Evaluation of Machine Learning and Rules-Based Approaches for Predicting Antimicrobial Resistance Profiles in Gram-negative Bacilli from Whole Genome Sequence Data

The time-to-result for culture-based microorganism recovery and phenotypic antimicrobial susceptibility testing necessitates initial use of empiric (frequently broad-spectrum) antimicrobial therapy. If the empiric therapy is not optimal, this can lead to adverse patient outcomes and contribute to increasing antibiotic resistance in pathogens. New, more rapid technologies are emerging to meet this need. Many of these are based on identifying resistance genes, rather than directly assaying resistance phenotypes, and thus require interpretation to translate the genotype into treatment recommendations. These interpretations, like other parts of clinical diagnostic workflows, are likely to be increasingly automated in the future. We set out to evaluate the two major approaches that could be amenable to automation pipelines: rules-based methods and machine learning methods. The rules-based algorithm makes predictions based upon current, curated knowledge of Enterobacteriaceae resistance genes. The machine-learning algorithm predicts resistance and susceptibility based on a model built from a training set of variably resistant isolates. As our test set, we used whole genome sequence data from 78 clinical Enterobacteriaceae isolates, previously identified to represent a variety of phenotypes, from fully-susceptible to pan-resistant strains for the antibiotics tested. We tested three antibiotic resistance determinant databases for their utility in identifying the complete resistome for each isolate. The predictions of the rules-based and machine learning algorithms for these isolates were compared to results of phenotype-based diagnostics. The rules based and machine-learning predictions achieved agreement with standard-of-care phenotypic diagnostics of 89.0 and 90.3%, respectively, across twelve antibiotic agents from six major antibiotic classes. Several sources of disagreement between the algorithms were identified. Novel variants of known resistance factors and incomplete genome assembly confounded the rules-based algorithm, resulting in predictions based on gene family, rather than on knowledge of the specific variant found. Low-frequency resistance caused errors in the machine-learning algorithm because those genes were not seen or seen infrequently in the test set. We also identified an example of variability in the phenotype-based results that led to disagreement with both genotype-based methods. Genotype-based antimicrobial susceptibility testing shows great promise as a diagnostic tool, and we outline specific research goals to further refine this methodology.