Antimicrobial resistance (AMR) is a growing burden in US hospital systems costing an estimated $2 billion annually.  Knowing which antibiotics will be effective reduces morbidity and mortality and improves overall antibiotic stewardship. The current gold standard diagnostic for AMR is culturing in a clinical pathology laboratory, which can take days.

As costs have gone down and technology has improved, genome sequencing has been touted for its potential as a diagnostic.  Previous studies have demonstrated that AMR phenotypes can be predicted by training machine learning models using genome sequence data as input. These studies have typically used input features that are based on whole genomes or curated sets of AMR genes to predict susceptible, intermediate, and resistant phenotypes or minimum inhibitory concentrations. Since whole genome sequencing requires culturing, it offers little improvement in turnaround times. A more appropriate strategy might be to perform metagenomic sequencing directly from the infection source. However, the individual genomes in metagenomic sequences are often incomplete, and it can be difficult to attribute mobile genetic elements carrying AMR genes with their appropriate host genomes.

In this talk, I will discuss the progress that we have made predicting AMR phenotypes from partial genome sequence data.