Developing an in silico minimum inhibitory concentration panel test for Klebsiella pneumoniae

Abstract

Antimicrobial resistant infections are a serious public health threat worldwide. Whole genome sequencing approaches to rapidly identify pathogens and predict antibiotic resistance phenotypes are becoming more feasible and may offer a way to reduce clinical test turnaround times compared to conventional culture-based methods, and in turn, improve patient outcomes. In this study, we use whole genome sequence data from 1668 clinical isolates of Klebsiella pneumoniae to develop a XGBoost-based machine learning model that accurately predicts minimum inhibitory concentrations (MICs) for 20 antibiotics. The overall accuracy of the model, within ±1 two-fold dilution factor, is 92%. Individual accuracies are ≥90% for 15/20 antibiotics. We show that the MICs predicted by the model correlate with known antimicrobial resistance genes. Importantly, the genome-wide approach described in this study offers a way to predict MICs for isolates without knowledge of the underlying gene content. This study shows that machine learning can be used to build a complete in silico MIC prediction panel for K. pneumoniae and provides a framework for building MIC prediction models for other pathogenic bacteria.

Precise prediction of antibiotic resistance in Escherichia coli from full genome sequences

“…used random forests to predict invasiveness of Salmonella enterica lineages [14]. In another study, a tree ensemble was trained with boosting to predict the minimum inhibitory concentration from DNA k-mers for a large-scale Klebsiella pneumoniae panel [10], but the value of using core genome compared to accessory genes was not investigated. In general, including variant data or k-mers in the model greatly increases the number of features.…”

“…As general-purpose methods, they are agnostic to the causal mechanisms, and learn useful features directly from data [7-9]. Already, decision tree based models have proven valuable for predicting resistance and pathogen invasiveness from genomic sequences [10-14]. However, these studies were limited in both the genetic features used and the methods applied.…”

Current Affairs of Microbial Genome-Wide Association Studies: Approaches, Bottlenecks and Analytical Pitfalls

“…Furthermore, we encourage open data sharing to improve the quality of testing and thus solve the problem of overfitting of tools to specific datasets or organisms. For machine learning, data sharing enables training of models that predict anti-microbial resistance (AMR) phenotypes without relying on a database of preexisting AMR genes or mutations (Nguyen et al, 2018). We recommend depositing of raw sequences in the sequence read archive (SRA), an international public archival resource for next generation sequencing data (Leinonen et al, 2011) and publishing the accession numbers.…”

Machine learning with random subspace ensembles identifies antimicrobial resistance determinants from pan-genomes of three pathogens

“…AMR prediction models have already been developed from genome sequence collections of many pathogens, such as Staphylococcus aureus [3,4,5], Mycobacterium tuberculosis [4,6,7], Salmonella [8,9], Klebsiella pneumoniae [10,11], and Neisseria gonorrhoeae [12,13]. However, these approaches are often designed to maximize accuracy in predicting AMR phenotypes, emphasizing their diagnostic capabilities over their capacity to uncover genetic mechanisms for resistance.…”