Chronic inflammation, neutrophil activity, and autoreactivity splits long COVID

The study addresses why some individuals experience persistent symptoms after acute SARS-CoV-2 infection—post-acute sequelae of COVID-19 (PASC, or long COVID)—and seeks immunologic correlates that distinguish PASC from uncomplicated recovery. COVID-19 exhibits heterogeneous clinical manifestations and outcomes, and while many immunologic features of acute disease have been described (e.g., myeloid and neutrophil activation, cytotoxic T-cell responses, collapse of germinal centers with extrafollicular B-cell responses and emergent autoreactivity), their resolution and contribution to PASC are unclear. PASC spans diverse symptoms (e.g., dyspnea, fatigue, brain fog) persisting months to years, with diagnostic thresholds variably defined by CDC and WHO. Prior studies show mixed results identifying stable biological distinctions for PASC across time. The purpose of this study is to integrate broad serological, proteomic, and cellular immune profiling with machine learning to define immunologic signatures of PASC, identify clinically meaningful subtypes, and inform diagnostic and therapeutic strategies.

Existing literature identifies immunotypes associated with acute COVID-19 severity, including prominent myeloid activation, neutrophil activity, and cytotoxic T-cell responses. Severe disease features include collapse of germinal centers and emergence of extrafollicular B-cell responses, linked to de novo autoreactivity seen in autoimmune diseases like lupus and persisting after severe COVID-19. Early recovery studies suggested inflammation-associated PASC immunotypes but noted waning associations over time. Other correlates (e.g., cortisol) have been reported, though their pathophysiologic significance remains uncertain. There is growing evidence implicating dysregulated innate and adaptive immunity, coagulation, antigen processing/presentation pathways, and autoantibodies in PASC; however, robust, generalizable classifiers and clinically actionable subtypes have been lacking.

Design and cohorts: Cross-sectional analysis of 97 PASC patients and 26 donors with uncomplicated COVID-19 recovery (CR) at similar days post symptom onset (mean ~140 days for PASC; CR ~110 days). Demographics and symptom profiles were collected; common PASC symptoms included dyspnea (≈69%), fatigue (64%), and brain fog (47%). Patients with suspected pre-COVID rheumatic disease were excluded. Proteomics: Plasma proteomics using Olink Explore 3072 measured ~2925–3072 protein features. Quality control used Olink tools and custom R pipelines. Differential abundance and pathway enrichment (Reactome/KEGG) were performed; PCA and clustering analyses were used to explore heterogeneity. Machine learning: Supervised Random Forest (RF) models were trained to classify PASC vs CR and to classify inflammatory PASC (infPASC) vs others. A consensus approach trained up to 10,000 models with randomized 80/20 train/test splits; 1000 trees per model with tuned hyperparameters. Feature potency scoring incorporated frequency of selection, feature importance (Gini), and model AUC performance. A reduced 12-marker panel derived from top discriminators was also evaluated. Clinical labs and biomarkers: Routine clinical testing assessed inflammatory and coagulation markers (e.g., CRP, fibrinogen), complete blood counts (neutrophils), and neutrophil extracellular trap (NET) activity markers (e.g., citrullinated histone H3). Associations between proteomics and clinical measurements were examined. B-cell phenotyping: Antigen-specific flow cytometry on PBMCs from PASC subtypes (n≈14 infPASC; n≈24 niPASC) quantified B-cell subsets including extrafollicular lineage (DN2), plasmablast/ASC, naive, and memory compartments; antigen-specific frequencies to SARS-CoV-2 antigens were assessed and clustered. Serology: Luminex-based multiplex assays quantified IgM/IgA/IgG against SARS-CoV-2 spike (RBD) and non-spike (e.g., nucleocapsid) antigens. Analyses considered time since diagnosis (>120 days) to address waning. Autoreactivity: Clinical autoantibody panels (Exagen, CLIA-certified) screened for ANA (ELISA/IF), anti-dsDNA, extractable nuclear antigens, RF, anti-cardiolipin, anti-β2-glycoprotein 1, ANCA (PR3/MPO), anti-GBM, anti-CCP, and others across IgG/IgM/IgA isotypes. Follow-up ANA testing ~1 year later was performed in subsets. Broad antiviral serology: VirScan phage immunoprecipitation sequencing profiled antibodies against >14,000 peptides from >450 human pathogens to assess non–SARS-CoV-2 viral reactivity trends. Statistics and software: Analyses conducted in R (v3.6.2), GraphPad Prism; heatmaps via pheatmap, clustering by Ward’s method; UMAP; GSEA/CSEA for pathway analyses. RF modeling via MLJAR/DecisionTrees; model AUCs and feature importance calculated; feature potency Score(f,M)=Gini(f)*AUC(M).

- Differential proteomics: >700 proteins had significantly increased abundance in PASC vs CR; ~20 decreased. Elevated proteins were largely inflammatory, including IL-6 and NF-κB pathway components; neutrophil-related signatures were prominent in differential expression. - ML discriminators: While classical proinflammatory cytokines (IL-6, IL-8) were not the strongest RF discriminators of PASC vs CR, top features emphasized pathways related to coagulation cascades, EGFR/EGF ligand signaling (e.g., EREG), antiviral sensing/antigen processing (e.g., IFI30), and antigen presentation. - PASC heterogeneity and subtypes: Unsupervised clustering of recovery cohorts revealed two PASC subgroups: an inflammatory PASC (infPASC) and a non-inflammatory PASC (niPASC). infPASC showed markedly elevated IL-6, IL-8, and IL-1β; in infPASC these cytokines exceeded levels observed in severe/critical acute COVID-19. - Clinical correlates of infPASC: infPASC had higher CRP, fibrinogen, absolute neutrophil counts, and NET markers (e.g., citrullinated histone H3). Neutrophil expansion and inflammatory/coagulation markers were correlated (e.g., reported R²≈0.44 for specific associations). - B-cell alterations: PASC patients exhibited signs of ongoing extrafollicular activity, including elevated DN2 B cells, enriched in infPASC. Antigen-specific B cells were increased across PASC subtypes, with infPASC enriched for antigen-specific naive and class-switched memory populations, indicating ongoing B-cell activation/memory remodeling. - Humoral targeting: Spike RBD-binding titers did not clearly differ between subtypes, with slightly higher IgM/IgA in infPASC. Anti-nucleocapsid antibodies were enriched across isotypes in the infPASC cohort and remained elevated beyond 120 days, indicating stable differences in targeting over time windows assessed. - Autoreactivity: >75% of PASC patients had reactivity to at least one autoantigen; >33% had two or more positive autoantibodies. ANAs were more frequent and at higher titers in niPASC; ANCA positivity occurred only in niPASC (4/44). Of six anti-β2-glycoprotein 1–positive patients, five were niPASC. Longitudinally (~1 year), ANA titers tended to decline in niPASC (5/7) but increased in infPASC (6/8), suggesting building autoreactivity in infPASC. - ML classification of infPASC: RF models distinguishing infPASC from all other recoveries achieved mean ROC AUC ≈0.95 (SD ±0.04) using the full proteome. A curated 12-marker panel maintained high performance (mean ROC AUC ≈0.94, SD ±0.05), indicating feasible identification without full proteomics. - Overall, findings support a subdivision of PASC into inflammatory and non-inflammatory types, each with distinct innate and adaptive immune features and autoreactive profiles, with implications for diagnostics and treatment.

The data support that PASC encompasses at least two biologically distinct subtypes. An inflammatory subtype (infPASC) shows broad inflammatory signaling, pronounced neutrophil activation and NET formation, and qualitative alterations in B-cell memory and antigen-specific responses. A non-inflammatory subtype (niPASC) exhibits weaker systemic inflammation but greater ANA incidence and diverse autoreactivity, including ANCA in a subset. These distinctions help reconcile inconsistent findings across prior PASC studies by accounting for immunologic heterogeneity. Machine-learning analyses reveal that global proteomic patterns, rather than single canonical cytokines, best discriminate PASC, highlighting pathways in coagulation, EGFR signaling, and antigen processing/presentation. For infPASC, strong models using either full proteomics or a 12-marker subset enable practical identification strategies potentially implementable with standard clinical assays when combined. Notably, single markers may fall within normal ranges; a multifeature pattern provides better classification. The persistence of inflammatory and neutrophil signatures well beyond acute disease, coupled with ongoing antigen-specific B-cell activation and evolving autoreactivity, suggests sustained or recurrent antigenic stimulation and dysregulated immune homeostasis in PASC. These features imply that infPASC patients may preferentially benefit from therapies targeting neutrophil activity, inflammation, and pathological B-cell/autoimmune pathways. The difficulty of differentiating subtypes by symptoms alone underscores the need for biomarker-based classification to guide treatment and clinical trial design. The temporal dynamics remain uncertain; infPASC and niPASC may represent distinct conditions or phases within a disease continuum, potentially influenced by viral reservoirs’ location and immunologic context. Longitudinal, larger studies are required to clarify trajectories, mechanisms linking neutrophil dysregulation with coagulation abnormalities and autoimmunity, and to validate biomarker panels for clinical deployment.

This study integrates high-dimensional proteomics, cellular immunophenotyping, serology, autoantibody profiling, and machine learning to define robust immunologic signatures of PASC and to delineate two biologically distinct subtypes: inflammatory (infPASC) and non-inflammatory (niPASC). It identifies actionable pathways (neutrophil activity/NETs, coagulation, EGFR signaling, antigen processing) and demonstrates that a compact 12-marker panel can accurately classify infPASC, enabling scalable diagnostics and rational patient stratification for clinical trials. The findings suggest that targeted immunomodulation—particularly against neutrophil-driven inflammation and maladaptive B-cell responses—may benefit specific PASC patients. Future research should focus on longitudinal validation, mechanistic studies of viral persistence and immune dysregulation, refinement of classifier panels across diverse populations, and interventional trials stratified by immunologic subtype.

The study is cross-sectional, limiting causal and temporal inferences regarding whether infPASC and niPASC are distinct subtypes or phases in a disease continuum. Certain analyses involve modest subgroup sizes (e.g., follow-up ANA testing), and some reported associations rely on high-dimensional proteomics not universally available, though a reduced marker panel mitigates this. Symptomatology did not reliably distinguish subtypes, increasing risk of misclassification without biomarker data. Heterogeneity in time post–COVID-19 onset and potential confounders may influence immune signatures despite efforts to control for them. Generalizability requires validation in larger, multi-center, longitudinal cohorts with standardized diagnostic criteria.