Immune-metabolic trajectories delineate subgroups in paediatric long COVID

Categories: Disease & Virus

May 4, 2026

To interrogate paediatric LC heterogeneity, we analysed biologically motivated subgroups based on disease phase, EBV exposure, immune polarisation, and organ involvement, and assessed complementary clinical, immunological, metabolic, and functional readouts within each subgroup. We analysed 74 children with active LC symptoms and 27 controls, defining the index date as the last documented SARS-CoV-2 infection preceding LC onset; time since index infection at study visit (TsinceIndex) ranged up to 166 weeks (~3.2 years) (Fig. 1a–e; Supplementary Table 1a, b). Controls comprised healthy children (n = 14) and clinically stable children with cystic fibrosis (CF; n = 13), included as a respiratory infection-exposed comparator group²⁶. As pre-specified, both control subgroups were analysed as a single control arm; baseline inflammatory profiles were comparable between healthy and CF controls (Supplementary Table 1c), supporting pooling for subsequent analyses. Controls were slightly younger (median difference of 3.1 years). LC participants were re-assessed 3–6 months later (LC visit 2, follow-up). Symptom burden and functional impairment were quantified at each visit using validated patient- and parent-reported instruments and performance tests (Fig. 1e; Supplementary Fig. 1)²⁷. As an a priori marker of antiviral activity, we measured serum IFNα. IFNα was elevated in paediatric LC within the first year after the index infection compared with controls, whereas complement parameters, including C3, C4, and total haemolytic complement activity (CH50), remained within age-adjusted laboratory reference ranges across the observation period (Fig. 1d). Beyond 1 year of LC, IFNα levels were significantly lower (estimate = −6.02, 95% CI: −10.96 to −1.09; F(1, 95.39) = 5.83, P = 0.0177; R²m = 0.0196), consistent with waning antiviral signalling at later disease stages.

**Fig. 1: Demographic and clinical characterisation of paediatric participants.**

Patients assessed within 1 year of active LC showed no group-level improvement in physical or mental health measures (Fig. 1c, e; linear mixed-effects models (LMMs) details in Supplementary Fig. 1a, b and Supplementary Table 1b). Across the broader post-infection window (up to 3 years), we similarly observed no consistent improvement at the cohort level, although individual trajectories varied, with some patients showing modest improvement or decline. We next tested whether previously proposed risk factors for paediatric LC were associated with symptom severity. We fitted a pre-specified LMM with patient ID as a random intercept and included comorbidity status, sex and SARS-CoV-2 variant-of-concern (VOC) wave as fixed effects, adjusting for age, number of vaccinations and number of infections prior to LC onset (Supplementary Table 1d, lower panel). Across multiple pre-defined severity readouts, including the sit-to-stand test and questionnaire-based measures such as the Bell score, we found no evidence that these variables were associated with disease severity in this cohort (all P > 0.6). Thus, within our study, these proposed risk factors did not explain inter-individual differences in the severity of the LC phenotype.

Given the high prevalence of fatigue and dyspnoea in LC, we performed a protocolised assessment of potential cardiac and pulmonary involvement in paediatric LC (Figs. 1c and 2a, b)^8,28. All LC participants (n = 74) underwent a 12-lead electrocardiogram (ECG) and transthoracic echocardiography at presentation. ECG showed supraventricular extrasystoles in one participant. Echocardiography revealed findings in three patients (4.1%): mild mitral regurgitation; a bicuspid aortic valve with a coronary artery fistula; and mild diastolic dysfunction that resolved at follow-up. Paediatric cardiology review deemed these findings incidental and not explanatory of the LC phenotype.

**Fig. 2: Immune, pulmonary and endothelial features of paediatric LC.**

Pulmonary involvement was assessed using FEV1 z-scores standardised to paediatric reference values. Values were largely within the normal range, with only a minority at or below the lower limit of normal (z ≤ −1.64) (Fig. 2b), arguing against a cohort-level obstructive ventilatory defect. In an LMM accounting for repeated measures (random intercept: patient ID), FEV1 z-scores were not associated with comorbidity status, age, sex, VOC wave, number of infections prior LC onset, vaccination prior LC onset, or TsinceIndex window (2.1).

We next tested an a priori immunological framework centred on epithelial/type-2 signalling implicated in airway biology (IL-33, IL-13)²⁹ and systemic inflammation (IL-6) (Fig. 2c). Within the first year after TsinceIndex (index infection preceding LC onset), IL-13 and IL-33 were significantly higher in paediatric LC compared with controls after Holm–Bonferroni adjustment (P z-scores using LMM, including TsinceIndex and the selected cytokines (Fig. 2d; Supplementary Table 2.2). To avoid multicollinearity and maintain model interpretability, we applied a pre-specified rule to retain one representative of highly correlated predictors (r > 0.7); accordingly, IL-13 was kept as the representative type-2 effector signal, and IL-33 was excluded from the final model (Supplementary Table 2.2)^15,16. The final model explained 16% of the variance through fixed effects (R²m = 0.1622), whereas R²c (0.7210) highlighted substantial between-individual variability captured by the random intercept. Additional adjustment for the pre-specified covariates did not materially change the associations (Supplementary Table 2.2b). Notably, higher FEV1 z-scores were associated with lower IL-6 and higher IL-13, consistent with concurrent pro-inflammatory signalling and type-2–linked tissue-remodelling programmes potentially contributing to inter-individual heterogeneity in pulmonary function.

To characterise systemic immune activation in paediatric LC during the pre-specified primary analysis window of the first year after the LC-index SARS-CoV-2 infection, we quantified circulating serum cytokines and grouped them into pre-defined functional families, each represented by an a priori selected lead cytokine (SARS-CoV-2-associated, Th1/Th2 balance, innate-like, Th17/Th22-based and regulatory-like; Table 1, Supplementary Fig. 2). These cytokine family analyses within the first-year window constituted the pre-specified primary cytokine objective. Multiple testing was controlled within each family (Supplementary Table 2.3). The 1-year window was chosen to maximise comparability across studies and to capture the phase in which SARS-CoV-2-specific immune signals are expected to wane^28,30.

Table 1 Lead cytokines across pre-defined functional cytokine families in paediatric LC

Within the SARS-CoV-2-associated family, the lead cytokine IL-13 was elevated in LC within the first year and declined thereafter to levels comparable to controls (Fig. 2c). IL-33 showed a similar transient increase, whereas IL-6 did not differ significantly from controls. Based on prior reports in young adults³¹, we additionally assessed anti-CCP and anti-Transglutaminase (anti-TransG) autoantibodies; both remained within the reference range and were not increased in paediatric LC compared with pooled controls (Fig. 2e; Supplementary Table 3.3a). These findings were robust in sensitivity analyses evaluating healthy and CF control subgroups separately against LC (Supplementary Table 3.3b, c; Supplementary Fig. 4).

Across families, additional pre-specified lead markers were significant within the first year, including the IL-4/IFNγ ratio (Th2/Th1 balance), IL-1β (innate-like), IL-11 (regulatory-like) and IL-12p40 (Th17/Th22-based) (Fig. 2f; Table 1). Family-level analyses revealed increased Th2-associated cytokines (IL-4 and IL-5), whereas Th1-associated cytokines (IFNγ, IL-2 and IL-12p70) were not elevated (Supplementary Fig. 2a, b). Among innate-like cytokines, IL-1α increased while GM-CSF was unchanged (Fig. 2g; Supplementary Fig. 2c). Within the regulatory-like family, IL-11 increased, whereas IL-10, IL-15, IL-18 and IL-27 did not (Fig. 2h; Supplementary Fig. 2c). Within the Th17/Th22-based family, IL-12p40 remained significant after Holm–Bonferroni adjustment (adjusted P = 0.048), while IL-17A, IL-17F, IL-22 and IL-23 did not (Fig. 2i; Supplementary Table 2.3e). Collectively, these data define a first-year LC signature characterised by immunological dysregulation with distinct cytokine patterns.

Paediatric LC beyond the first year

Subsequently, we monitored systemic parameters after 1 year of LC to delineate the evolving immunopathological phases. Therefore, we performed two complementary analyses: a cross-sectional comparison of cytokine levels in the LC 1–3 year group versus controls (Fig. 2c, f–i; Supplementary Fig. 2; Supplementary Table 2.3 for within-family multiple comparisons), and a LMM within the LC cohort with time window based on TsinceIndex (2c, f–i; Supplementary Table 2.4). Lead-cytokine comparisons are summarised in Table 1. The previously elevated IFNα levels declined, indicating a waning antiviral response (Fig. 1d), in parallel with a decrease in the lead cytokine IL-13 within the SARS-CoV-2-associated cytokine family (Fig. 2c). In addition, the innate-like lead cytokine IL-1β was significantly upregulated, as was IL-1α within this family, but not GM-CSF. Within the Th17 axis, lead IL-12p40 and IL-23 were increased, but IL-22 significantly decreased after 1 year of LC. Still, the IL-4/IFNγ ratio (Fig. 2f) was increased and remained elevated throughout, suggesting a transition towards a Th2/17-driven state. Taken together, these findings outline an early antiviral and Th2-skewed response within the first year shifts towards a waning SARS-CoV-2 response showing innate- and Th2/17-oriented responses. This is consistent with the progressive immunopathological remodelling associated with chronicity (Fig. 2j).

Next, we fitted a pre-specified LMM (random intercept: patient ID) to quantify determinants of LC severity, using the Bell score as the dependent variable (Fig. 2k; Supplementary Table 2.5). We hypothesised that functional impairment in paediatric LC reflects sustained immune activation and granulocyte perturbations, and that prior EBV infection may modify these associations through immune imprinting. Accordingly, covariates were defined a priori to represent complementary biological axes rather than being selected by univariable screening: a marker of prior EBV exposure (anti-EBV EBNA; exposure history, not reactivation), mean corpuscular haemoglobin concentration (MCHC)³², as a haematological readout, IL-12p40 as a robust proxy of IL-12/IL-23–axis activity, and basophil granulocyte counts as an indicator of granulocyte lineage shifts. The fixed effects explained 21% of Bell score variance (R²m = 0.2121), and each covariate was significantly associated with the Bell score. Notably, anti-EBV EBNA and MCHC decreased with clinical improvement, whereas basophil counts and IL-12p40 increased. To assess robustness and potential effect modification, we next fitted an expanded model that adjusted for key confounders (age, sex, comorbidity, and TsinceIndex; Supplementary Table 2.5b). Based on accumulating evidence implicating mitochondrial dysfunction in LC and the essential role of thiamine (vitamin B1) in mitochondrial energy metabolism^33,34, we additionally included vitamin B1 as a pre-specified metabolic covariate. Finally, to directly test the hypothesis that prior EBV exposure modifies specific biological correlates of severity, we incorporated a priori anti-EBV EBNA interaction terms. In this fully adjusted interaction model, most interaction terms were not statistically significant; however, anti-EBV EBNA showed significant interactions with MCHC and vitamin B1 (Supplementary Table 2.5b), motivating targeted follow-up analyses stratified by EBV exposure.

Autoantibodies (aAb) and organ Injury in children with active LC

To further test the hypothesis that paediatric LC comprises biologically heterogeneous strata, including subsets with potential CNS-related symptomatology, we measured serum neurofilament light chain (NfL) as a pre-specified exploratory circulating marker (Fig. 3a; Supplementary Fig. 3a). In parallel, we assessed disease-associated autoantibody profiles as a pre-specified autoantibody axis to evaluate autoreactive immune signatures. NfL was interpreted as a hypothesis-generating marker of potential neuro-axonal injury. NfL concentrations were converted to age-adjusted z-scores using the Basel paediatric reference dataset and expressed as percentiles relative to healthy children³⁵. In a one-sample comparison against the reference median (P50), we did not observe a significant upward shift at the cohort level (W = 1084, n = 73, P = 0.20; r = 0.15). A subset of children fell into the upper tail of the age-referenced distribution (12.3%; 9/73 > P90; Supplementary Fig. 3a), consistent with inter-individual heterogeneity rather than a uniform cohort-wide shift. In exploratory analyses, NfL percentile was inversely associated with functional status (Pearson’s r = −0.3536, P = 0.0060; Fig. 3a). Using a clinically anchored threshold for severe impairment (Bell score ≤40), children below this threshold displayed higher NfL percentiles (P = 0.0058; Supplementary Fig. 3a, left). NfL percentiles were not associated with TsinceIndex (P = 0.8727; Supplementary Fig. 3a, right). Collectively, these findings indicate heterogeneity in NfL and an exploratory association with clinical impairment, motivating mechanistic follow-up to clarify drivers and clinical relevance.

**Fig. 3: Systemic autoantibody profiles and endothelial features of paediatric LC.**

In separate analyses, to probe a potential autoimmune contribution to autonomic involvement, aAbs against GPCR (GPCR-aAb) targeting β1/β2-adrenergic and M3/M4-muscarinic receptors were quantified by ELISA (Supplementary Fig. 3b). LMMs showed that these aAb were not associated with the pre-specified early TsinceIndex window (≤1 year), systemic cytokine levels, or disease severity (Supplementary Table 3.1). Moreover, aAb levels did not differ between LC TsinceIndex windows (3b; Supplementary Table 3.2). Next, we analysed additional aAbs as surrogate markers of vasculitis (anti-proteinase 3 (PR3) and anti-myeloperoxidase (MPO)) and antiphospholipid syndrome (anti-cardiolipin and anti–β2-glycoprotein I (anti-β2GPI)) (Figs. 2e and 3b; Supplementary Table 3.3). None of these markers, including anti-CCP and anti-Transglutaminase (anti-TransG) assessed in Fig. 2e, were elevated relative to controls, and all remained below the assay cut-off during the first year of paediatric LC (Supplementary Table 3.3a). These null findings were unchanged in sensitivity analyses using healthy and clinically stable cystic fibrosis controls as separate comparator groups (Supplementary Table 3.3b, c; Supplementary Fig. 4). Within-LC comparisons using LMMs, with TsinceIndex window as a fixed effect and participant ID as a random intercept, likewise showed no increase in aAb levels between LC 2e and 3b and Supplementary Table 2.4f). To exploratorily assess whether enhanced autoreactivity is generally present in active paediatric LC, we quantified the prevalence of aAb positivity across individuals (Fig. 3c; Supplementary Tables 2.3 and 3.3) and observed no significant increase compared with controls, arguing against an aAb-mediated organ-injury phenotype in this cohort.

Within the pre-specified aAb axis, we additionally evaluated isolated anti-DFS70 reactivity (anti-DFS70 (LEDGF/p75) aAbs) as a potential marker of a distinct, potentially benign autoreactivity signature (Fig. 3d). Notably, none of the controls were anti-DFS70-positive, whereas 11% of patients with active paediatric LC tested positive; anti-DFS70 positivity persisted at a follow-up visit months later. We therefore stratified LC patients by anti-DFS70 status (Fig. 3e; Supplementary Fig. 3c–e; Supplementary Table 3.4). Within the LC cohort, anti-DFS70-negative patients showed higher von Willebrand factor activity/frequency and elevated factor VIII levels, while other haemostatic parameters (antithrombin III, fibrinogen, D-dimer, aPTT, protein C, free protein S) and complement components (C3, C4) remained unchanged. Because the extended coagulation panel was not available for controls, these analyses reflect within-cohort stratification rather than case–control differences. EBV-related antibodies were independent of anti-DFS70 status in LC patients. Together, these data indicate a coagulation-factor signature in a defined LC subgroup and suggest that anti-DFS70 reactivity, classically associated with non-systemic autoimmunity, may capture a distinct, potentially less-pathogenic immune state in paediatric LC³⁶.

Molecular landscape of EBV-associated paediatric LC

EBV has been implicated in post-viral immune dysregulation and is a plausible modifier of paediatric LC biology. We therefore assessed whether EBV exposure contributes to the molecular and clinical landscape of paediatric LC by modelling EBV serostatus (anti-EBV EBNA binary) as a covariate in LMMs, with repeated measures clustered by patient ID, across pre-defined mechanistic readout families, adjusting for sex, age, TsinceIndex, vaccination status and comorbidity (Fig. 4, left; Table 2; Supplementary Table 4). Multiple testing was controlled within each family (Holm–Bonferroni), and the composite summary highlights only readouts with significant overall model fit and a significant positive contribution of EBV exposure (Fig. 4, left; Table 2; Supplementary Table 4).

**Fig. 4: Biological EBV-linked subgroups of paediatric LC.**

Table 2 Associations of EBV exposure status with cytokines, granulocytes, and autoantibody levels in paediatric LC

Using this framework, EBV exposure was associated with a pronounced inflammatory signature in paediatric LC (Fig. 4, left; Table 2; Supplementary Table 4). EBV-experienced patients showed an innate-inflammatory cytokine profile with elevated IL-1α and IL-15, accompanied by higher IFNα and IL-18 (Table 2; Supplementary Table 4). EBV exposure was further linked to a Th17/Th22-associated pattern characterised by increased IL-12p40 and IL-22, whereas IL-17A/F did not show EBV-dependent differences after Holm–Bonferroni adjustment (Table 2; Supplementary Table 4). This was paralleled by increased regulatory cytokines (IL-11, IL-10 and IL-27), consistent with a concomitant counter-regulatory response (Table 2; Supplementary Table 4). Cellular innate activation was supported by higher neutrophil counts, with no increase in eosinophils or basophils (Table 2; Supplementary Table 4); across EBV-associated readouts, TsinceIndex, age and sex frequently contributed as covariates, whereas comorbidity did not, and vaccination status contributed only to neutrophil variation (Fig. 4, left; Supplementary Table 4). In contrast, SARS-CoV-2-specific humoral readouts and complement measures, as well as aAbs, did not show EBV-dependent effects in this cohort/TsinceIndex window (Table 2; Supplementary Table 4). As chronic EBV responses are known to increase the risk of depression in adolescents³⁷, and depression and anxiety negatively impact recovery in children with chronic fatigue syndrome³⁸, the model was applied to mental health scores (Table 2; Supplementary Table 4). EBV serostatus was not associated with worse mental health scores after Holm–Bonferroni adjustment (Table 2). To synthesise these findings, we generated a composite overview integrating readouts with a significant overall model fit and a significant positive contribution of EBV exposure (Fig. 4, left; Table 2; Supplementary Table 4). This composite highlights a predominantly innate-inflammatory and IL-12p40/IL-22-skewed cytokine landscape in EBV-experienced paediatric LC patients, accompanied by increased neutrophils and regulatory cytokines.

Verification of subgroups of paediatric LC

We identified three features by which LC subgroups can be categorised: The first relates to TsinceIndex, where LC in the first year is characterised by Th2-, innate and viral-associated cytokines, and LC persisting for 1–3.2 years is characterised by Th2-biased, Th17-related and innate-like systemic cytokines. A second relates to anti-DFS70-positivity, which was associated with markedly fewer coagulation abnormalities. The third involves EBV serostatus, which was linked to a pro-inflammatory cytokine profile and granulocytic dysregulation. To further explore these separately immuno-clinical subgroups in paediatric LC, we systematically analysed a panel of 43 haematopoietic, coagulation, electrolyte, and vitamin-related biomarkers (Table 3, Supplementary Table 5a, b). Using LMMs for each parameter, we tested for associations with subgroup assignment, EBV exposure status, anti-DFS70 autoantibody status, and time window based on TsinceIndex (−, Ca²⁺, bicarbonate), vitamins (B1, B6, B12, D, folic acid), metabolic markers (glucose, lactate, creatinine, eGFR), and inflammation/coagulation markers (CRP, D-dimer, fibrinogen, Factor VIII, protein S). After correcting for multiple testing, three metabolic markers were found to be significantly associated with the defined subgroups: TSH (adjusted P = 0.0043), aPTT (adjusted P = 0.0043) and LP(a) (adjusted P = 0.0123) (Table 3). LP(a) and TSH levels increased over time, whereas aPTT decreased (Supplementary Table 5b). Significant models were adjusted for age and sex (Supplementary Table 5b). In the aPTT model, male sex was associated with higher aPTT values and improved the R²m by 0.084. In the LP(a) model, sex was statistically significant but contributed little to model fit, whereas sex was not significant in the TSH model. Age was not a significant predictor in any model. These biomarkers exhibited non-overlapping association patterns, supporting the concept of pathophysiological divergence. Together, these findings provide evidence for biologically differentiated subgroup structure within paediatric LC, although discrete endotypes remain to be established.

Table 3 Three of 43 clinical metabolic and coagulation parameters remain significant after mixed-effects modelling and multiple-testing adjustment using the Holm–Bonferroni method

EBV-naïve/low subgroup reveals immunometabolic correlates of functional status

Leveraging the anti-EBV EBNA–negative/low status observed in ~50% of our paediatric LC cohort, we applied LMMs used for disease severity in Fig. 2k (Bell score as dependent variable) to this subgroup, omitting EBV status as a covariate and retaining IL-12p40, MCHC, basophil counts and vitamin B1 (thiamine) as predictors (Supplementary Table 6). The model was highly significant (P R²m of 0.3180 (Fig. 4, right; Supplementary Table 6a), supporting an immunometabolic contribution to clinical severity within EBV-naïve/low participants. IL-12p40 levels and basophil counts were the strongest positive predictors of Bell score, whereas higher MCHC was associated with lower Bell scores, consistent with poorer functional status at higher MCHC. Adding potential confounders (comorbidity, vaccination, sex, age and TsinceIndex) did not materially improve the model fit (R²m = 0.3169; P 6b). Together, these data identify an EBV-naïve/low subgroup within paediatric LC with severity-linked immunometabolic and haematological features, including higher IL-12p40, basophil counts, vitamin B1 and anti-DFS70 positivity in participants with better functional status.

Source link