Distinguishing features of long COVID identified through immune profiling

Post-acute infection syndromes (PAIS) are chronic symptoms that can persist for months or years after acute viral infections. Long COVID, a PAIS following SARS-CoV-2 infection, is characterized by debilitating fatigue, post-exertional malaise, cognitive dysfunction, and autonomic dysfunction. The underlying biological mechanisms are still unclear. While PAIS have been recognized for over a century, their underlying biology remains largely unknown, even for well-studied syndromes like myalgic encephalomyelitis/chronic fatigue syndrome. SARS-CoV-2, the virus causing COVID-19, has resulted in millions of deaths worldwide and while many individuals recover fully, a significant proportion experiences long COVID, significantly impacting their quality of life. This study aims to use comprehensive immune profiling and machine learning to identify the biological signatures of long COVID, advancing our understanding of its pathophysiology and potential biomarker discovery.

The existing literature on long COVID highlights the heterogeneity of symptoms and the lack of a clear understanding of the underlying mechanisms. Several hypotheses have been proposed, including persistent viral antigens, autoimmunity, dysbiosis, reactivation of latent viruses, and chronic inflammation. Studies have reported diverse changes in immune and inflammatory factors in long COVID patients. However, a comprehensive understanding of the immunological features that specifically define long COVID and distinguish it from other post-viral conditions is still lacking. This necessitates a study that employs robust methodologies such as multidimensional immune phenotyping and unbiased machine learning to determine the unique biological signatures characteristic of long COVID.

The Mount Sinai-Yale long COVID (MY-LC) study was a cross-sectional study involving 275 participants: healthcare workers infected with SARS-CoV-2 before vaccination (HCW); healthy, uninfected, vaccinated controls (HC); previously infected, vaccinated controls without persistent symptoms (CC); individuals with persistent symptoms after acute infection (LC); and an external LC group (EXT-LC). Participants in the CC and LC groups had primarily mild acute COVID-19, with samples collected more than a year after acute infection. The HC, CC, and LC groups underwent multidimensional immunophenotyping, including flow cytometry to analyze peripheral blood mononuclear cell (PBMC) populations and assess cytokine production; ELISA to measure SARS-CoV-2 specific antibody responses; and multiplex analysis of circulating hormones and immune mediators. Peptide display and rapid extracellular antigen profiling (REAP) were used to assess antibody reactivity against SARS-CoV-2 and other antigens. A serum epitope repertoire analysis (SERA) and ELISA examined antibody reactivity against herpesviruses. Unbiased machine learning, specifically using logistic regression and LASSO models, integrated the data to identify biomarkers for long COVID. Participants with LC were matched to controls using the Gale-Shapley procedure to account for demographic and clinical variables. Principal component analysis (PCA) and k-nearest neighbor (k-NN) classification were employed for data visualization and classification accuracy assessment.

The study revealed significant differences in circulating immune cell populations between LC and CC groups. Long COVID was associated with higher levels of non-conventional monocytes (CD14lowCD16high), activated B lymphocytes (CD86highHLA-DRhigh), and double-negative B cells (IgD−CD27−CD24−CD38−), while levels of conventional type 1 dendritic (cDC1) cells and central memory CD4+ T cells were lower. Long COVID patients showed increased intracellular IL-2 and IL-4 production in CD4+ T cells and IL-2 and IL-6 production in CD8+ T cells after stimulation. Exaggerated humoral responses against SARS-CoV-2 were observed, particularly against specific spike protein epitopes. Importantly, higher antibody responses to non-SARS-CoV-2 viral pathogens, particularly Epstein-Barr virus (EBV), were detected in long COVID patients. Circulating levels of cortisol were significantly lower in LC individuals compared to controls. Machine learning models demonstrated that serum cortisol, galectin-1 concentration, IgG responses against EBV epitopes, and the number of PD-1+CD4+ T central memory cells were strong predictors of long COVID status. The models achieved high AUC values (around 0.94-0.96), indicating strong discriminatory power.

The findings suggest a complex interplay of factors in long COVID pathophysiology. The persistence of SARS-CoV-2 antigens, reactivation of latent herpesviruses (EBV and VZV), and chronic inflammation appear to contribute to the development of long COVID. The lower cortisol levels observed in long COVID patients may reflect an impaired stress response or a dysregulation of the hypothalamic-pituitary-adrenal (HPA) axis. The elevated antibody responses to EBV suggest that viral reactivation might be a common feature of long COVID. The lack of significant differences in autoantibody levels against the human exoproteome suggests that autoimmunity may not be the dominant driver of long COVID, although the role of autoreactive T cells remains to be investigated. The high predictive power of the machine learning models highlights the potential of using these immunological features as biomarkers for long COVID.

This study identifies distinct immunological features that differentiate individuals with long COVID from controls more than a year after acute infection. The findings support the roles of persistent SARS-CoV-2 antigens, latent herpesvirus reactivation, and chronic inflammation in the pathophysiology of long COVID. Lower cortisol levels and alterations in specific immune cell populations are also significant predictors. The robust predictive models developed offer potential biomarkers for long COVID diagnosis and future research should focus on validating these biomarkers in larger, independent cohorts and investigating the underlying mechanisms of these immunological alterations.

The study's limitations include convenience sampling, potentially leading to biases in participant selection; a relatively small sample size compared to traditional machine learning studies; and focus on peripheral immune factors, neglecting local immune responses within affected organs. The analysis of autoantibodies was limited to the exoproteome, excluding intracellular or non-protein antigens. Future studies are needed to address these limitations and validate the findings in more diverse populations.