#3120 PREDICTION OF COVID-19 SEVERITY IN HOSPITALIZED PATIENTS USING URINARY PROTEOMICS AND MACHINE LEARNING

Abstract

Abstract Background and Aims COVID-19 has been a significant public health concern for the last three years; however, not much is known about the mechanisms that lead to severe kidney outcomes in patients hospitalized with COVID-19. In this multicenter study, we combine isobaric TMT-tagged urinary proteomics and machine learning to predict severe kidney outcomes in hospitalized COVID-19 patients. Method Urine samples from hospitalized COVID-19 patients in two medical centers (Mount Sinai Hospital and University of Michigan) were used in this study in adherence with proper consenting protocols. Urine samples were prepared for LC-MS/MS analysis as previously reported [1]. The obtained spectra were analyzed using Proteome discoverer software and matched against Uniprot human database. For constructing the ML algorithm, the samples were randomly divided into discovery and validation set at a 2:1 ratio. Severe outcomes were defined as ICU admission, mechanical ventilation, acute kidney injury (AKI), death, or length of stay more than 21 days. Limma test was used on the discovery set to identify differentially expressed proteins and then features were selected using Boruta feature selection method. 10-fold cross validation on a random forest model was then applied to obtain receiver operating characteristic (ROC) curves. Results Urine samples from 120 PCR-positive COVID-19 patients from two different medical centers were collected within one week of hospitalization. More than 3,000 unique urinary proteins were identified using TMT-tagged mass spectrometry. For constructing a predictive algorithm, patients were stratified into severe and mild outcomes. Using Limma test on the discovery set, we identified differentially expressed proteins (DEPs) in severe outcome cohort vs the mild outcome cohort (Figure 1A). A set of 12 top features were identified using Boruta feature selection method and used for random forest model construction within the discovery set with 10-fold cross validation (Figure 1B). The generated ROC curves show that the algorithm demonstrated good predictive power for both discovery and validation set with 87% and 79% accuracy, respectively and close to 90% specificity (Figure 1C, D). On average, major adverse kidney events were observed in patients within 5-13 days after hospitalization. Enrichment analysis of DEP in COVID-19 patients compared to healthy patients showed significant upregulation of immune related processes and downregulation of proteolytic and metabolic processes. Enrichment analysis of DEPs in severe COVID-19 patients compared to mild COVID-19 patients showed significant upregulation of exocytosis and some immune related processes and downregulation of cell adhesion and extracellular matrix organization related processes (Figure 2A, B). Upregulated proteins were associated with kidney proximal tubular cells in addition to pulmonary alveolar cells (Figure 2C). Downregulated proteins were associated strongly with kidney cells such as podocytes and mesangial cells in addition to endothelial cells (Figure 2D). Conclusion Here, we developed an algorithm for prediction of severity in COVID-19 patients within 5-13 days after hospitalization. We further delineate potential mechanisms that drive severe outcomes in COVID-19 patients. Learnings from this study can be used for developing therapeutic options for long COVID, in addition to better preparedness in the event of other respiratory illnesses in the future.