1352. A Precision Medicine Approach to Predicting “Long COVID” through Machine Learning Analysis of Whole Blood Transcriptome Data

Abstract

Background. Predicting, preventing, and treating post-COVID condition (PCC; 'Long COVID) is challenging due to a limited understanding of PCC mechanisms. To address this, we analyzed whole blood transcriptomic data using machine learning to identify candidate gene and gene expression pathways associated with the development of PCC. Methods. The Epidemiology, Immunology, and Clinical Characteristics of Emerging Infectious Diseases with Pandemic Potential (EPICC) study is a longitudinal cohort study exploring the impact of SARS-CoV-2 infection inmilitary health system beneficiaries. We collected demographic and clinical data through surveys, interviews and medical record reviews. We conducted transcriptome profiling (bulk RNA-seq) on early post-infection whole blood samples fromunvaccinated individuals with and without PCC at 6-months, as well as uninfected, unvaccinated controls. These three groups were matched on sex, age, and comorbidities. We identified transcriptomic signatures associated with group classes using t-tests with correction for multiple comparisons. Molecular signatures and candidate markers for PCC were developed using gene set enrichment analysis (GSEA) and machine learning, including random forest (RF) and support vector machine (SVM) methods. Results. Out of the 5289 SARS-CoV-2 positive Military Health System beneficiaries enrolled in EPICC, 814 participants had at least one sample with blood transcriptomic profiling data. Among these, 171 unvaccinated, SARS-CoV-2 positive participants completed a 6-month follow up survey and were included in the analysis. Ninety-one percent did not report any chronic symptoms, while 8% reported moderate to severe symptoms at six months and were classified as PCC group. A control group of 51 participants was also identified. Machine learning approaches (RF: AUROC = 0.90, CI = 0.79 to 1; and SVM: AUROC = 0.89, CI = 0.76 to 1) identified three candidate biomarkers (Figure 1), including TSSK4, TUFT1, and XBP1 genes, which could be a predictive or mechanistic marker for PCC with further validation. Conclusion. Our study demonstrates the potential of using machine learning to translate transcriptomic data into precision medicine applications, specifically for predicting the development of PCC after COVID-19. (Table Presented).