Oral and gut microbiota relate to symptom subphenotypes in long COVID, independent of viral persistence

Summary

Long COVID presents a significant public health challenge, complicating diagnosis and treatment. In a prospective study of 349 individuals with long COVID (March 2021–December 2023), latent class analysis identified three symptom subphenotypes: high constitutional symptom burden (21%), predominant smell/taste disturbances (17%), and minimal persisting symptoms (62%). While viral persistence in saliva and stool was limited, 16S rRNA gene sequencing revealed microbiota associations with symptomatology. Alpha diversity was lower in individuals with high symptom burden, and specific taxa correlated with nausea and smell/taste disturbances. Distinct oral and gut microbiota patterns emerged across symptom clusters, with microbiota profiles also linked to patient-reported outcomes, including employment and overall health impact. These findings suggest that bacterial dysbiosis may contribute to long COVID symptom variability and highlight the microbiome’s potential role in its pathophysiology. Understanding microbial influences on symptom persistence may inform microbiome-targeted therapeutic strategies and improve long COVID management.

Graphical abstract

Highlights

• •Long COVID symptoms grouped into three distinct clinical subphenotypes
• •Oral and gut microbiota diversity was lower in participants with higher symptom burden
• •Specific microbial taxa were linked to nausea, taste/smell loss, diarrhea, and fever
• •Microbiota patterns differed from healthy controls and predicted employment outcomes

Microbiome; Viral microbiology; Oral microbiology

Introduction

The global impact of post-acute sequelae of COVID-19, commonly referred to as long COVID, has emerged as a major public health challenge, affecting millions worldwide.^1,2,3 Long COVID encompasses a broad spectrum of new and persistent symptoms that can last for weeks or months after the acute infection phase, including fatigue, dyspnea, cognitive impairment, and chemosensory dysfunction.^4,5 This symptom heterogeneity complicates diagnosis and treatment, and different symptoms or long COVID phenotypes may result from different mechanisms.^6,7 Proposed contributors include persistent viral replication, immune dysregulation, mitochondrial dysfunction, thromboinflammation, and localized tissue inflammation.^7,8,9,10,11 However, the diversity of clinical presentations suggests that a more nuanced understanding is needed.

Recent studies have identified distinct long COVID subphenotypes.^5,12 Using latent class analysis (LCA) of self-reported symptoms, we described three subphenotypes: (1) a high burden of constitutional symptoms, (2) persistent loss or change of smell and taste, and (3) minimal residual symptoms.¹² Subphenotypes were strongly associated with self-assessments of overall health, recovery, and long COVID’s impact on employment. However, the biological drivers of this symptom heterogeneity remain unclear, underscoring the complexity of long COVID and the likelihood that multiple biological pathways are involved.

Emerging evidence points to the microbiome as a key factor in long COVID’s pathophysiology.^3,13,14 The human microbiome, which consists of diverse microbial communities, is known to influence immune function, inflammation, and overall health.¹⁵ Alterations in the gut and oral microbiota have been implicated in chronic conditions, and recent studies suggest they may also play a role in long COVID.^13,14,16,17 However, significant gaps remain in understanding how microbial dysbiosis might contribute to the variability in the persistence of specific long COVID subphenotypes. The interplay between SARS-CoV-2 infection, microbiota changes, and clinical presentations of long COVID warrants further investigation.

This prospective observational cohort study explores the relationships between viral persistence, oral and gut microbiota, and symptom subphenotypes in individuals with long COVID. While we found limited evidence of viral persistence in non-invasive saliva and stool samples, microbiota profiling via 16S rRNA gene sequencing revealed significant associations between microbial community features, symptom subphenotypes, and patient-reported outcomes. These findings suggest that bacterial dysbiosis may contribute to clinical features of long COVID, highlighting the potential of microbiome-targeted therapeutic strategies.

Results

Cohort description

From March 2021 to December 2023, we enrolled 349 individuals with long COVID in the prospective Post-COVID Impairment Phenotyping and Outcomes (Post-CIPO) study.¹² Participants were referred to the study through various sources when they reported symptoms consistent with, or sought care for, long COVID (Table 1; Figure S1). Of the 329 participants who completed the study questionnaires, 76 (22%) had been hospitalized during their acute SARS-CoV-2 infection, while 253 (78%) were non-hospitalized. Hospitalized individuals were older, had a higher BMI, and exhibited a greater burden of comorbidities, including diabetes, obstructive sleep apnea, and airway disease, compared with non-hospitalized participants (Table 1). The median enrollment time was 240 days post-COVID-19 diagnosis (interquartile range [IQR]: 152–413), with hospitalized participants enrolling closer to their acute illness than non-hospitalized participants (p < 0.0001, Table 1).

Table 1.

Clinical characteristics of the 349 individuals with long COVID, stratified by inpatient vs. outpatient status during acute COVID-19

Variable	All	Outpatientsa	Inpatientsa	p Value
Participants, n (%)	349	253 (72.5)	76 (21.8)	NA
Age (median, [IQR]), years	47.7 [34.1, 60.2]	43.2 [31.2, 54.8]	60.9 [49.4, 68.3]	<0.0001
Men, n (%)	92 (26.4)	56 (22.1)	28 (36.8)	0.01
White race, n (%)	308 (88.3)	225 (88.9)	67 (88.2)	0.78
Body mass index (BMI) (median [IQR])	29.4 [25.1, 34.4]	28.1 [24.4, 33.6]	31.4 [27.1, 37.1]	<0.0001
No college-level degree, n (%)	138 (41.9)	92 (36.4)	46 (60.5)	<0.0001
Hypertension, n (%)	110 (33.4)	76 (30.0)	34 (44.7)	0.02
Diabetes, n (%)	43 (13.1)	17 (6.7)	26 (34.2)	<0.0001
Obstructive sleep apnea, n (%)	77 (23.4)	44 (17.4)	33 (43.4)	<0.0001
Obstructive airways disease, n (%)	89 (25.5)	59 (23.3)	30 (39.5)	0.008
History of immunosuppression, n (%)	55 (16.7)	37 (14.6)	18 (23.7)	0.09
Ever smoker, n (%)	114 (34.7)	81 (32.0)	33 (43.4)	0.09
Vaccinated for Influenza, n (%)	217 (66.2)	170 (67.5)	47 (61.8)	0.44
Vaccinated for COVID-19, n (%)	229 (65.6)	184 (72.7)	46 (60.5)	0.06
No of COVID-19 vaccinations (median [IQR])	2.0 [0.0, 3.0]	2.0 [0.0, 3.0]	2.0 [0.0, 3.0]	0.01
Prevalent SARS-CoV-2 variant during each acute infection period	–	–	–	<0.0001
Wild type, n (%)	85 (24.4)	40 (15.8)	41 (53.9)	–
Alpha, n (%)	88 (25.2)	61 (24.1)	21 (27.6)	–
Delta, n (%)	58 (16.6)	47 (18.6)	8 (10.5)	–
Omicron, n (%)	118 (33.8)	105 (41.5)	6 (7.9)	–
Days since acute COVID-19 infection date (median [IQR])	240.0 [152.0, 413.0]	293.0 [171.0, 509.0]	167.5 [124.8, 241.5]	<0.0001

We present continuous variables as median and interquartile range (IQR) and categorical variables as number (%). We compared continuous variables with Wilcoxon tests and categorical variables with Fisher’s exact tests. We consider p < 0.05 as statistically significant..

^aData presented for the 329 participants with completed study questionnaires (253 outpatients and 76 inpatients).

Subphenotypes of self-reported symptoms

Among the 329 participants who completed the study questionnaires, a median of three symptoms (IQR: 1–5) persisted at a median of 240 days post COVID-19 (Figure 1A). Of the 16 assessed symptoms, fatigue, cognitive impairment, and shortness of breath were the most common (Figure 1B). LCA identified a 3-class model as the best fit (Table S1): cluster 1 (21%) had a high overall symptom burden, cluster 2 (17%) was characterized by predominant taste and smell disturbances, and cluster 3 represented a low-symptom burden group (62%) (Figures 1C and 1D). Cluster membership was not significantly associated with hospitalization during acute COVID-19 or with clinical and demographic factors typically linked to acute disease severity, such as age, sex, or comorbidities (Table 2). However, participants in cluster 3 had a higher proportion of COVID-19 vaccination (≥1 dose) and a greater cumulative number of vaccine doses compared to the other clusters, suggesting a potential association between COVID-19 vaccination and reduced long COVID symptom burden (Table 2).

Figure 1.

Distribution of symptoms and numerical scales by long COVID subphenotypes

(A) Distribution of present symptoms at the baseline visit across all subjects.

(B) Stacked bar showing the proportions of presence (“Yes” in orange) vs. absence (“No” in gray) for each of the 16 interviewed symptoms.

(C) Comparison of total symptom burden across clusters: cluster 1 (21%) had the highest number of active symptoms, followed by cluster 2 (17%), while cluster 3 (62%) had the fewest symptoms.

(D) Stacked bar plots illustrating symptom prevalence within each cluster, with symptom presence shown in cluster-specific colors and absence in gray.

(E) Cluster 1 subjects had much higher scores for the numerical scales generalized anxiety scale-7 (GAD7) for anxiety, patient health questionnaire-9 (PHQ9) for depression, and insomnia severity index (ISI) for insomnia, but no difference in the Montreal Cognitive Assessment/MoCA-BLIND score for neurocognitive functioning.

(F) Self-reported outcomes by clusters: comparisons of symptom counts and numerical scale scores between clusters were conducted using a global Kruskal-Wallis test, followed by pairwise Wilcoxon tests adjusted for multiple comparisons using the Bonferroni correction.

Only Bonferroni-adjusted p values are displayed. Data in boxplots (A and C) are presented as medians with interquartile range. We conducted LCA using the 16 active symptoms reported at the baseline visit as input variables.

Table 2.

Clinical characteristics of Long COVID subphenotypes

Variable	Cluster 1	Cluster 2	Cluster 3	p Value
Participants	70	55	204	–
Age (median, [IQR]), years	47.7 [34.1, 60.2]	43.2 [31.2, 54.8]	60.9 [49.4, 68.3]	<0.0001
Men, n (%)	17 (24.3)	13 (23.6)	54 (26.5)	0.88
White race, n (%)	66 (94.3)	52 (94.5)	174 (85.3)	0.41
Body mass index (BMI) (median [IQR])	29.4 [25.1, 34.4]	28.1 [24.4, 33.6]	31.4 [27.1, 37.1]	<0.0001
No college-level degree, n (%)	32 (45.7)	25 (45.5)	81 (39.7)	0.58
Hypertension, n (%)	30 (42.9)	16 (29.1)	64 (31.4)	0.16
Diabetes, n (%)	13 (18.6)	6 (10.9)	24 (11.8)	0.30
Obstructive sleep apnea, n (%)	21 (30.0)	15 (27.3)	41 (20.1)	0.18
Obstructive airways disease, n (%)	26 (37.1)	15 (27.3)	48 (23.5)	0.09
History of immunosuppression, n (%)	18 (25.7)	7 (12.7)	30 (14.7)	0.07
Ever smoker, n (%)	26 (37.1)	24 (43.6)	64 (31.4)	0.21
Hospitalized for COVID-19, n (%)	11 (15.7)	14 (25.5)	51 (25.0)	0.25
Vaccinated for Influenza, n (%)	40 (57.1)	38 (70.4)	139 (68.1)	0.19
Vaccinated for COVID-19, n (%)	42 (60.0)	34 (61.8)	154 (75.5)	0.02
No of COVID-19 vaccinations (median [IQR])	1.0 [0.0, 3.0]	2.0 [0.0, 3.0]	2.0 [0.0, 3.0]	0.008
Prevalent SARS-CoV-2 variant during each acute infection period	–	–	–	0.004
Wild type, n (%)	18 (25.7)	16 (29.1)	47 (23.0)	–
Alpha, n (%)	15 (21.4)	17 (30.9)	50 (24.5)	–
Delta, n (%)	15 (21.4)	15 (27.3)	25 (12.3)	–
Omicron, n (%)	22 (31.4)	7 (12.7)	82 (40.2)	–
Days since acute COVID-19 infection date (median [IQR])	249.5 [146.2, 422.0]	266.0 [172.0, 486.5]	246.5 [154.8, 395.0]	0.78

We present continuous variables as median and interquartile range (IQR) and categorical variables as number (%). We compared continuous variables with Kruskal-Wallis tests and categorical variables with Fisher’s exact tests. We consider p < 0.05 as statistically significant.

Cluster 1 participants reported higher anxiety, depression, and insomnia scores, though no difference was observed on the Montreal Cognitive Assessment (MoCA)-BLIND neurocognitive scale (Figure 1E), despite 86% self-reporting cognitive impairment. Cluster membership was significantly associated with self-assessed health, long COVID recovery, and employment impact, with cluster 1 reporting the poorest outcomes—70% reported a negative impact on employment (Figure 1F).

Viral persistence in long COVID

We measured SARS-CoV-2 RNA (vRNA) in 143 saliva and 141 stool samples, detecting vRNA in 8 saliva samples (5.5%) and 1 stool sample (1.0%). Given this low prevalence, further statistical analyses were not performed. These findings suggest that viral persistence in saliva or stool samples, collected at a median of 240 days post-acute COVID-19, is rare in individuals with long COVID. Consequently, the symptom burden, long COVID subphenotypes, and outcomes do not appear to be driven by detectable viral persistence in non-invasive, easily accessible biospecimens.

Saliva microbiota and long COVID

We profiled saliva bacterial communities using 16S rRNA gene sequencing of the V4 region in samples from 239 participants. Analyses were performed at the genus level, assessing alpha diversity (Shannon index), additive log-ratio (ALR)-transformed abundances of the top 25 taxa, and beta diversity for compositional differences. We investigated relationships between saliva microbiota, long COVID clusters, individual symptoms, and self-reported outcomes. To further contextualize our findings in long COVID patients, we compared their saliva-community profiles with available salivary microbiota data from 37 healthy controls, who had been enrolled in the MEDBIO cohort at the University of Pittsburgh and had their 16S rRNA gene sequencing data available in our database. Healthy controls were defined as participants in the MEDBIO cohort who had been enrolled as non-diseased participants and with an Elixhauser score of 0, i.e., absence of chronic comorbid conditions, such as diabetes, chronic obstructive pulmonary disease, or hypertension.

Core oral microbiome members (e.g., Streptococcus, Prevotella, Veillonella, and Rothia) were highly abundant across all clusters (Figure 2A). Alpha diversity, measured by the Shannon index, was significantly lower in cluster 1 (the globally symptomatic group) compared with cluster 2 (p = 0.05, Figure 2B). Cluster membership was the strongest predictor of beta diversity (Manhattan distances, PERMANOVA R² = 0.18, p = 0.02), even after adjusting for clinical variables, including sex, age, vaccination status, and time since acute COVID-19 (Figure 2C). Using ALR-transformed abundances and multivariate models, we identified significant associations between taxa and specific symptoms, with the strongest links for nausea and taste/smell disturbances (8 taxa each) and “runny nose” (6 taxa) (Figure S2). Including microbiota data in regression models improved symptom prediction for throat pain, nausea, “runny nose,” headache, aches, and taste/smell disturbances—five of which localize to the head/neck region near the site of sampling (Figure S3).

Figure 2.

Saliva microbiota analyses in long COVID

(A) Bacterial community composition stratified by long COVID clusters showed that core oral microbiota members (Streptococcus, Prevotella, Veillonella, and Rothia) were the most abundant across all clusters.

(B) Alpha diversity differed significantly by cluster, with the globally symptomatic cluster 1 displaying the lowest Shannon index (p = 0.05 vs. cluster 2).

(C) Beta diversity was assessed using a permutational analysis of variance (PERMANOVA) model, adjusting for sex, age, vaccination status, days since acute COVID-19, and other covariates. Cluster membership was the strongest factor explaining between-sample differences (R² = 0.18, p = 0.02), exceeding the impact of clinical covariates.

Based on distance-based analyses, salivary microbiota from long COVID participants showed significant differences compared to healthy controls (R² = 0.062, p = 0.0001, Figure S4). Long COVID participants demonstrated enrichment for Streptococcus, Gemella, Actinomyces and Staphylococcus taxa compared to healthy controls based on ALR-transformed abundance comparisons (Figure S4). Taken together, our salivary microbiota analyses demonstrate that long COVID participants had significantly different microbial communities compared to healthy controls, and that these microbiota differences are most strongly associated with anatomically proximate symptoms in the head and neck region, suggesting a potential role for oral dysbiosis in the pathophysiology of upper respiratory and neurological manifestations of long COVID.

Stool microbiota and long COVID

Microbiota analysis of 238 stool samples from long COVID participants identified Bacteroides, Blautia, Faecalibacterium, and Lachnospiraceae as the most abundant taxa across all long COVID clusters (Figure 3A). Alpha diversity (Shannon index) was significantly lower in cluster 1 (high symptom burden) compared to cluster 3 (p = 0.03, Figure 3B). In PERMANOVA models, beta diversity was not significantly associated with long COVID clusters but was influenced by diabetes history and White race. Symptom-based analysis of ALR-transformed abundances revealed associations between specific taxa and symptoms, with the strongest links to diarrhea and fever (11/25 taxa) and taste/smell disturbances (7 taxa) (Figure S5). Given the low fever prevalence (1%), further analyses focused on diarrhea (13%). After adjusting for clinical covariates, Ruminococcaceae abundance was associated with reduced diarrhea risk, while Bifidobacterium, Lachnospiraceae, and Akkermansia were linked to increased risk (Figure S6).

Figure 3.

Associations between stool microbiota and long COVID clusters

(A) Stratification by long COVID clusters revealed that common gut taxa, including Bacteroides, Blautia, Faecalibacterium, and Lachnospiraceae, were most abundant across all clusters.

(B) Alpha diversity analysis showed that cluster 1, characterized by globally symptomatic individuals, had a significantly lower Shannon diversity index compared to the less symptomatic cluster 3 (p = 0.03).

To examine broader microbiota-symptom relationships, hierarchical cluster analysis with multinomial logistic regression (HCAMLR) defined microbiota-derived clusters distinct from long COVID symptom clusters. A two-cluster HCAMLR model emerged (Figure 4): HCAMLR-1 (11%) was Prevotella-enriched and associated with greater dyspnea severity (p < 0.01, Modified Medical Research Council [MMRC] scale p < 0.05), while HCAMLR-2 (89%) was Bacteroides-dominant and linked to increased cognitive impairment risk (p < 0.05, Figure 4).

Figure 4.

Hierarchical microbiota clusters and associations with long COVID symptoms

Using an unsupervised approach, we performed hierarchical clustering based on taxa abundance and applied multinomial logistic regression to identify relationships between microbiota-derived clusters and individual symptoms.

(A) Non-metric multidimensional scaling (NMDS) plot illustrating the two-cluster model.

(B) Cluster 1 (11% of the cohort) exhibited a higher abundance of Prevotella taxa, while cluster 2 (89%) was characterized by Bacteroides predominance. Participants in the Prevotella-driven cluster had a significantly higher risk of reporting dyspnea (p < 0.01), whereas individuals in the Bacteroides-driven cluster were more likely to report cognitive impairment (p < 0.05).

Stool microbiota also correlated with patient-reported outcomes, including employment status and symptom severity. In a PERMANOVA model adjusting for clinical factors, beta diversity was significantly associated with negative employment outcomes (Figure 5A). Among individual taxa, 14 of 25 were linked to employment impact, and 6 were associated with MMRC dyspnea scores (Figure S7). Including ALR-transformed taxa improved predictive models for employment, GAD-7, and ISI, whereas clinical variables better predicted overall health and MoCA scores (Figure 5B). In multivariate models, protective factors against employment impact included total vaccine doses and higher Ruminococcus and Veillonella abundances, while Lachnospiraceae was associated with negative employment outcomes (Figure 5C). Unsupervised clustering linked HCAMLR-2 to poorer self-reported health.

Figure 5.

Associations between stool microbiota and self-reported outcomes and clinical scales in long COVID

(A) In a PERMANOVA model adjusting for clinical covariates, beta diversity was significantly associated with negative impact on employment. The model controlled for the following variables: long COVID prior to vaccination, White race, smoking history, autoimmune disease history, education level, total vaccine doses, MoCA score, days post-COVID, age, chronic obstructive pulmonary diseaase (COPD), BMI, GAD-7, PHQ-9, ISI scale, OSA, and hospitalization status during acute COVID.

(B) Comparison of full versus reduced models (with and without taxa abundances) demonstrated that incorporating microbiota data improved predictive accuracy for employment, the generalized anxiety disorder scale (GAD-7), and the insomnia severity index (ISI), while clinical variables alone provided better predictions for overall health and MoCA scores, as indicated by R² ratios.

(C) Multivariate regression models adjusting for clinical factors indicated that more COVID-19 vaccination doses, as well as greater abundances of Ruminococcus and Veillonella, were associated with fewer negative employment impacts, while Lachnospiraceae abundance was linked to a detrimental effect.

Finally, to contextualize stool microbiota findings in long COVID participants, we compared them with available stool microbiota data from 141 healthy control participants from the MEDBIO cohort, similar to our selection of healthy controls for the salivary microbiota comparisons. Distance-based analyses revealed significant differences in beta diversity between long COVID and healthy controls (Figure S8), while abundance comparisons showed enrichment of Blautia and Lachnospiraceae taxa in long COVID participants, with depletion of Barnesiella, Akkermansia, and Parasutterella (Figure S8). Taken together, stool microbiota analyses demonstrate that long COVID participants exhibit distinct gut microbial communities compared to healthy controls, with microbiota-derived clustering patterns that predict specific symptom domains (respiratory and cognitive) and functional outcomes.

Discussion

Our study provides valuable insights into the clinical heterogeneity of long COVID and the potential role of the microbiome in its pathophysiology. Through LCA, we identified three distinct symptom-based long COVID subphenotypes, varying in symptom burden and composition. The globally symptomatic group (cluster 1) reported the poorest outcomes across all measures, underscoring the significant impact of long COVID on quality of life and functional status. These findings align with prior subphenotyping efforts^2,5 and are consistent with a previous analysis of our own,¹² further validating observed symptom patterns.

Although we found minimal detectable SARS-CoV-2 RNA in saliva or stool samples at a median of 240 days post-acute infection, this does not fully exclude viral persistence, which may occur at low titers or inaccessible tissue reservoirs.¹¹ More importantly, we observed significant associations between microbiome features and long COVID subphenotypes, suggesting a role for microbial dysbiosis in symptom persistence. Saliva- and stool-microbiota analyses revealed lower alpha diversity in cluster 1 and associations between specific taxa and symptom domains. While these findings are consistent with emerging literature linking microbial dysbiosis to chronic conditions, including post-viral syndromes,^{3,13,14,18,19} they should be interpreted with caution. A substantial proportion of our participants had histories of smoking or immunosuppression, both of which are known to influence microbial communities and may confound observed associations. Additionally, unmeasured factors such as diet, medication use, and social determinants of health—including socioeconomic status, healthcare access, nutrition, and living conditions—likely influenced both symptom expression and microbiota profiles. Specific long COVID symptoms, particularly taste and smell disturbances, could also affect dietary intake and feeding behavior, further complicating the causal interpretation. While our analyses adjusted for select clinical variables, residual confounding from these unmeasured factors cannot be excluded. Thus, although our results suggest that microbiome alterations may contribute to symptom persistence in long COVID, we cannot exclude the possibility that underlying host factors or preexisting conditions also played a role.

Notably, we also observed associations between stool microbiota and patient-reported outcomes, including employment impacts and symptom severity scales. These associations underscore potential links between the gut microbiome and functional status in Long COVID.

Given the limitations and cross-sectional design of our study, these associations should be viewed as hypothesis generating. Longitudinal studies with comprehensive clinical, dietary, and social data will be essential to determine whether microbiome alterations contribute to long COVID pathophysiology or reflect broader host and environmental influences, though further work is needed to clarify causality and directionality.

To further contextualize these microbiota alterations, we compared salivary and stool microbial profiles in long COVID participants with healthy control cohorts without chronic comorbidities and collected using similar protocols. Long COVID participants exhibited significant compositional differences in both oral and gut microbiota compared to healthy controls, with enrichment of potentially pro-inflammatory or opportunistic taxa (e.g., Streptococcus, Gemella, Lachnospiraceae) and depletion of taxa associated with gut health (e.g., Barnesiella, Akkermansia). While differences in timing of sample collection and pandemic-era exposures may have contributed to these findings, the consistent signal of dysbiosis across compartments supports the hypothesis that long COVID involves persistent microbiome alterations that may contribute to the ongoing symptoms.

Recent studies, including a randomized controlled trial demonstrating symptom improvement with synbiotic supplementation (SIM01), suggest that microbiome-targeted therapies may be a promising avenue for long COVID management.^16,20,21 Although our study was not designed to evaluate interventions and cannot directly support therapeutic claims, our findings may help inform future trial designs by identifying microbiome features and symptom clusters that could be targeted in microbiome-modulating interventions. Future clinical trials could benefit from stratifying participants by symptom clusters—targeting gut microbiota for dyspnea and cognitive impairment, while focusing on oral microbiota in patients with head and neck-predominant symptoms, such as those in cluster 2 with significant taste and smell disturbances.

Limitations of the study

Despite these promising findings, our study has limitations. While it represents one of the largest cohorts combining saliva- and stool-microbiota analysis in long COVID, partitioning into symptom-based clusters reduced sample sizes for individual subgroup analyses, necessitating cautious interpretation of some associations. We did not collect information on dietary or drinking behaviors, which limits our ability to evaluate the role of baseline nutrition in shaping microbiota differences or influencing long COVID symptoms. Moreover, we did not capture whether long COVID-related changes in taste, smell, gastrointestinal function, or energy levels prompted alterations in diet or beverage consumption, which could also have affected microbiota composition and recovery trajectories. In addition to potential confounding from unmeasured variables such as diet and antibiotic use, a sizable proportion of participants reported prior immunosuppression or smoking, both of which are known to influence the microbiome and could confound associations with long COVID symptoms. Additionally, while healthy controls were carefully selected for the absence of chronic comorbidities at the time of sampling and underwent microbiota profiling using identical protocols, we did not have data on persistent symptoms or functional outcomes (e.g., employment status) in this group, limiting our ability to draw direct inferences about microbiota-symptom or microbiota-outcome associations. Moreover, although most control samples were obtained prior to the COVID-19 pandemic and the remainder before any reported infection, we cannot fully exclude heterogeneity in prior SARS-CoV-2 exposure within the control group. The cross-sectional design further precludes causal inferences. A further challenge, common to long COVID studies, is that the condition arises in the context of evolving SARS-CoV-2 variants, changing acute-phase treatments, widespread vaccination, and increasing herd immunity. As a result, some associations may be influenced by the timing of participant enrollment during different phases of the pandemic. Longitudinal studies are needed to determine whether microbiome alterations drive long COVID pathophysiology or arise as a consequence.

Future research should focus on characterizing microbial changes across long COVID subphenotypes and exploring targeted interventions to modulate the microbiome and alleviate symptoms. Larger, multicenter studies are essential to validate microbiome-targeted interventions and assess their long-term benefits across diverse populations. While mechanisms linking microbiome alterations to long COVID remain to be fully elucidated, our study provides a foundation for further investigation. Advancing our understanding of these associations and developing microbiome-focused treatments will be critical to addressing the ongoing burden of long COVID.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Dr. Georgios D. Kitsios (kitsiosg@upmc.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

The microbiome sequencing data supporting this study have been deposited in the NCBI Sequence Read Archive (SRA) under BioProject accession PRJNA1187611. This accession should be used in citations to facilitate searching via Entrez. The data will be publicly available upon release on December 1, 2026, or the publication of this article, whichever comes first.

Acknowledgments

The investigators would like to thank Dr. Michael Flock, PhD, University of Pittsburgh, who provided invaluable assistance in project coordination and grant contracting. We also thank all study participants for their contribution to this study.

This work was supported by a research grant from Pfizer Inc to the University of Pittsburgh.

Author contributions

G.D.K. wrote the first draft of the manuscript and led the data analysis. K.L. performed the microbiome analysis. S.B., B.Z., H.G., and C.M. were involved in cohort assembly and database management. X.W. and A.P. conducted microbiome experiments. A.F., J.J., A.N., and J.M. led the viral persistence analysis. L.P., A.R., and F.D. contributed expertise in statistics and epidemiology. B.M. oversaw microbiome experiments and analyses. S.M.N. led latent class analysis (LCA) and other statistical modeling. A.M. supervised the study and secured funding. F.S. and J.M. provided clinical oversight. All authors contributed to data interpretation, manuscript revision, and approved the final version for submission.

Declaration of interests

G.D.K. has received research funding from Genentech, Inc, has received consultant fees from InflaRx, Inc, and serves on the Advisory Board for KeepBio, Inc. G.D.K., A.M., J.M., F.S., and M.N. have received funding from Pfizer, Inc. J.M. is a consultant to Gilead Sciences and owns shares or share options in Co-Crystal Pharmaceuticals, ID Connect, and Abound Bio, all unrelated to the current work.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, peptides, and recombinant proteins
Saliva and stool samples	University of Pittsburgh, Division of Pulmonary, Allergy and Critical Care Medicine	NA
ZymoBIOMICS Microbial Community DNA Standard	Zymo Research	D6305
Ambion Proteinase K	ThermoFisher Scientific	AM2546
Critical commercial assays
Dneasy Powersoil DNA Kit	QIAGEN	#12888-100
MiSeq Reagent Kit v2 (300-cycles)	Illumina	MS-102-2002
MagMAX Viral Pathogens Kit	ThermoFisher Scientific	A42352
TaqPath™ 1-Step RT-qPCR Master Mix	ThermoFisher Scientific	A15300
TaqMan 2019-nCoV Assay Kit v1 (RNaseP target)	ThermoFisher Scientific	A47532
KingFisher Flex Automated Purification System	ThermoFisher Scientific	5400630
SARS-CoV-2 RNA Control 2 (Wuhan-1, MN908947.3)	Twist Biosciences	SKU: 102024
Heat-inactivated SARS-CoV-2 (Isolate USA-WA1/2020)	BEI Resources	NR-52350
Deposited data
16S rRNA gene sequences	Sequence Resource Archive	BioProject accession PRJNA1187611 for Long COVID participants; and PRJNA1187611 and PRJNA847970 for healthy controls
Software and algorithms
R version v4.4.3	The R Foundation for Statistical Computing	Not applicable
Prism 10 for macOS	GraphPad	Not applicable
Mothur	Mothur.org	Not applicable
Other
Primary Code	GitHub - Zenodo	https://doi.org/10.5281/zenodo.16545341

Experimental model and study participant details

Study cohort

We conducted the Post-COVID Impairment Phenotyping and Outcomes [Post-CIPO] study, a prospective, observational cohort study with longitudinal follow-up of adults (≥18 years old) experiencing Long COVID symptoms. The study was approved by the University of Pittsburgh IRB (STUDY21010001). We used an inclusive case definition of Long COVID, defined as the presence of any new or persistent symptoms for at least 20 days following a documented SARS-CoV-2 infection by positive quantitative real-time-PCR (qPCR). To ensure a diverse cohort, we enrolled 349 participants from different sources, including individuals previously enrolled in acute COVID-19 studies, patients self-referred to this Long COVID investigation, referrals from physicians, and individuals identified as eligible during clinical encounters in a Post-COVID clinic at UPMC. This broad enrollment strategy aimed to capture a wide range of patients with varying risks of Long COVID subtypes and severity. Participants completed structured telephone interviews to collect demographic data (including sex and race), comorbid conditions, details on time since their most recent COVID-19 diagnosis, prior COVID-19 infections, vaccination history, treatments received during the acute COVID-19 phase. We did not capture self-identified gender information. We then recorded the presence, duration, and severity of 16 Long COVID symptoms. These symptoms included fever, chills, muscle aches, “runny nose”, sore throat, cough, dyspnea (“shortness of breath”), nausea or vomiting, headache, abdominal pain, diarrhea, loss/change of smell, loss/change of taste, cognitive impairment (“brain fog”), fatigue, and chest issues (pain or palpitations). Symptoms were selected based on expert consensus and existing knowledge of Long COVID symptomatology at the time of cohort initiation.

Ethics approval and consent to participate

The University of Pittsburgh Institutional Review Board (IRB) approved the study protocol STUDY21010001. We obtained written or electronic informed consent by all participants in accordance with the Declaration of Helsinki. Participants for the MEDBIO cohort had been enrolled under study protocol CR19030104-017.

Consent for publication

We obtained the necessary participant consent and the appropriate institutional forms have been archived. Any patient/participant/sample identifiers included were not known to anyone outside the research group so cannot be used to identify individuals.

Method details

Study cohort

Standardized assessments using validated clinical numerical scales were conducted to capture the following domains: i) psychological symptoms: Generalized Anxiety Scale-7 [GAD7] for anxiety²²; patient health questionnaire-9 [PHQ9] for depression²³; insomnia severity index [ISI] for insomnia,²⁴ ii) neurocognitive functioning: Montreal Cognitive Assessment / MoCA-BLIND,²⁵ and iii) cardiopulmonary function: Modified Medical Research Council [MMRC] Dyspnea scale.²⁶ Participants also provided self-assessments of their overall Long COVID outcomes in terms of global health, recovery, and impact on employment. Data from the initial survey at the time of enrollment in the Long COVID portion of the study were included in this analysis. Detailed questions from the structured interviews have been previously described.¹²

To contextualize our microbiota analysis findings in Long COVID, we performed comparisons with healthy control salivary and stool microbiota data that were available in our database. For that purpose, we selected healthy control participants from the MEDBIO cohort, a prospective observational study at the University of Pittsburgh. Inclusion criteria for healthy controls required enrollment as non-diseased participants (i.e., no index diagnosis such as COPD, idiopathic pulmonary fibrosis, or inflammatory bowel disease) and an Elixhauser comorbidity index of 0, indicating no recorded chronic comorbidities. Additionally, participants were required to have salivary or stool microbiota data generated using the same 16S rRNA gene sequencing protocols as used for the Long COVID cohort. Of the included control samples, approximately 80% were collected prior to 2020, before the emergence of COVID-19, while the remaining ∼20% were collected after the onset of the pandemic but prior to any known or reported SARS-CoV-2 infection. Thus, these controls represent individuals without known exposure to SARS-CoV-2, providing a pre-pandemic and uninfected comparison group for our analyses.

Biospecimen acquisition and molecular analyses

Following the initial study visit, Long COVID participants self-collected stool and saliva samples at a single time point, as follows:

• a.Stool specimens were self-collected using the DNA/RNA Shield Fecal Collection tubes (Zymo) for nucleic acid preservation and short-term (two to four weeks) storage at ambient temperature.
• b.Saliva specimens were self-collected using the OMNIgene·ORAL OM-505 devices. 2 mL of saliva were collected for nucleic acid preservation and short-term storage at ambient temperature.

Specimens were mailed to the University of Pittsburgh Center for Medicine and the Microbiome. Upon receipt, specimens were sub-aliquoted prior to long-term storage at -80°C.

SARS-CoV-2 viral load quantification: We performed qRT-PCR for SARS-CoV-2 as follows. Proportional amounts of a qualitative internal extraction control (replication-competent avian leukosis virus (ALV) long terminal repeat (LTR) with a splice adaptor (RCAS) were added to 0.35 mL of inactivated saliva and stool samples (when available), and total RNA extracted using the MagMax Viral Pathogens Kit (kit-supplied Proteinase K replaced with 200μg Ambion ProteinaseK) and the Kingfisher Flex automated extractor (Thermofisher). One step quantitative RT-PCR was performed on RNA extracts using primer/probe sets to detect RCAS SARS-CoV-2 N, and human RNaseP (Thermofisher TaqPath™ 1-Step RT-qPCR Master Mix, Cat#A15300; RnaseP target from TaqMan 2019nCoV Assay Kit v1, Cat#A47532; N target (Fwd: 5’-GTTTGGTGGACCCTCAGATT-3’, Rev: 5’-CGCAGTATTATTGGGTAAACCTTG-3’, Probe: 5’6-FAM-TAACCAGAATGGAGAACGCAGTGGG-3’BHQ1). RNA was quantified using an in-run standard curve constructed from full genome SARS-CoV-2 RNA (Wuhan-1, Twist Biosciences, Twist Synthetic SARS-CoV-2 RNA Control 2 (MN908947.3) - SKU: 102024) that was concentration-verified by endpoint dilution. Each run contained a positive control (100 copies of heat inactivated SARS-CoV-2 (Isolate USA-WA1/2020, BEI Resources Catalog No. NR-52350) in healthy human pre-COVID specimen), and negative control (healthy human pre-COVID specimen).^27,28

16S rRNA gene sequencing

From separate aliquots of stool and saliva samples, we extracted genomic DNA and performed Illumina sequencing of the V4 hypervariable region of the bacterial 16S rRNA gene as previously described.^29,30,31 Genomic DNA (gDNA) was extracted and amplified for the V4 region using Q5 HS High-Fidelity polymerase (New England BioLabs, Ipswich, MA). Inline barcode primers were designed based on the method described by Caporaso et al. (2012). The V4 primer sequences used were 515f (5’-GTGCCAGCMGCCGCGGTAA-3’) and 806r (5’-GGACTACHVGGGTWTCTAAT-3’). Approximately 5-10 ng of each sample was amplified in 25 μL reactions. The cycling conditions consisted of an initial denaturation at 98°C for 30 seconds, followed by 30 cycles of 98°C for 10 seconds, 57°C for 30 seconds, and 72°C for 30 seconds, culminating in a final extension step at 72°C for 2 minutes. Amplicons were purified using AMPure XP beads (Beckman Coulter, Indianapolis, IN) at a 0.8:1 (beads:DNA) ratio to eliminate primer dimers. The eluted DNA was quantified using a Qubit fluorimeter (Life Technologies, Grand Island, NY). Sample pooling was conducted on ice by combining 40 ng of each purified band. For negative controls and samples with low performance, 20 μL from each was utilized. The sample pool was then purified with the MinElute PCR purification kit (Qiagen, Germantown, MD). Two additional purifications were performed: the first using AMPure XP beads at a 0.8:1 ratio to further remove primer dimers, followed by a final cleanup with the PureLink PCR Purification Kit (Life Technologies). The purified pool was quantified in triplicate using the Qubit fluorimeter prior to sequencing. The sequencing pool was prepared according to Illumina’s recommendations (Illumina, Inc., San Diego, CA), with an additional incubation at 95°C for 2 minutes immediately after the initial dilution to 20 pM. The pool was then adjusted to a final concentration of 7 pM, supplemented with 20% PhiX control (Illumina). Sequencing was conducted using an Illumina MiSeq 500-cycle V2 kit.

Quantification and statistical analysis

Numerical data are presented as medians with interquartile ranges. To identify distinct Long COVID symptom-based subphenotypes (classes), we conducted unsupervised classification with Latent Class Analysis (LCA) of self-reported symptoms at the initial assessment. We examined model fit by calculating membership probability, entropy, and the parametric bootstrapped log likelihood ratio between different classes, as well as a clinical relevance criterion of ensuring that each class has at least 5% of observations from the cohort. Results from the quantitative scales (GAD7, PHQ9, ISI, MoCA-BLIND, MMRC) were then mapped to each cluster for a descriptive assessment of those data by cluster assignment. We performed non-parametric comparisons for continuous (described as median and interquartile range – IQR) and categorical variables between different groups and reported the nominal p-values for all tests performed. We conducted analyses in R v4.2.0 (statistical and microbiota analyses) and Mplus 8.8 (LCA).

Sequences obtained from the Illumina MiSeq were deconvoluted and processed through the Center for Medicine and the Microbiome (CMM) in-house sequence quality control pipeline. This pipeline includes low-complexity filtering, trimming of quality values (QV<30), and the removal of primers used for 16S rRNA gene amplification, along with minimum read length filtering. Using scripts from Hannon’s Cold Spring Harbor Laboratory FASTA-Toolkit (http://hannonlab.cshl.edu/fastx_toolkit/), reads were trimmed until the quality value reached 30 or higher. Trimmed reads shorter than 75 bp or with less than 95% of bases above a QV of 30 were discarded. Forward and reverse paired reads were merged with a minimum required overlap of 25 bp, allowing a proportion overlap mismatch of > 0.2, a maximum of 4 N’s, and a minimum read length of 125 bp. The merged reads were processed through the CMM’s Mothur-based (v1.44.1) pipeline³² for 16S rRNA gene sequence clustering and annotation. Taxonomic classification of sequences was performed using the Ribosomal Database Project’s (RDP) naïve Bayesian classifier, referencing the SILVA 16S rRNA database (v138).³³

Analyses of the 16S rRNA gene sequence taxonomic profiles were conducted using three distinct approaches: abundance-based, distance-based, and distribution-based methods.

• 1.Abundance-based Methods: We applied the additive-log ratio (ALR) transformation³⁴ to relative abundances to mitigate spurious correlations among taxa in compositional data. This transformation allowed taxa to be analyzed independently and treated as normally distributed.
• 2.Distance-based Methods: Beta diversity, or inter-sample differences, was assessed using the Manhattan distance metric, which quantifies dissimilarities in microbiota composition between samples.
• 3.Distribution-based Methods: Within-sample diversity (alpha diversity) was evaluated using multiple diversity indices, including the Tail statistic and Shannon index, to capture richness and evenness within individual samples.³⁵

Two separate regression analyses were performed to examine associations between microbiota profiles and target variables, which included LCA cluster classifications and symptom variables (all 16 PASC-related symptoms). These variables were derived from the initial study visit data (self-reported or clinician-assigned). For each analysis, microbiota profiles (quantified using ALR-transformed abundances, inter-sample distances, or alpha diversity indices) were modeled both as predictors and responders to the target variables. Covariates, which are unlikely to be influenced by the microbiome, were included in both models to control for potential confounding. These covariates included: BMI, age, COVID-19 vaccination status, COVID-19 variant, smoking status, use of anti-viral treatments, sex, and race.

Predictor/Response Analysis

A Predictor/Response Analysis³⁶ was conducted to assess the relative strength of the microbiota as a predictor versus a response to the target variables. This analysis compared the fit of the two regression models by examining the microbiota's contribution to model outcomes. When microbiota was modeled as a predictor, both full models (including microbiota and covariates) and reduced models (covariates only) were compared using ANOVA and R², evaluating the microbiota’s contribution to the overall model fit.

Hierarchical Cluster Analysis with multinomial logistic regression (HCAMLR)

To identify targeted variables associated with specific microbiota clusters, we used Hierarchical Cluster Analysis with Multinomial Logistic Regression (HCAMLR).³⁷ HCAMLR first employed hierarchical clustering with the Manhattan distance and Ward’s linkage criterion to define clusters. The clusters were then associated with the target variables through multinomial logistic regression at each level of the tree. To identify the taxa driving cluster formation, we calculated the log(R² ratio). The R² statistic measures the separation between two groups based on taxa composition. Full R² is the value obtained when all taxa are included in the distance calculation, while reduced R² is calculated by excluding a specific taxon. Taxa that contributed to cluster separation were identified when the ratio of reduced R² to full R² was less than 1, indicating that the excluded taxon significantly influenced the clustering.

Published: September 23, 2025