Abstract 1231 Using Machine Learning to Predict Antimicrobial Resistance in Water-Derived E. coli

Antimicrobial resistance (AMR) is a serious public health problem worldwide leading to an estimated 1.27 million deaths in 2019 and is projected to result in one trillion dollars of additional medical costs through 2050. Whole genome sequencing is becoming more common to characterize bacterial pathogens in clinical settings and advances in machine learning are being applied to these data to predict antibiotic resistance profiles of bacterial pathogens. As part of longitudinal monitoring of recreational waters in the Macatawa watershed (Holland, MI), we have isolated over 10,000 strains of from lake and riverine sites using EPA Method 1603 and have sequenced (Illumina MiSeq) over 500 strains, allowing characterization of the genetic diversity of freshwater Escherichia populations. Diverse data sets such as this have the potential to improve informatics approaches applied in clinical settings. The objectives of this study were to 1) experimentally determine antibiotic resistance profiles of freshwater strains, 2) evaluate the accuracy of machine learning AMR prediction algorithms on these strains, 3) assess the impact of including potentially diverse genetic information into existing algorithm training sets that primarily comprise clinically isolated strain information, and 4) link short genomic sequences (k-mers) used by the algorithms to make AMR predictions to known and potentially novel genetic loci in these genomes. We have experimentally determined antibiotic resistance profiles for 158 of the sequenced watershed strains using a MicroScan WalkAway 96 Plus system that screens for 20 clinically relevant antibiotics. In collaboration with Argonne National Laboratory, we have used a machine learning approach (Nguyen et al., 2019; https://doi.org/10.1128/JCM.01260-18) to classify 517 sequenced watershed strains into resistance categories (susceptible, intermediate, resistant) for 19 antibiotics. Results indicate that the freshwater E. coli strains have lower numbers of resistances per strain than clinical strains, with resistance to tetracycline being the most common. Machine learning predictions for 19 antibiotic models have accuracies ranging from 87.6% to 88.2% when compared to the experimentally determined AMR profiles for watershed strains; model accuracies were improved with the addition of freshwater strain genomic and AMR profile information into model training sets. In order to further understand the modeling predictions, we have begun to interrogate which genetic sequences the algorithms are using to classify a strain as resistant or susceptible to an antibiotic using comparative genomics approaches. Initial results indicate that genes associated with known antibiotic resistance mechanisms are identified (e.g., TetR, TetA, TetB), increasing confidence in model predictions and potentially allowing the discovery of novel resistance mechanisms in this population of Escherichia. This work has produced a data set of linked freshwater Escherichia genomes and experimentally determined antibiotic resistance profiles that can be used to improve machine learning approaches to predicting resistance profiles and provide insights into the genomic context of resistances. Future applications could include accurate identification of antimicrobial resistance in clinical settings, potentially reducing overuse and misuse of antibacterial drugs and refining patient care decisions. Funding support is gratefully acknowledged from the National Science Foundation (MRI and MCB), Project Clarity/Outdoor Discovery Center, Herbert H. and Grace A. Dow Foundation, and the A. Paul Schaap endowed research fund.