One in ten severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infections results in post-acute sequelae of coronavirus disease 2019 (PASC) or long coronavirus disease (COVID), which affects 65 million people worldwide1. Long COVID (LC) remains common, even after mild acute infection with recent variants2, and it is likely LC will continue to cause substantial long-term ill health, requiring targeted management based on an understanding of how disease phenotypes relate to underlying mechanisms. Persistent inflammation has been reported in adults with LC1,3, but studies have been limited in size, timing of samples or breadth of immune mediators measured, leading to inconsistent or absent associations with symptoms. Markers of oxidative stress, metabolic disturbance, vasculoproliferative processes and IFN-, NF-κB- or monocyte-related inflammation have been suggested3,4,5,6.

The PHOSP-COVID study, a multicenter United Kingdom study of patients previously hospitalized with COVID-19, has reported inflammatory profiles in 626 adults with health impairment after COVID-19, identified through clustering. Elevated IL-6 and markers of mucosal inflammation were observed in those with severe impairment compared with individuals with milder impairment7. However, LC is a heterogeneous condition that may be a distinct form of health impairment after COVID-19, and it remains unclear whether there are inflammatory changes specific to LC symptom subtypes. Determining whether activated inflammatory pathways underlie all cases of LC or if mechanisms differ according to clinical presentation is essential for developing effective therapies and has been highlighted as a top research priority by patients and clinicians8.

In this Letter, in a prospective multicenter study, we measured 368 plasma proteins in 657 adults previously hospitalized for COVID-19 (Fig. 1a and Table 1). Individuals in our cohort experienced a range of acute COVID-19 severities based on World Health Organization (WHO) progression scores9; WHO 3–4 (no oxygen support, n = 133 and median age of 55 years), WHO 5–6 (oxygen support, n = 353 and median age of 59 years) and WHO 7–9 (critical care, n = 171 and median age of 57 years). Participants were hospitalized for COVID-19 ≥3 months before sample collection (median 6.1 months, interquartile range (IQR) 5.1–6.8 months and range 3.0–8.3 months) and confirmed clinically (n = 36/657) or by PCR (n = 621/657). Symptom data indicated 233/657 (35%) felt fully recovered at 6 months (hereafter ‘recovered’) and the remaining 424 (65%) reported symptoms consistent with the WHO definition for LC (symptoms ≥3 months post infection10). Given the diversity of LC presentations, patients were grouped according to symptom type (Fig. 1b). Groups were defined using symptoms and health deficits that have been commonly reported in the literature1 (Methods). A multivariate penalized logistic regression model (PLR) was used to explore associations of clinical covariates and immune mediators at 6 months between recovered patients (n = 233) and each LC group (cardiorespiratory symptoms, cardioresp, n = 398, Fig. 1c; fatigue, n = 384, Fig. 1d; affective symptoms, anxiety/depression, n = 202, Fig. 1e; gastrointestinal symptoms, GI, n = 132, Fig. 1f; and cognitive impairment, cognitive, n = 61, Fig. 1g). Women (n = 239) were more likely to experience CardioResp (odds ratio (OR 1.14), Fatigue (OR 1.22), GI (OR 1.13) and Cognitive (OR 1.03) outcomes (Fig. 1c,d,f,g). Repeated cross-validation was used to optimize and assess model performance (Methods and Extended Data Fig. 1). Pre-existing conditions, such as chronic lung disease, neurological disease and cardiovascular disease (Supplementary Table 1), were associated with all LC groups (Fig. 1c–g). Age, C-reactive protein (CRP) and acute disease severity were not associated with any LC group (Table 1).

Fig. 1: Subtypes of LC are associated with distinct inflammatory profiles.
figure 1

a, Distribution of time from COVID-19 hospitalization at sample collection. All samples were cross-sectional. The vertical red line indicates the 3 month cutoff used to define our final cohort and samples collected before 3 months were excluded. b, An UpSet plot describing pooled LC groups. The horizontal colored bars represent the number of patients in each symptom group: cardiorespiratory (Cardio_Resp), fatigue, cognitive, GI and anxiety/depression (Anx_Dep). Vertical black bars represent the number of patients in each symptom combination group. To prevent patient identification, where less than five patients belong to a combination group, this has been represented as ‘<5’. The recovered group (n = 233) were used as controls. cg, Forest plots of Olink protein concentrations (NPX) associated with Cardio_Resp (n = 365) (c), fatigue (n = 314) (d), Anx_Dep (n = 202) (e), GI (n = 124) (f) and cognitive (n = 60) (g). Neuro_Psych, neuropsychiatric. The error bars represent the median accuracy of the model. h,i, Distribution of Olink values (NPX) for IL-1R2 (h) and MATN2, neurofascin and sCD58 (i) measured between symptomatic and recovered individuals in recovered (n = 233), Cardio_Resp (n = 365), fatigue (n = 314) and Anx_Dep (n = 202) groups (h) and MATN2 in GI (n = 124), neurofascin in cognitive (n = 60) and sCD58 in Cardio_Resp and recovered groups (i). The box plot center line represents the median, the boundaries represent IQR and the whisker length represents 1.5× IQR. The median values were compared between groups using two-sided Wilcoxon signed-rank test, *P < 0.05, **P < 0.01, ***P < 0.001 and ****P < 0.0001.

Table 1 Cohort demographics

To study the association of peripheral inflammation with symptoms, we analyzed cross-sectional data collected approximately 6 months after hospitalizations. We measured 368 immune mediators from plasma collected contemporaneously with symptom data. Mediators suggestive of myeloid inflammation were associated with all symptoms (Fig. 1c–h). Elevated IL-1R2, an IL-1 receptor expressed by monocytes and macrophages modulating inflammation11 and MATN2, an extracellular matrix protein that modulates tissue inflammation through recruitment of innate immune cells12, were associated with cardioresp (IL-1R2 OR 1.14, Fig. 1c,h), fatigue (IL-1R2 OR 1.45, Fig. 1d,h), anxiety/depression (IL-1R2 OR 1.34. Fig. 1e,h) and GI (MATN2 OR 1.08, Fig. 1f). IL-3RA, an IL-3 receptor, was associated with cardioresp (OR 1.07, Fig. 1c), fatigue (OR 1.21, Fig. 1d), anxiety/depression (OR 1.12, Fig. 1e) and GI (OR 1.06, Fig. 1f) groups, while CSF3, a cytokine promoting neutrophilic inflammation13, was elevated in cardioresp (OR 1.06, Fig. 1c), fatigue (OR 1.12, Fig. 1d) and GI (OR 1.08, Fig. 1f).

Elevated COLEC12, which initiates inflammation in tissues by activating the alternative complement pathway14, associated with cardioresp (OR 1.09, Fig. 1c), fatigue (OR 1.19, Fig. 1d) and anxiety/depression (OR 1.11, Fig. 1e), but not with GI (Fig. 1f) and only weakly with cognitive (OR 1.02, Fig. 1g). C1QA, a degradation product released by complement activation15 was associated with GI (OR 1.08, Fig. 1f) and cognitive (OR 1.03, Fig. 1g). C1QA, which is known to mediate dementia-related neuroinflammation16, had the third strongest association with cognitive (Fig. 1g). These observations indicated that myeloid inflammation and complement activation were associated with LC.

Increased expression of DPP10 and SCG3 was observed in the GI group compared with recovered (DPP10 OR 1.07 and SCG3 OR 1.08, Fig. 1f). DPP10 is a membrane protein that modulates tissue inflammation, and increased DPP10 expression is associated with inflammatory bowel disease17,18, suggesting that GI symptoms may result from enteric inflammation. Elevated SCG3, a multifunctional protein that has been associated with irritable bowel syndrome19, suggested that noninflammatory disturbance of the brain–gut axis or dysbiosis, may occur in the GI group. The cognitive group was associated with elevated CTSO (OR 1.04), NFASC (OR 1.03) and SPON-1 (OR 1.02, Fig. 1g,i). NFASC and SPON-1 regulate neural growth20,21, while CTSO is a cysteine proteinase supporting tissue turnover22. The increased expression of these three proteins as well as C1QA and DPP10 in the cognitive group (Fig. 1g) suggested neuroinflammation and alterations in nerve tissue repair, possibly resulting in neurodegeneration. Together, our findings indicated that complement activation and myeloid inflammation were common to all LC groups, but subtle differences were observed in the GI and cognitive groups, which may have mechanistic importance. Acutely elevated fibrinogen during hospitalization has been reported to be predictive of LC cognitive deficits23. We found elevated fibrinogen in LC relative to recovered (Extended Data Fig. 2a; P = 0.0077), although this was not significant when restricted to the cognitive group (P = 0.074), supporting our observation of complement pathway activation in LC and in keeping with reports that complement dysregulation and thrombosis drive severe COVID-19 (ref. 24).

Elevated sCD58 was associated with lower odds of all LC symptoms and was most pronounced in cardioresp (OR 0.85, Fig. 1c,i), fatigue (OR 0.80, Fig. 1d) and anxiety/depression (OR 0.83, Fig. 1e). IL-2 was negatively associated with the cardioresp (Fig. 1c, OR 0.87), fatigue (Fig. 1d, OR 0.80), anxiety/depression (Fig. 1e, OR 0.84) and cognitive (Fig. 1g, OR 0.96) groups. Both IL-2 and sCD58 have immunoregulatory functions25,26. Specifically, sCD58 suppresses IL-1- or IL-6-dependent interactions between CD2+ monocytes and CD58+ T or natural killer cells26. The association of sCD58 with recovered suggests a central role of dysregulated myeloid inflammation in LC. Elevated markers of tissue repair, IDS and DNER27,28, were also associated with recovered relative to all LC groups (Fig. 1c–g). Taken together, our data suggest that suppression of myeloid inflammation and enhanced tissue repair were associated with recovered, supporting the use of immunomodulatory agents in therapeutic trials29 (Supplementary Table 2).

We next sought to validate the experimental and analytical approaches used. Although Olink has been validated against other immunoassay platforms, showing superior sensitivity and specificity30,31, we confirmed the performance of Olink against chemiluminescent immunoassays within our cohort. We performed chemiluminescent immunoassays on plasma from a subgroup of 58 participants (recovered n = 13 and LC n = 45). There were good correlations between results from Olink (normalized protein expression (NPX)) and chemiluminescent immunoassays (pg ml−1) for CSF3, IL-1R2, IL-3RA, TNF and TFF2 (Extended Data Fig. 3). Most samples did not have concentrations of IL-2 detectable using a mesoscale discovery chemiluminescent assay, limiting this analysis to 14 samples (recovered n = 4, LC n = 10, R = 0.55 and P = 0.053, Extended Data Fig. 3). We next repeated our analysis using alternative definitions of LC. The Centers for Disease Control and Prevention and National Institute for Health and Care Excellence definitions for LC include symptoms occurring 1 month post infection32,33. Using the 1 month post-infection definition included 62 additional participants to our analysis (recovered n = 21, 3 females and median age 61 years and LC n = 41, 15 females and median age 60 years, Extended Data Fig. 2c) and found that inflammatory associations with each LC group were consistent with our analysis based on the WHO definition (Extended Data Fig. 2d–h). Finally, to validate the analytical approach (PLR) we examined the distribution of data, prioritizing proteins that were most strongly associated with each LC/recovered group (IL-1R2, MATN2, NFASC and sCD58). Each protein was significantly elevated in the LC group compared with recovered (Fig. 1h,i and Extended Data Fig. 4), consistent with the PLR. Alternative regression approaches (unadjusted regression models and partial least squares, PLS) reported results consistent with the original analysis of protein associations and LC outcome in the WHO-defined cohort (Fig. 1c–g, Supplementary Table 3 and Extended Data Figs. 5 and 6). The standard errors of PLS estimates were wide (Extended Data Fig. 6), consistent with previous demonstrations that PLR is the optimal method to analyze high-dimensional data where variables may have combined effects34. As inflammatory proteins are often colinear, working in-tandem to mediate effects, we prioritized PLR results to draw conclusions.

To explore the relationship between inflammatory mediators associated with different LC symptoms, we performed a network analysis of Olink mediators highlighted by PLR within each LC group. COLEC12 and markers of endothelial and mucosal inflammation (MATN2, PCDH1, ROBO1, ISM1, ANGPTL2, TGF-α and TFF2) were highly correlated within the cardioresp, fatigue and anxiety/depression groups (Fig. 2 and Extended Data Fig. 7). Elevated PCDH1, an adhesion protein modulating airway inflammation35, was highly correlated with other inflammatory proteins associated with the cardioresp group (Fig. 2), suggesting that systemic inflammation may arise from the lung in these individuals. This was supported by increased expression of IL-3RA, which regulates innate immune responses in the lung through interactions with circulating IL-3 (ref. 36), in fatigue (Figs. 1d and 2), which correlated with markers of tissue inflammation, including PCDH1 (Fig. 2). MATN2 and ISM1, mucosal proteins that enhance inflammation37,38, were highly correlated in the GI group (Fig. 2), highlighting the role of tissue-specific inflammation in different LC groups. SCG3 correlated less closely with mediators in the GI group (Fig. 2), suggesting that the brain–gut axis may contribute separately to some GI symptoms. SPON-1, which regulates neural growth21, was the most highly correlated mediator in the cognitive group (Fig. 2 and Extended Data Fig. 7), highlighting that processes within nerve tissue may underlie this group. These observations suggested that inflammation might arise from mucosal tissues and that additional mechanisms may contribute to pathophysiology underlying the GI and cognitive groups.

Fig. 2: Network analyses define key immune mediators in LC symptom groups.
figure 2

Network analysis of Olink mediators associated with cardioresp (n = 365), fatigue (n = 314), anxiety/depression (n = 202), GI (n = 124) and cognitive groups (n = 60). Each node corresponds to a protein mediator identified by PLR. The edges (blue lines) were weighted according to the size of Spearman’s rank correlation coefficient between proteins. All edges represent positive and significant correlations (P < 0.05) after FDR adjustment.

Women were more likely to experience LC (Table 1), as found in previous studies1. As estrogen can influence immunological responses39, we investigated whether hormonal differences between men and women with LC in our cohort explained this trend. We grouped men and women with LC symptoms into two age groups (those younger than 50 years and those 50 years and older, using age as a proxy for menopause status in women) and compared mediator levels between men and women in each age group, prioritizing those identified by PLR to be higher in LC compared with recovered. As we aimed to understand whether women with LC had stronger inflammatory responses than men with LC, we did not assess differences in men and women in the recovered group. IL-1R2 and MATN2 were significantly higher in women ≥50 years than men ≥50 years in the cardioresp group (Fig. 3a, IL-1R2 and MATN2) and the fatigue group (Fig. 3b). In the GI group, CSF3 was higher in women ≥50 years compared with men ≥50 years (Fig. 3c), indicating that the inflammatory markers observed in women were not likely to be estrogen-dependent. Women have been reported to have stronger innate immune responses to infection and to be at greater risk of autoimmunity39, possibly explaining why some women in the ≥50 years group had higher inflammatory proteins than men the same group. Proteins associated with the anxiety/depression (IL-1R2 P = 0.11 and MATN2 P = 0.61, Extended Data Fig. 8a) and cognitive groups (CTSO P = 0.64 and NFASC P = 0.41, Extended Data Fig. 8b) were not different between men and women in either age group, consistent with the absent/weak association between sex and these outcomes identified by PLR (Fig. 1e,g). Though our findings suggested that nonhormonal differences in inflammatory responses may explain why some women are more likely to have LC, they require confirmation in adequately powered studies.

Fig. 3: Elevated immune mediator levels are most pronounced in older women with LC.
figure 3

ac, Olink-measured plasma protein levels (NPX) of IL-1R2 and MATN2 (a and b) and CSF3 (c) between LC men and LC women divided by age (<50 or ≥50 years) in the cardiorespiratory group (<50 years n = 8 and ≥50 years n = 270) (a), fatigue group (<50 years n = 81 and ≥50 years n = 227) (b) and GI group (<50 years n = 34 and ≥50 years n = 82) (c). the median values were compared between men and women using two-sided Wilcoxon signed-rank test, *P < 0.05, **P < 0.01, ***P < 0.001 and ****P < 0.0001. The box plot center line represents the median, the boundaries represent IQR and the whisker length represents 1.5× IQR.

To test whether local respiratory tract inflammation persisted after COVID-19, we compared nasosorption samples from 89 participants (recovered, n = 31; LC, n = 33; and healthy SARS-CoV-2 naive controls, n = 25, Supplementary Tables 4 and 5). Several inflammatory markers were elevated in the upper respiratory tract post COVID (including IL-1α, CXCL10, CXCL11, TNF, VEGF and TFF2) when compared with naive controls, but similar between recovered and LC (Fig. 4a). In the cardioresp group (n = 29), inflammatory mediators elevated in plasma (for example, IL-6, APO-2, TGF-α and TFF2) were not elevated in the upper respiratory tract (Extended Data Fig. 9a) and there was no correlation between plasma and nasal mediator levels (Extended Data Fig. 9b). This exploratory analysis suggested upper respiratory tract inflammation post COVID was not specifically associated with cardiorespiratory symptoms.

Fig. 4: Pronounced mucosal inflammation after COVID-19 is not associated with LC.
figure 4

a, Nasal cytokines measured by immunoassay in post-COVID participants (n = 64) compared with healthy SARS-CoV-2 naive controls (n = 25), and between the the cardioresp group (n = 29) and the recovered group (n = 31). The red values indicate significantly increased cytokine levels after FDR adjustment (P < 0.05) using two-tailed Wilcoxon signed-rank test. b, SARS-CoV-2 N antigen measured in sputum by electrochemiluminescence from recovered (n = 17) and pooled LC (n = 23) groups, compared with BALF from SARS-CoV-2 naive controls (n = 9). The horizontal dashed line indicates the lower limit of detection of the assay. c, Plasma S- and N-specific IgG responses measured by electrochemiluminescence in the LC (n = 35) and recovered (n = 19) groups. The median values were compared using two-sided Wilcoxon signed-rank tests, NS P > 0.05, *P < 0.05, **P < 0.01, ***P < 0.001 and ****P < 0.0001. The box plot center lines represent the median, the boundaries represent IQR and the whisker length represents 1.5× IQR.

To explore whether SARS-CoV-2 persistence might explain the inflammatory profiles observed in the cardioresp group, we measured SARS-CoV-2 nucleocapsid (N) antigen in sputum from 40 participants (recovered n = 17 and LC n = 23) collected approximately 6 months post hospitalization (Supplementary Table 6). All samples were compared with prepandemic bronchoalveolar lavage fluid (n = 9, Supplementary Table 4). Only four samples (recovered n = 2 and LC n = 2) had N antigen above the assay’s lower limit of detection, and there was no difference in N antigen concentrations between LC and recovered (Fig. 4b, P = 0.78). These observations did not exclude viral persistence, which might require tissues samples for detection40,41. On the basis of the hypothesis that persistent viral antigen might prevent a decline in antibody levels over time, we examined the titers of SARS-CoV-2-specific antibodies in unvaccinated individuals (recovered n = 19 and LC n = 35). SARS-CoV-2 N-specific (P = 0.023) and spike (S)-specific (P = 0.0040) immunoglobulin G (IgG) levels were elevated in LC compared with recovered (Fig. 4c).

Overall, we identified myeloid inflammation and complement activation in the cardioresp, fatigue, anxiety/depression, cognitive and GI groups 6 months after hospitalization (Extended Data Fig. 10). Our findings build on results of smaller studies5,6,42 and are consistent with a genome-wide association study that identified an independent association between LC and FOXP4, which modulates neutrophilic inflammation and immune cell function43,44. In addition, we identified tissue-specific inflammatory elements, indicating that myeloid disturbance in different tissues may result in distinct symptoms. Multiple mechanisms for LC have been suggested, including autoimmunity, thrombosis, vascular dysfunction, SARS-CoV-2 persistence and latent virus reactivation1. All these processes involve myeloid inflammation and complement activation45. Complement activation in LC has been suggested in a proteomic study in 97 mostly nonhospitalized COVID-19 cases42 and a study of 48 LC patients, of which one-third experienced severe acute disease46. As components of the complement system are known to have a short half-life47, ongoing complement activation suggests active inflammation rather than past tissue damage from acute infection.

Despite the heterogeneity of LC and the likelihood of coexisting or multiple etiologies, our work suggests some common pathways that might be targeted therapeutically and supports the rationale for several drugs currently under trial. Our finding of increased sCD58 levels (associated with suppression of monocyte–lymphocyte interactions26) in the recovered group, strengthens our conclusion that myeloid inflammation is central to the biology of LC and that trials of steroids, IL-1 antagonists, JAK inhibitors, naltrexone and colchicine are justified. Although anticoagulants such as apixaban might prevent thrombosis downstream of complement dysregulation, they can also increase the risk of serious bleeding when given after COVID-19 hospitalization48. Thus, clinical trials, already underway, need to carefully assess the risks and benefits of anticoagulants (Supplementary Table 2).

Our finding of elevated S- and N-specific IgG in LC could suggest viral persistence, as found in other studies6,42,49. Our network analysis indicated that inflammatory proteins in the cardioresp group interacted strongly with ISM1 and ROBO1, which are expressed during respiratory tract infection and regulate lung inflammation50,51. Although we were unable to find SARS-CoV-2 antigen in sputum from our LC cases, we did not test for viral persistence in GI tract and lung tissue40,41 or in plasma52. Evidence of SARS-CoV-2 persistence would justify trials of antiviral drugs (singly or in combination) in LC. It is also possible that autoimmune processes could result in an innate inflammatory profile in LC. Autoreactive B cells have been identified in LC patients with higher SARS-CoV-2-specific antibody titers in a study of mostly mild acute COVID cases (59% WHO 2–3)42, a different population from our study of hospitalized cases.

Our observations of distinct protein profiles in GI and cognitive groups support previous reports on distinct associations between Epstein–Barr virus reactivation and neurological symptoms, or autoantibodies and GI symptoms relative to other forms of LC49,53. We did not assess autoantibody induction but found evidence of brain–gut axis disturbance (SCG3) in the GI group, which occurs in many autoimmune diseases54. We found signatures suggestive of neuroinflammation (C1QA) in the cognitive group, consistent with findings of brain abnormalities on magnetic resonance imaging after COVID-19 hospitalization55, as well as findings of microglial activation in mice after COVID-19 (ref. 56). Proinflammatory signatures dominated in the cardioresp, fatigue and anxiety/depression groups and were consistent with those seen in non-COVID depression, suggesting shared mechanisms57. The association between markers of myeloid inflammation, including IL-3RA, and symptoms was greatest for fatigue. Whilst membrane-bound IL-3RA facilitates IL-3 signaling upstream of myelopoesis36 its soluble form (measured in plasma) can bind IL-3 and can act as a decoy receptor, preventing monocyte maturation and enhancing immunopathology58. Monocytes from individuals with post-COVID fatigue are reported to have abnormal expression profiles (including reduced CXCR2), suggestive of altered maturation and migration5,59. Lung-specific inflammation was suggested by the association between PCDH1 (an airway epithelial adhesion molecule35) and cardioresp symptoms.

Our observations do not align with all published observations on LC. One proteomic study of 55 LC cases after generally mild (WHO 2–3) acute disease found that TNF and IFN signatures were elevated in LC3. Vasculoproliferative processes and metabolic disturbance have been reported in LC4,60, but these studies used uninfected healthy individuals for comparison and cannot distinguish between LC-specific phenomena and residual post-COVID inflammation. A study of 63 adults (LC, n = 50 and recovered, n = 13) reported no association between immune cell activation and LC 3 months after infection61, though myeloid inflammation was not directly measured, and 3 months post infection may be too early to detect subtle differences between LC and recovered cases due to residual acute inflammation.

Our study has limitations. We designed the study to identify inflammatory markers identifying pathways underlying LC subgroups rather than diagnostic biomarkers. The ORs we report are small, but associations were consistent across alternative methods of analysis and when using different LC definitions. Small effect sizes can be expected when using PLR, which shrinks correlated mediator coefficients to reflect combined effects and prevent colinear inflation62, and could also result from measurement of plasma mediators that may underestimate tissue inflammation. Although our LC cohort is large compared with most other published studies, some of our subgroups are small (only 60 cases were designated cognitive). Though the performance of the cognitive PLR model was adequate, our findings should be validated in larger studies. It should be noted that our cohort of hospitalized cases may not represent all types of LC, especially those occurring after mild infection. We looked for an effect of acute disease severity within our study and did not find it, and are reassured that the inflammatory profiles we observed were consistent with those seen in smaller studies including nonhospitalized cases42,46. Studies of posthospital LC may be confounded by ‘posthospital syndrome’, which encompasses general and nonspecific effects of hospitalization (particularly intensive care)63.

In conclusion, we found markers of myeloid inflammation and complement activation in our large prospective posthospital cohort of patients with LC, in addition to distinct inflammatory patterns in patients with cognitive impairment or gastrointestinal symptoms. These findings show the need to consider subphenotypes in managing patients with LC and support the use of antiviral or immunomodulatory agents in controlled therapeutic trials.


Study design and ethics

After hospitalization for COVID-19, adults who had no comorbidity resulting in a prognosis of less than 6 months were recruited to the PHOSP-COVID study (n = 719). Patients hospitalized between February 2020 and January 2021 were recruited. Both sexes were recruited and gender was self-reported (female, n = 257 and male, n = 462). Written informed consent was obtained from all patients. Ethical approvals for the PHOSP-COVID study were given by Leeds West Research Ethics Committee (20/YH/0225).

Symptom data and samples were prospectively collected from individuals approximately 6 months (IQR 5.1–6.8 months and range 3.0–8.3 months) post hospitalization (Fig. 1a), via the PHOSP-COVID multicenter United Kingdom study64. Data relating to patient demographics and acute admission were collected via the International Severe Acute Respiratory and Emerging Infection Consortium World Health Organization Clinical Characterisation Protocol United Kingdom (ISARIC4C study; IRAS260007/IRAS126600) (ref. 65). Adults hospitalized during the SARS-CoV-2 pandemic were systematically recruited into ISARIC4C. Written informed consent was obtained from all patients. Ethical approval was given by the South Central–Oxford C Research Ethics Committee in England (reference 13:/SC/0149), Scotland A Research Ethics Committee (20/SS/0028) and WHO Ethics Review Committee (RPC571 and RPC572l, 25 April 2013).

Data were collected to account for variables affecting symptom outcome, via hospital records and self-reporting. Acute disease severity was classified according to the WHO clinical progression score: WHO class 3–4: no oxygen therapy; class 5: oxygen therapy; class 6: noninvasive ventilation or high-flow nasal oxygen; and class 7–9: managed in critical care9. Clinical data were used to place patients into six categories: ‘recovered’, ‘GI’, ‘cardiorespiratory’, ‘fatigue’, ‘cognitive impairment’ and ‘anxiety/depression’ (Supplementary Table 7). Patient-reported symptoms and validated clinical scores were used when feasible, including Medical Research Council (MRC) breathlessness score, dyspnea-12 score, Functional Assessment of Chronic Illness Therapy (FACIT) score, Patient Health Questionnaire (PHQ)-9 and Generalized Anxiety Disorder (GAD)-7. Cognitive impairment was defined as a Montreal Cognitive Assessment score <26. GI symptoms were defined as answering ‘Yes’ to the presence of at least two of the listed symptoms. ‘Recovered’ was defined by self-reporting. Patients were placed in multiple groups if they experienced a combination of symptoms.

Matched nasal fluid and sputum samples were prospectively collected from a subgroup of convalescent patients approximately 6 months after hospitalization via the PHOSP-COVID study. Nasal and bronchoalveolar lavage fluid (BALF) collected from healthy volunteers before the COVID-19 pandemic were used as controls (Supplementary Table 4). Written consent was obtained for all individuals and ethical approvals were given by London–Harrow Research Ethics Committee (13/LO/1899) for the collection of nasal samples and the Health Research Authority London–Fulham Research Ethics Committee (IRAS project ID 154109; references 14/LO/1023, 10/H0711/94 and 11/LO/1826) for BALF samples.


Ethylenediaminetetraacetic acid plasma was collected from whole blood taken by venepuncture and frozen at −80 °C as previously described7,66. Nasal fluid was collected using a NasosorptionTM FX·I device (Hunt Developments), which uses a synthetic absorptive matrix to collect concentrated nasal fluid. Samples were eluted and stored as previously described67. Sputum samples were collected via passive expectoration and frozen at −80 °C without the addition of buffers. Sputum samples from convalescent individuals were compared with BALF from healthy SARS-CoV-2-naive controls, collected before the pandemic. BALF samples were used to act as a comparison for lower respiratory tract samples since passively expectorated sputum from healthy SARS-CoV-2-naive individuals was not available. BALF samples were obtained by instillation and recovery of up to 240 ml of normal saline via a fiberoptic bronchoscope. BALF was filtered through 100 µM strainers into sterile 50 ml Falcon tubes, then centrifuged for 10 min at 400 g at 4 °C. The resulting supernatant was transferred into sterile 50 ml Falcon tubes and frozen at −80 °C until use. The full methods for BALF collection and processing have been described previously68,69.


To determine inflammatory signatures that associated with symptom outcomes, plasma samples were analyzed on an Olink Explore 384 Inflammation panel70. Supplementary Table 8 (Appendix 1) lists all the analytes measured. To ensure the validity of results, samples were run in a single batch with the use of negative controls, plate controls in triplicate and repeated measurement of patient samples between plates in duplicate. Samples were randomized between plates according to site and sample collection date. Randomization between plates was blind to LC/recovered outcome. Data were first normalized to an internal extension control that was included in each sample well. Plates were standardized by normalizing to interplate controls, run in triplicate on each plate. Each plate contained a minimum of four patient samples, which were duplicates on another plate; these duplicate pairs allowed any plate to be linked to any other through the duplicates. Data were then intensity normalized across all cohort samples. Finally, Olink results underwent quality control processing and samples or analytes that did not reach quality control standards were excluded. Final normalized relative protein quantities were reported as log2 NPX values.

To further validate our findings, we performed conventional electrochemiluminescence (ECL) assays and enzyme-linked immunosorbent assay for Olink mediators that were associated with symptom outcome (Supplementary Methods). Contemporaneously collected plasma samples were available from 58 individuals. Like most omics platforms, Olink measures relative quantities, so perfect agreement with conventional assays that measure absolute concentrations is not expected.

Sputum samples were thawed before analysis and sputum plugs were extracted with the addition of 0.1% dithiothreitol creating a one in two sample dilution, as previously described71. SARS-CoV-2 S and N proteins were measured by ECL S-plex assay at a fixed dilution of one in two (Mesoscale Diagnostics), as per the manufacturers protocol72. Control BALF samples were thawed and measured on the same plate, neat. The S-plex assay is highly sensitive in detecting viral antigen in respiratory tract samples73.

Nasal cytokines were measured by ECL (mesoscale discovery) and Luminex bead multiplex assays (Biotechne). The full methods and list of analytes are detailed in Supplementary Methods.

Statistics and reproducibility

Clinical data was collected via the PHOSP REDCap database, to which access is available under reasonable request as per the data sharing statement in the manuscript. All analyses were performed within the Outbreak Data Analysis Platform (ODAP). All data and code can be accessed using information in the ‘Data sharing’ and ‘Code sharing’ statements at the end of the manuscript. No statistical method was used to predetermine sample size. Data distribution was assumed to be normal but this was not formally tested. Olink assays and immunoassays were randomized and investigators were blinded to outcomes.

To determine protein signatures that associated with each symptom outcome, a ridge PLR was used. PLR shrinks coefficients to account for combined effects within high-dimensional data, preventing false discovery while managing multicollinearity34. Thus, PLR was chosen a priori as the most appropriate model to assess associations between a large number of explanatory variables (that may work together to mediate effects) and symptom outcome34,62,70,74. In keeping with our aim to perform an unbiased exploration of inflammatory process, the model alpha was set to zero, facilitating regularization without complete penalization of any mediator. This enabled review of all possible mediators that might associate with LC62.

A 50 repeats tenfold nested cross-validation was used to select the optimal lambda for each model and assess its accuracy (Extended Data Fig. 1). The performance of the cognitive impairment model was influenced by the imbalance in size of the symptom group (n = 60) relative to recovered (n = 250). The model was weighted to account for this imbalance resulting in a sensitivity of 0.98, indicating its validity. We have expanded on the model performance and validation approaches in Supplementary Information.

Age, sex, acute disease severity and preexisting comorbidities were included as covariates in the PLR analysis (Supplementary Tables 1 and 3). Covariates were selected a priori using features reported to influence the risk of LC and inflammatory responses1,39,64,75. Ethnicity was not included since it has been shown not to predict symptom outcome in this cohort64. Individuals with missing data were excluded from the regression analysis. Each symptom group was compared with the ‘recovered’ group. The model coefficients of each covariate were converted into ORs for each outcome and visualized in a forest plot, after removing variables associated with regularized OR between 0.98 and 1.02 or in cases where most variables fell outside of this range, using mediators associated with the highest decile of coefficients either side of this range. This enabled exclusion of mediators with effect sizes that were unlikely to have clinical or mechanistic importance since the ridge PLR shrinks and orders coefficients according to their relative importance rather than making estimates with standard error. Thus, confidence intervals cannot be appropriately derived from PLR, and forest plot error bars were calculated using the median accuracy of the model generated by the nested cross-validation. To verify observations made through PLR analysis, we also performed an unadjusted PLR, an unadjusted logistic regression and a PLS analysis. Univariate analyses using Wilcoxon signed-rank test was also performed (Supplementary Table 8, Appendix 1). Analyses were performed in R version 4.2.0 using ‘data.table v1.14.2’, ‘EnvStats v2.7.0’ ‘tidyverse v1.3.2’, ‘lme4 v1.1-32’, ‘caret v6.0-93’, ‘glmnet v4.1-6’, ‘mdatools v0.14.0’, ‘ggpubbr v0.4.0’ and ‘ggplot2 v3.3.6’ packages.

To further investigate the relationship between proteins elevated in each symptom group, we performed a correlation network analysis using Spearman’s rank correlation coefficient and false discovery rate (FDR) thresholding. The mediators visualized in the PLR forest plots, which were associated with cardiorespiratory symptoms, fatigue, anxiety/depression GI symptoms and cognitive impairment were used, respectively. Analyses were performed in R version 4.2.0 using ‘bootnet v1.5.6 and ‘qgraph v1.9.8 packages.

To determine whether differences in protein levels between men and women related to hormonal differences, we divided each symptom group into premenopausal and postmenopausal groups using an age cutoff of 50 years old. Differences between sexes in each group were determined using the Wilcoxon signed-rank test. To understand whether antigen persistence contributed to inflammation in adults with LC, the median viral antigen concentration from sputum/BALF samples and cytokine concentrations from nasal samples were compared using the Wilcoxon signed-rank test. All tests were two-tailed and statistical significance was defined as a P value < 0.05 after adjustment for FDR (q-value of 0.05). Analyses were performed in R version 4.2.0 using ‘bootnet v1.5.6’ and ‘qgraph v1.9.8’ packages.

Extended Data Fig. 10 was made using Biorender, accessed at

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.