Background
Long COVID (LC) is a newly defined syndrome linking COVID infection with an umbrella of long-lasting symptoms. Due to this variety of presentation and lack of understanding of the mechanism involved, no reliable biomarkers have been found so far, and diagnostic uncertainty remains high. In the prospect of personalized medicine, there is a dire need for diagnostic and prognostic markers. To address this issue, we designed a multi-approach exploratory study on a well characterized cohort of LC patients using Resolved COVID infections (RC) and Uninfected Controls (UC).
Methods
Plasma samples from 100 LC, 60 RC and 60 UC were analyzed by using state-of-the-art IVDr NMR allowing to harvest detailed lipidomic and metabolomic data and immune and mitochondrial stress markers explored by several multiplex ELISAs and electrochemiluminescent assays. On these data, we performed multivariate dimensionality reduction models (PCA, OPLS-DA) to visualize data distribution and discriminant variables between groups. Candidate biomarkers were validated with univariate analysis and ROC curve analysis, and then used to construct a machine deep learning model able to sort patients into diagnostic groups. Finally, LC symptoms were correlated with the most significant variables.
Results
We identified several potential biomarkers related to platelet and neutrophil activation, inflammation, TCA cycle, vascularization, lipoproteins. Models fed by those biomarkers alone were sufficient to cluster the LC group from the two other groups. We also found that both previously infected groups (RC and LC) had a specific signature distinguishing them from UC. SARS-CoV-2 infection induced host immuno-metabolic signatures that persisted after full recovery (RC group). However, this signature was different in the LC group and was associated with LC pathophysiology. We created several machine learning models to test whether these biomarkers along with the other variables are enough to predict LC patients. Our model was able to correctly predict 99% of LC patients and had about 70% accuracy in distinguishing RC from UC patients.
Conclusions
In this study, we identified potential biomarkers sufficient to diagnose LC with 99% accuracy. These biomarkers were associated with immune and metabolic dysregulation, as well as content symptoms, providing clues to future therapeutic targets. We are now testing these markers in a prospective setting to validate their use in clinical practice.