antibiotic minimum inhibitory concentrations in previously unseen Escherichia coli pathogens
Related Experiment Video
Updated: Nov 5, 2026
Characterization of a Pathogenic Escherichia coli Strain Derived from Oreochromis spp. Farms Using Whole-Genome Sequencing
Published on: December 23, 2022
1Independent Researcher, Cambridge, United Kingdom.
Introduction:
When an Escherichia coli pathogen is newly encountered, antimicrobial susceptibility information may be incomplete or unavailable. We investigated whether minimum inhibitory concentrations (MICs) could be predicted for previously unseen pathogens by jointly using pathogen genomic sequence information, antibiotic characteristics, and observed pathogen-antibiotic MIC patterns.
Hypothesis Or Gap Statement:
The predictive value of integrating pathogen genomic information, antibiotic characteristics, and observed pathogen-antibiotic MIC patterns within a hybrid collaborative-filtering framework for previously unseen E. coli pathogens remains unclear.
Aim:
To develop and evaluate Model C for predicting MIC values in previously unseen, genomically characterised E. coli pathogens.
Methodology:
The study used 9377 E. coli pathogens, 26 antibiotics, and 51,554 exact MIC measurements among 243,802 possible pathogen-antibiotic combinations. Model C combined observed MIC patterns, genomic features related to antimicrobial resistance (AMR), nucleotide-sequence comparisons across 265 AMR-related loci, and antibiotic characteristics. Five-fold pathogen-out validation excluded all MIC measurements from each test pathogen from training and model selection. The common validation cohort contained 9058 pathogens and 50,460 observed MIC values; Models 3B and 3 were evaluated as comparators.
Results:
In pathogen-out validation of the 50,460 observed MIC values, Model C achieved a mean absolute error of 0.5931 log₂ MIC units, root mean squared error of 0.8477, and Pearson correlation coefficient of 0.9055. Model C performed best overall. A final model fitted to all 50,460 observed MICs predicted the remaining 185,048 unmeasured MICs.
Conclusion:
Model C estimated MIC values for previously unseen E. coli pathogens using genomic, antibiotic, and observed interaction information. Independent external validation is required. Predictions may help prioritise confirmatory testing but should not replace antimicrobial susceptibility testing or be used alone to guide treatment.
