Considerations for the Development of Reliable and Sustainable Genotype-Based Diagnostics for Antimicrobial Resistance

Abstract

Rapid diagnostics for antimicrobial resistance (AMR) offer one of the most promising methods for reducing inappropriate and ineffective antibiotic use, quelling the selective pressures for resistance, and ultimately reducing the public health burden imposed by AMR. With the increasing speed and decreasing cost of sequencing, genotype-based assays for AMR prediction are more practical in the diagnostic setting than the relatively slow and sometimes labor intensive phenotypic AMR tests. While there has been considerable emphasis on the development on machine learning (ML)-based AMR diagnostics, there has previously been little focus on biological and technical factors that might influence the performance and, perhaps more importantly, the reliability of these predictive models. Further, even for single-locus, highly accurate genetic markers of resistance, it can be expected that sensitivity of genotype-based diagnostics will wane with the emergence of novel resistance mechanisms, but there has previously been no quantification of the sampling rates necessary to ensure timely detection of such novel variants. It has also not been assessed how such surveillance strategies may be optimized in order to increase detection efficiency of novel variants over random sampling, minimizing the number of treatment failures that may be attributed to them, as well as the cost associated with surveillance. Here, we present an evaluation of factors that may influence the performance and reliability of ML-based AMR testing from whole genome sequencing data. We further present a model quantifying the surveillance required to maintain the sensitivity of genotype-based AMR diagnostics, along with evaluations of targeted surveillance strategies that may increase the detection efficiency of novel AMR variants.