118. Machine Learning Approaches to Predicting Treatment Outcomes for Carbapenem-Resistant Enterobacterales in a Region with High Prevalence of Non-Carbapenemase Producers

Abstract Background Carbapenem-resistant Enterobacterales are a growing threat globally. Early detection of CRE is necessary for appropriate treatment and infection control measures. Many hospital labs can test for carbapenemase production; however, in some regions, including South Texas, the majority of CRE are non-carbapenemase producing (NCPE). This study had two interrelated aims to develop decision rules tailored to a region with high prevalence of NCPE to predict 1) antimicrobial resistance (AMR) from whole genome sequencing (WGS) data and 2) CRE treatment outcomes. Methods To better understand links between resistome, phenotypic AMR, and prediction of outcomes for CRE, we developed decision rules to build machine learning prediction models. We conducted WGS and antibiotic susceptibility testing (21 antibiotics) on CRE isolates from unique patients across 5 hospitals in the South Texas region between 2013 and 2020. Day 30 outcomes were based on desirability of outcome ranking (DOOR). The overall classification accuracies of the models are reported. Results Overall 146 CRE isolates were included, 97 were used to train each model, and 49 were used for validation. Among the K. pneumoniae and E. coli CRE isolates that were available with susceptibility data, the majority (62%) were NCPE. For the clinical recovery model (DOOR), the combination of admission ICU status, piperacillin-tazobactam (PT) MIC > 16, presence of sul gene, and polymyxin-sparring regimens associated with an overall accuracy of 95% for having a worse DOOR. Majority (60%) of patients were empirically treated with piperacillin-tazobactam; notably, less than 33% isolates had PT MIC ≤ 16. Interestingly, combined effects of isolates that did not harbor carbapenemases, blaOXA-1, blaCTX-M-15, blaCMY, or aac(6’)ib-cr genes resulted in a decision rule with a 95.7% overall accuracy for susceptibility to PT (MIC < 16 ug/mL). Conclusion Herein, we used machine learning approaches to predict AMR and treatment-based outcomes linked with WGS data in a region with predominantly NCPE infections. Machine learning can obtain a model that can make repeatable predictions, further data can improve the accuracy and guide tailored clinical decision-making. Disclosures Grace Lee, PharmD, PhD, BCPS, Merck Co. (Grant/Research Support)NIA/NIH (Research Grant or Support)


Session: O-24. New Developments in Infectious Diseases Diagnostics
Background. Carbapenem-resistant Enterobacterales are a growing threat globally. Early detection of CRE is necessary for appropriate treatment and infection control measures. Many hospital labs can test for carbapenemase production; however, in some regions, including South Texas, the majority of CRE are non-carbapenemase producing (NCPE). This study had two interrelated aims to develop decision rules tailored to a region with high prevalence of NCPE to predict 1) antimicrobial resistance (AMR) from whole genome sequencing (WGS) data and 2) CRE treatment outcomes.
Methods. To better understand links between resistome, phenotypic AMR, and prediction of outcomes for CRE, we developed decision rules to build machine learning prediction models. We conducted WGS and antibiotic susceptibility testing (21 antibiotics) on CRE isolates from unique patients across 5 hospitals in the South Texas region between 2013 and 2020. Day 30 outcomes were based on desirability of outcome ranking (DOOR). The overall classification accuracies of the models are reported.
Results. Overall 146 CRE isolates were included, 97 were used to train each model, and 49 were used for validation. Among the K. pneumoniae and E. coli CRE isolates that were available with susceptibility data, the majority (62%) were NCPE. For the clinical recovery model (DOOR), the combination of admission ICU status, piperacillin-tazobactam (PT) MIC > 16, presence of sul gene, and polymyxin-sparring regimens associated with an overall accuracy of 95% for having a worse DOOR. Majority (60%) of patients were empirically treated with piperacillin-tazobactam; notably, less than 33% isolates had PT MIC ≤ 16. Interestingly, combined effects of isolates that did not harbor carbapenemases, blaOXA-1, blaCTX-M-15, blaCMY, or aac(6')ib-cr genes resulted in a decision rule with a 95.7% overall accuracy for susceptibility to PT (MIC < 16 ug/mL).
Conclusion. Herein, we used machine learning approaches to predict AMR and treatment-based outcomes linked with WGS data in a region with predominantly NCPE infections. Machine learning can obtain a model that can make repeatable predictions, further data can improve the accuracy and guide tailored clinical decision-making.
Disclosures. Grace Lee, PharmD, PhD, BCPS, Merck Co. (Grant/Research Support)NIA/NIH (Research Grant or Support) Background. Despite antifungal therapy and surgical debridement, overall mortality of invasive mucormycosis is >40%. Currently the world is witnessing an explosion in mucormycosis in India among COVID-19 patients with an official count of 28,252 cases as of 06/07/2021. Thus, novel therapeutic modalities are needed. We previously reported on a mouse monoclonal antibody (C2) targeting CotH invasins being protective against mucormycosis. Here, we humanized C2 MAb and assessed its efficacy in vitro and in vivo.

A Humanized Antibody Targeting the CotH Invasins is Protective Against Murine Mucormycosis
Methods. The C2 (IgG1) paratopes of the heavy chain and light chain were grafted on the most suitable human IgG1 with back mutations in the paratopes needed to restore binding of humanized clones to CotH3 (by biolayer interferometry using Gator). Clones were compared to C2 in their ability to prevent Rhizopus delemar-induced injury to A549 alveolar epithelial and primary human endothelial cells and for enhancing human neutrophil killing of the fungus in vitro. C2 and the humanized clones were also compared for their ability to protect neutropenic mice from mucormycosis induced by R. delemar or Mucor cicrinelloides with and without antifungal therapy.
Results. Three humanized clones showed 10-fold enhanced binding affinity to CotH3 protein (~5 nM for humanized vs. ~50 nM for C2). One humanized clone (VX01) doubled the ability of neutrophils to kill R. delemar and resulted in ~50% reduction in host cell damage. A single low dose of VX01 (30 µg) given 24 h post infection resulted in comparable survival of 60-70% in mice infected intratracheally