MSACL 2024 Abstract

INTRODUCTION:
Bloodstream infections (BSIs) are responsible for over 500,000 hospitalizations and 90,000 deaths each year in North America alone. Our ability to determine the clinical trajectory of BSIs and formulate an effective treatment plan relies on a comprehensive understanding of the molecular characteristics of the infecting pathogen. Currently, our healthcare system is able to systematically track antimicrobial resistance for BSIs, yet, we do not have the molecular level data needed to fully realize all of the virulence factors within microbial populations. This means that the emergence of new, and potentially more virulent, isolates will go largely unrecognized until an evident pattern in the clinical progression of infections emerges. A multi-omics analysis of these microbial pathogens can enable us to determine the risk profiles of microbes and elucidate epidemiological trends. This information can then be used to evaluate Infection Prevention and Control (IPC) policies and implemented into clinical decision-making.

METHODS:
Multi-omics data were collected from 38,000 BSI isolates spanning 12 species. Genomes were collected via Illumina whole genome sequencing, metabolomics via untargeted LC-MS on a Thermo Q Exactive HF, and quantitative proteomics were collected using our automated TMT11-based workflow on a Thermo Orbitrap Fusion Lumos. Each isolate was then matched to electronic medical records from each patient covering over 1,000 factors (e.g. age, sex, comorbidity, antibiotic history, hospital procedure codes, etc.). These data were then used to build supervised machine learning models (XGBoost, deep learning algorithms etc.) to predict 30-day mortality. The predictive influence of each gene, protein, metabolite, and patient factor was then determined using the models to assess the relative contributions of each factor to infection severity.

RESULTS:
Our machine learning models were able to predict 30-day survival with a high degree of accuracy with an area under the receiver operator characteristic curve of over 0.8. A systematic analysis of the contributors to these models showed that patient factors, genomics data, and proteomics data could each be used to build predictors of outcome. Moreover, we found that many of these factors were highly covariant between the models, suggesting that microbial factors and patient factors are not orthogonal. In addition, in assessing the impact of these models we showed that many of the previously reported virulence factors had minimal prognostic value whereas certain factors that have not been previously identified appear to be significant contributors.

DISCUSSION:
We will describe the systematic contributors to the predictive power of our ML approaches and discuss the relative importance of patient versus microbial proteomic features. Overall, our analysis shows that LC-MS proteomics can provide a robust mechanism for classifying microbial species and that our models are able to accurately detect the expression of virulence factors and resistance traits. Beyond this, we show that patient factors have a significant influence on the 30-day mortality and that microbially-linked traits—though important for guiding clinical therapy—are ultimately less important than patient factors when determining clinical outcome.