Peripheral immune progression to long COVID is associated with mitochondrial gene transcription: A meta-analysis

Abstract

SARS-CoV-2, the virus responsible for the COVID-19 pandemic, has led to millions of cases of Long COVID worldwide. Long COVID is a phenomenon characterized by persistent and debilitating mental and physical symptoms following acute infection. Despite ongoing research, trials, and considerable progress in understanding Long COVID, its exact causes remain only partially understood, with current hypotheses addressing specific aspects of the condition. We conducted one of the most comprehensive meta-analyses to date of all quality bulk RNA-seq studies worldwide from the COVID-19 pandemic and show significant mitochondrial transcript changes in the peripheral immune system of people with Long COVID, with unexpectedly low levels of intracellular viral RNA in Long COVID. This extensive analysis, which includes 26 studies and 1,272 individuals, shows that mononuclear cells, PBMC, and granulocytes from Long COVID patients exhibit significant alterations in mitochondrial genes and related processes. These findings likely represent the true transcriptomic landscape of Long COVID across diverse datasets, highlighting the long-lasting impacts of SARS-CoV-2 on peripheral immune function. In combination with other ex vivo and proteomics studies showing mitochondrial dysfunction, our results suggest critical new directions, such as the potential role of clonal hematopoiesis and infected seed cells. This work highlights the need for further investigation into the mechanisms underlying these immune changes and persistent symptoms in people with Long COVID. These findings will serve as a foundation for defining the paradigm underlying the biological mechanisms of Long COVID, driving research into the peripheral immune system, bone marrow, and mitochondria.

1. Introduction

SARS-CoV-2 (severe acute respiratory syndrome coronavirus 2), the causative virus of COVID-19 (coronavirus disease 2019), caused a major global pandemic. (Zhou, 2020; Maison, 2025) The Baltimore class IV virus has caused hundreds of millions of infections and millions of deaths worldwide. Recently, the WHO has declared an end to the global public health emergency. Although the acute phase of the pandemic has ended, long-term complications remain a significant public health challenge. Globally, more than 60 million people are experiencing Long COVID (LC). LC manifests following approximately 10 % of acute SARS-CoV-2 infections, and in the United States alone, at least 15 to 20 million people are affected by LC, with 5,000 deaths attributed to the illness. (Davis et al., 2023; Ely et al., 2024) There are also many unresolved questions surrounding the virus that require further clarification. In particular, the molecular mechanisms and natural history of the post-acute sequelae of chronic conditions that result from SARS-CoV-2 infection and, specifically, LC, are still unknown. (Davis et al., 2023; Al-Aly and Topol, 2024).

LC manifests with an array of debilitating symptoms across many organ systems, including relentless fatigue and persistent pain. These symptoms significantly intrude on everyday life, compromising the mental and physical health of those afflicted. (Groff, 2021; Cervia, 2022; Pinto, 2022; Ballering et al., 2022; Wanga, 2021; Davis, 2021) LC can manifest in many different ways, and is diagnosable when symptoms or conditions are periodic or constant for a minimum of three months. (Ely et al., 2024) There are many different hypotheses for the cause of LC, including viral persistence, immune dysregulation, microbiome dysbiosis, neuronal inflammation, endothelial abnormalities, and mitochondrial dysfunction.

We hypothesized that peripheral immune cells in LC would have a unique signature, and we observed mitochondrial dysfunction in aggregated mononuclear cels, peripheral blood mononuclear cells (PBMC), and granulocytes in people experiencing LC. First, to validate this cohort and these findings, we show that the transcriptome of Acute COVID (AC) is defined by known transcripts, pathways, and biomarkers. Next, we show LC differs from AC and no known COVID (NC) in many metabolic systems, including oxidative phosphorylation and mitochondrial gene expression. Then, between COVID Recovered (CR) and LC in mononuclear cells and PBMC, we found 13 and 65 differential transcripts, respectively, with eight mitochondrial genes (five nuclear-encoded and three mitochondrial-encoded). We demonstrate significant transcriptomic overlap between aggregated mononuclear cells and isolated PBMCs, despite some cell-type-specific differences. These findings, alongside previous evidence of mitochondrial dysfunction, suggest an altered state of mitochondrial function in myeloid and lymphoid cells at both transcript and functional levels. One possible explanation for such dysfunction is that, during viremic states of acute infection, SARS-CoV-2 RNA could infect hematopoietic stem cells (HSC), resulting in the destruction of evolutionarily-selected seed cells, (Bernstein, 2024) thus compromising the host’s adaptations. Crucially, the peripheral immune cells investigated in this study lack angiotensin-converting enzyme 2 (ACE2) and transmembrane protease, serine 2 (TMPRSS2), and are traditionally non-permissive to SARS-CoV-2 productive infection. Conversely, HSCs themselves are permissive to infection, and alterations in HSCs might be visible in their differentiated peripheral states. (Kazmierski, 2022) However, the possibility remains that SARS-CoV-2 may impact cells even after differentiation from HSC precursors, as there is still compelling evidence for long-term immune dysregulation and viral RNA persistence. As such, other possible explanations include persistent apoptosis or viral RNA or antigen uptake by immune cells, resulting in the chronic activation of mitochondrial antiviral pathways, increased ATP production, and mitochondrial dysfunction.

Herein, we evaluate bulk RNA-Seq data from 26 studies on peripheral immune cells. We identify the human transcriptome contributions to this disease and the viral RNA reservoir resident in these cells. We demonstrate considerably altered mitochondrial transcription from pre-COVID through post-COVID states. We show that intracellular viral RNA levels in peripheral immune cells are not higher in recovered individuals compared to LC patients. As such, we advance previous studies postulating mitochondrial dysfunction as the root cause of LC. (Al-Aly and Topol, 2024; Dirajlal-Fargo, 2024; Molnar, 2024) We explore mechanisms by which this mitochondrial dysfunction could occur, including persistent viral RNA activation, clonal hematopoiesis (CH) disruption, and ongoing apoptotic signaling. We demonstrate transcriptome differences between peripheral immune responses in LC patients, which may offer insights into the cause of LC and provide avenues for new research, including whole-genome sequencing of peripheral immune cells comparing pre-COVID to LC, mitochondrial staining, and apoptosis evaluation.

2. Methods

2.1. Protocol and search Strategy

Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) guidelines were used to perform this meta-analysis of publicly available RNAseq data acquired from immune cells in human studies. (Page, 2021; Haddaway et al., 2022) The Gene Expression Omnibus (GEO) was explored using the search term ((SARS-CoV-2 OR COVID) AND (rna-seq OR rna seq) AND Homo sapiens) to identify all relevant studies.

2.2. Eligibility Criteria, selection Process, data Collection, and risk of Bias assessment

All published studies that met the search criteria from the GEO search were explored for applicable datasets. Studies were further explored in the Sequence Read Archive (SRA). Two researchers independently screened the literature, GEO results, and SRA against inclusion criteria, and a senior author settled any disputes. Inclusion criteria were studies conducted with paired-end bulk RNA-seq on cells derived from HSC, including PBMC, monocytes, neutrophils, leukocytes, buffy coat, etc. The raw sequence reads for each GEO dataset were obtained from the corresponding SRA Accession using the SRA-Toolkit. A formal Risk of Bias assessment was not conducted due to the nature of the studies. Potential biases were mitigated by including the ‘study ID’ as a covariate in batch-correction to help control for any variability introduced through study design, cell type, execution, and reporting.

2.3. Exclusion criteria

We excluded runs that used single-end library preparations, did not disclose disease status, treated cells prior to RNA-Seq, or did not provide raw data via the SRA (Fig. 1).

Fig. 1. PRISMA.

Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) guidelines were followed for this study. Of 236 studies identified via the Gene Expression Omnibus (GEO) database search using ((SARS-CoV-2 OR COVID) AND (rna-seq OR rna seq) AND Homo sapiens), 26 were utilized. 194 were excluded due to cell type, 3 for not providing raw data in the Sequence Read Archive (SRA), 7 for utilizing single-end libraries, 5 for not specifying disease status in SRA files, and 1 for treating samples prior to RNAseq.

2.4. Data Items

Within the 26 RNA-Seq experiments, 59 unique conditions were assigned to runs. (Overmyer, 2021; Rother, 2020; Lee, 2021; Li, 2022; Caushi, 2021; Dean, 2021; Galván-Peña, 2021; Zhang, 2021; Lo Tartaro, 2022; Knabl, 2022; Lee, 2022; Lee, 2022; Lee, 2022; Brauns, 2022; Knabl et al., 2022; Lee et al., 2022; Lee, 2022; Lee, 2023; Giroux, 2022; Boribong, 2022; Yin, 2024; Rybkina, 2023) For analysis, these were re-categorized into 7 categories (NC, AC, Vaccinated, LC, CR, Multisystem Inflammatory Syndrome in Children (MISC), and MISC Recovered (Supplementary Table 1). The NC, AC, LC, and CR categories were utilized for subsequent analyses herein, and further divided into mononuclear cells, PBMC, and granulocytes (Table 1).

Table 1.

Unique Run ID Per Condition by Study from Sequence Read Archive.

BioProject	GEO ID	AC	NC	CR	LC	cell type	cell classification
PRJNA660067	GSE157103	200	52	0	0	Leukocytes	Mononuclear
PRJNA670179	GSE159678	0	3	0	0	Monocytes	Mononuclear
PRJNA682211	GSE151513	47	57	0	0	PBMC	Mononuclear
PRJNA708208	GSE168527	0	62	0	0	PBMC	Mononuclear
PRJNA730810	GSE174621	25	0	0	0	PBMC	Mononuclear
PRJNA741195	GSE178824	8	4	4	0	G-MDSC	Granulocyte
PRJNA743862	GSE179448	34	15	37	0	Treg	Mononuclear
PRJNA744408	GSE179627	33	22	15	0	PBMC	Mononuclear
PRJNA750398	GSE181005	6	6	0	0	memory CD4+	Mononuclear
PRJNA781246	GSE189039	24	0	0	0	Buffy Coat	Granulocyte
PRJNA787951	GSE190680	100	0	0	0	Buffy Coat	Granulocyte
PRJNA788374	GSE190747	0	68	47	0	Buffy Coat	Granulocyte
PRJNA789645	GSE191088	0	18	0	0	CD4,CD8,B	Mononuclear
PRJNA814290	GSE198256	7	11	16	0	Monocytes	Mononuclear
PRJNA832132	GSE201530	47	8	0	0	PBMC	Mononuclear
PRJNA832479	GSE201642	0	48	0	0	PBMC	Mononuclear
PRJNA844257	GSE205244	56	0	0	0	Buffy Coat	Granulocyte
PRJNA849921	GSE206263	55	34	0	0	PBMC	Mononuclear
PRJNA849954	GSE206283	12	24	0	0	Monocytes	Mononuclear
PRJNA898633	GSE217370	19	12	0	0	Neutrophil	Granulocyte
PRJNA931952	GSE224615	0	0	13	23	PBMC	Mononuclear
PBMC		207	231	28	23
Granulocyte		207	84	51	0
Mononuclear		466	360	81	23
TOTAL		673	444	132	23	1272

2.5. Human RNA-Seq analysis

Human reference sequence (GRCh38.p13) and annotations (release 109) were downloaded from the Ensembl database. The STAR aligner (version 2.7.9a) was used to generate genome index using the reference sequences and with the parameters runMode = ‘genomeGenerate’ and sjdbOverhang = 100. (Dobin, 2013) Quality control checks on raw sequence data were conducted using FastQC (v0.11.8). Following this, the data was cleaned and adapter sequences were removed with Trimmomatic. (Bolger et al., 2014) The cleaned reads were then reanalyzed with FastQC and aligned using STAR against the human reference genomes. The output was configured to generate sorted BAM files (outSAMtype BAM SortedByCoordinate), run in two-pass mode (twopassMode Basic), limit multimapping reads to one (outFilterMultimapNmax 1), and generate gene counts (quantMode GeneCounts).

2.6. SARS-CoV-2 RNA-Seq analysis

The SARS-CoV-2 reference sequence (wuhCor1) and its annotation were obtained from the Ensembl database. The STAR aligner was again used to generate the genome index using runMode = ‘genomeGenerate’ and ‘genomeSAindexNbases = 6.’ Trimmed reads from the above human analysis were then aligned to the SARS-CoV-2 index using the same parameters as above.

2.7. Downstream RNA-Seq analysis

The ‘ReadsPerGene’ files generated by STAR for each human and SARS-CoV-2 were used for further analysis with DESeq2. Differential expression analyses and visualization for human transcriptome was executed in R utilizing the following packages: ‘data.table’ for data manipulation, (Barrett, 2024) ‘DESeq2′ for differential gene expression analysis, (Love et al., 2014) ‘ggplot2′, (Wickham, et al., 2023) and ‘EnhancedVolcano’ for data visualization. (Blighe and Rana, 2023) Additionally, ‘clusterProfiler’ with ‘org.Hs.eg.db’ and ‘biomaRt’ were used for gene annotation and mapping. We accounted for batch effects between the studies by including ‘study’ as a covariate in the DESeq2 analysis. This allowed us to remove cell-type-specific differences as a confounder, as each study contained a unique cell type (Table 1). Analysis was done separately for mononuclear cells, PBMC, and granulocytes. Differentially expressed genes (DEGs) were identified at an adjusted p-value (FDR) threshold of < 0.05. The R packages ‘ggplot2′ and ‘ggstatsplot’ were used for analyzing the SARS-CoV-2 transcripts. SARS-CoV-2 transcripts were normalized by taking the ratio of reads mapping to SARS-CoV-2 to the total number of reads for each sample. (Moustafa et al., 2021) Differentially expressed genes in the CR to LC comparison were analyzed against CH driver genes.

2.8. Overlap analysis of differentially expressed genes between PBMC and mononuclear cell populations

To evaluate whether differential gene expression observed in PBMC was consistently represented across aggregated mononuclear cell populations, we performed a comparative analysis of two RNA-seq differential expression studies contrasting individuals with NC and AC. The PBMC dataset comprised 438 samples (231 NC, 207 AC), whereas the aggregated mononuclear dataset comprised 826 samples (360 NC, 466 AC) representing multiple cell types, including leukocytes, monocytes, regulatory T cells (Tregs), memory CD⁴⁺ T cells, and PBMC. We assessed overlap between the two datasets using Fisher’s exact test, with all genes (significant and non-significant) as the background.

2.9. Linear regression analysis of age effects on gene expression

To assess whether age influenced the gene expression differences we observed between LC and CR, we acquired age information from the study PRJNA931952, which included LC and CR, and combined it with CR samples from PRJNA744408, which also included age. Both studies used PBMC-only cells, resulting in LC n = 23 and CR n = 28. For each of the differentially expressed genes between LC and CR, 13 for mononuclear cells and 65 for PBMC, we used multiple linear regression to evaluate the independent impact of age and disease status on normalized gene expression levels (expression ~ age + condition). Analyses were performed in R and with Benjamini-Hochberg correction.

3. Results

In this meta-analysis of human and viral transcriptome data from 26 RNA-seq studies (Fig. 1), five studies involved vaccination or MISC and were not analyzed here, leaving 21 studies involving a total of 1,272 individuals (75.6 billion reads) across a spectrum of COVID-19 conditions, including AC (N = 673), NC (N = 444), CR (N = 132), and LC (N = 23). CR was determined uniquely by study and included definitions of those meeting discharge standard, disease severity, and time-course, and PCR negative. LC was defined as having LC symptoms for over 8 months. The DEG results are presented by comparing each condition against every other condition, first in mononuclear cells (with a PBMC subanalysis) and then in granulocytes. Each section lists the top GO pathways, KEGG pathways, and DEGs for that comparison.

3.1. Mononuclear Cells: People with AC had 19,983 DEGs from NC

Gene ontology (GO) analysis between NC and AC mononuclear cell-specific comparison revealed cell leading edge (adj.p = 9.37E-06), small GTPase-mediated signal transduction (adj.p = 9.61E-05), and activation of innate immune response (adj.p = 0.042). KEGG analysis of the mononuclear cell-specific comparison between NC and AC revealed immune system and infectious disease pathways, including Human papillomavirus infection (adj.p = 2.38E-03), MAPK signaling pathway (adj.p = 3.58E-03), and viral carcinogenesis (adj.p = 3.58E-03). In the NC vs AC mononuclear cell-specific comparison, the top differentially expressed genes were RMRP (log2FC = 2.69; adj.p = 5.21E-28), FILIP1L (log2FC = 1.31; adj.p = 5.00E-27), RN7SL4P (log2FC = 2.03; adj.p = 2.43E-25), SCARNA7 (log2FC = 2.02; adj.p = 1.01E-24), and RN7SKP180 (log2FC = 1.58; adj.p = 7.70E-24) (Supplemental Fig. 1).

3.2. Mononuclear Cells: People with CR had 14,734 DEGs from NC

The GO analysis focusing on mononuclear cells from NC versus CR showed mitochondrial and immune terms. These include mitochondrial gene expression (adj.p = 8.67E-23), mitochondrial translation (adj.p = 3.89E-22), and T cell receptor complex (adj.p = 6.70E-11). Insights from the KEGG analysis of mononuclear cells between NC and CR indicated metabolic terms. These include nucleocytoplasmic transport (adj.p = 5.68E-09), oxidative phosphorylation (adj.p = 6.17E-04), and viral carcinogenesis (adj.p = 2.28E-03). In the comparison specific to mononuclear cells between NC and CR, the most significantly differentially expressed genes were H4C3 (log2FC = 6.36; adj.p = 3.08E-79), processed pseudogene (ENSG00000203396) (log2FC = 5.67; adj.p = 8.68E-63), lncRNA (ENSG00000253824) (log2FC = 3.53; adj.p = 7.16E-51), EML1 (log2FC = 3.48; adj.p = 1.26E-45), and CADM1-AS1 (log2FC = 3.32; adj.p = 1.26E-45) (Supplemental Fig. 2).

3.3. Mononuclear Cells: People with LC had 1,516 DEGs from NC

Insights from the GO analysis of mononuclear cells between NC and LC indicated mitochondrial terms. These included mitochondrial gene expression (adj.p = 1.17E-12), mitochondrial protein-containing complex (adj.p = 3.35E-11), and mitochondrial translation (adj.p = 7.20E-11). KEGG pathway assessment, specifically focusing on mononuclear cells between NC and LC, demonstrated metabolism and cell cycle. These terms included fatty acid metabolism (adj.p = 8.32E-05) and cell cycle (adj.p = 1.02E-02). When comparing mononuclear cells specifically between NC and LC, the top significantly expressed genes were ANO7 (log2FC = 4.05; adj.p = 1.60E-13) H4C3 (log2FC = 5.96; adj.p = 2.26E-09), TRBV15 (log2FC = 4.24; adj.p = 8.98E-07), processed pseudogene (ENSG00000203396) (log2FC = 5.66; adj.p = 9.52E-07), and PRSS35 (log2FC = 4.11; adj.p = 4.70E-06) (Fig. 2).

Fig. 2. RNA-seq Analysis of Mononuclear Cells between No known COVID and Long COVID.

This figure displays the volcano plot of the 1,516 differentially expressed genes between No COVID (NC) and Long COVID (LC) and the top 15 resultant Gene Ontology (GO) terms for biological processes (BP), cellular components (CC), and molecular function (MF).

3.4. Mononuclear Cells: People with AC had 9,999 DEGs from CR

Findings from the GO analysis comparing mononuclear cells of AC to CR suggested mitochondrial terms. These include mitochondrial gene expression (adj.p = 7.58E-28) and mitochondrial translation (adj.p = 7.58E-28). The study of KEGG pathways in mononuclear cells, specifically between AC and CR presented metabolic and neurodegenerative categories. These included oxidative phosphorylation (adj.p = 1.85E-11), amyotrophic lateral sclerosis (adj.p = 3.44E-09), and thermogenesis (adj.p = 7.54E-08). In the targeted comparison of mononuclear cells of AC versus CR, the top differentially expressed genes were H4C3 (log2FC = 6.53; adj.p = 6.76E-89), processed pseudogene (ENSG00000203396) (log2FC = 5.97; adj.p = 5.00E-75), ZNHIT2 (log2FC = 3.36; adj.p = 6.42E-55), NICOL1 (log2FC = 4.14; adj.p = 1.56E-54), andARL2BP (log2FC = 4.06, adj.p = 3.48E-53) (Supplemental Fig. 3).

3.5. Mononuclear Cells: People with AC had 1,667 DEGs from people with LC

Results from GO analysis specific to mononuclear cells from AC to LC displayed mitochondrial terms. These terms include mitochondrial gene expression (adj.p = 1.03E-19), mitochondrial protein-containing complex (adj.p = 1.73E-18), mitochondrial translation (adj.p = 9.05E-18), and.

cellular respiration (adj.p = 8.28E-11). Analysis of KEGG pathways for mononuclear cells between AC and LC highlighted environmental adaptation, metabolic, and neurodegenerative terms. Among these terms are oxidative phosphorylation (adj.p = 2.72E-06), thermogenesis (adj.p = 4.42E-06), Parkinson’s disease (adj.p = 8.29E-05), and nucleotide metabolism (adj.p = 8.29E-05). In the AC vs LC mononuclear cell-specific comparison, the top differentially expressed genes were shown to be ANO7 (log2FC = 3.76; adj.p = 2.53E-11), H4C3 (log2FC = 6.12; adj.p = 3.74E-10), processed pseudogene (ENSG00000203396) (log2FC = 5.95; adj.p = 1.12E-07), lncRNA (ENSG00000238142) (log2FC = 2.42; adj.p = 2.35E-07), and CCDC85B (log2FC = 3.88; adj.p = 2.51E-07) (Fig. 3).

Fig. 3. RNA-seq Analysis of Mononuclear Cells between Acute COVID and Long COVID.

This figure displays the volcano plot of the 1,667 differentially expressed genes between Acute COVID (AC) and Long COVID (LC) and the top 15 resultant Gene Ontology (GO) terms for biological processes (BP), cellular components (CC), and molecular function (MF).

3.6. Mononuclear Cells: People with LC had 13 DEGs from CR

Examination of GO within mononuclear cells comparing CR and LC exposed one significant term of copper ion binding (adj.p = 0.016). Examination of KEGG pathways within mononuclear cells comparing CR and LC exposed no significant pathways. In the specific evaluation of mononuclear cells in CR compared to LC, 13 differentially expressed transcripts were identified. The genes that are most differentially expressed are ANO7 (log2FC = 3.72; adj.p = 2.40E-13), TMSB4XP4 (log2FC = −2.75; adj.p = 2.83E-05), IGKV6D-41 (log2FC = 2.78; adj.p = 2.73E-04), SULT1C4 (log2FC = −1.78; adj.p = 4.12E-03), and P2RY14 (log2FC = 2.55; adj.p = 8.90E-03) (Fig. 4A).

Fig. 4. RNA-seq Analysis of Mononuclear Cells between COVID Recovered and Long COVID.

This figure displays the volcano plot of A) the 13 differentially expressed genes between COVID Recovered (CR) and Long COVID (LC) for the mononuclear cell analysis and B) the 65 differentially expressed genes for the PBMC analysis.

3.7. PBMC: People with LC had 65 DEGs from CR

Neither GO nor KEGG analysis revealed any significant results when evaluating the DEGs between CR and LC. The most significantly differentially expressed genes are SNORA74A (log2FC = 14.86; adj.p = 1.75E-47), lncRNA (ENSG00000287189) (log2FC = 17.43; adj.p = 6.25E-42), MAT1A (log2FC = 17.13; adj.p = 2.55E-16), KRT9 (log2FC = 19.34; adj. p = 1.72E-14), and CKS1BP7 (log2FC = −4.89; adj.p = 3.39E-07) (Fig. 4B).

When the differential expression analysis was restricted to PBMCs only, we observed a greater number of differentially expressed genes between CR and LC. This discrepancy may indicate that these genes are differentially regulated in specific mononuclear subtypes, and their signal is diluted in mixed-cell populations. Differences may also reflect the reduced sample size in the PBMC-only COVID Recovered group (n = 28 vs. n = 81 in mononuclear cells), limiting statistical power.

3.8. Significant concordance of differential gene expression in PBMC and broader mononuclear analyses

Between NC and AC, we identified 9,168 DEGs in the PBMC analysis and 19,983 DEGs in the mononuclear analysis (FDR < 0.05). The intersection of these two sets contained 4,257 DEGs (46.4 % of PBMC DEGs, 21.3 % of mononuclear DEGs). Fisher’s exact test confirmed significant enrichment of shared DEGs beyond random expectation (odds ratio = 2.08, 95 % CI: 2.01 − ∞, p < 2.2E-16). Despite this substantial overlap, large proportions of DEGs remained unique to PBMC (4,911 DEGs) or mononuclear cells (15,726 DEGs), possibly indicating cell type specific differences. Conversely, the smaller sample size of the PBMC dataset (n = 438) compared to the mononuclear dataset (n = 826) may reduce statistical power, potentially limiting the detection of subtler transcriptomic differences and partially explaining the observed variation in DEG profiles.

3.9. Granulocytes: People with AC had 1,832 DEGs from NC

The comparison of GO in granulocyte specific to NC and AC uncovered response to virus (adj.p = 2.22E-15), defense response to virus (adj. p = 7.85E-10), and negative regulation of viral genome replication (adj. p = 1.23E-06). The KEGG analysis, focusing on granulocytes from NC versus AC, revealed pathways related to the immune system and infectious diseases. These included TNF signaling pathway (adj.p = 2.08E-06), Hepatitis B (adj.p = 4.26E-05), and Viral protein interaction with cytokine and cytokine receptor (adj.p = 2.47E-03). In the NC vs AC granulocyte-specific comparison, the most significant differentially expressed genes were processed_pseudogene (ENSG00000242197) (log2FC = − 4.45; adj.p = 3.26E-39), lncRNA (ENSG00000255097) (log2FC = 4.10; adj.p = 7.87E-34), lncRNA (ENSG00000289013) (log2FC = 4.57; adj.p = 3.09E-33), LSM3P2 (log2FC = 3.68; adj.p = 6.03E-32), and lncRNA (ENSG00000288948) (log2FC = 3.64; adj.p = 7.68E-30) (Supplemental Fig. 4).

3.10. Granulocytes: People with CR had 17,342 DEGs from NC

GO analysis between NC and CR granulocyte-specific comparison revealed mitochondrial and immune terms. Including T cell receptor complex (adj.p = 6.63E-15), mitochondrial gene expression (adj.p = 2.18E-04), and macorautophagy (adj.p = 2.18E-04). The comparison of KEGG terms in granulocytes specific to NC and CR uncovered mitochondrial and viral terms. These terms include mitophagy (adj.p = 9.11E-06), human papillomavirus infection (adj.p = 2.17E-02), and Coronavirus disease – COVID-19 (adj.p = 2.75E-02). Focusing on the granulocyte-specific differentiation between NC and CR, the most significantly differentially expressed genes were KDM5D (log2FC = −16.25, 5.87E-56), MTATP8P1 (log2FC = 3.19; adj.p = 3.73E-55), TXLNGY (log2FC = −15.64, adj.p = 4.67E-53), NACA3P (log2FC = 2.68; adj.p = 1.81E-51), and RPS4Y1 (log2FC = −14.82; adj.p = 4.75E-50) (Supplemental Fig. 5).

3.11. Granulocytes: People with AC had 2,377 DEGs from CR

GO analysis between AC and CR granulocyte-specific comparison revealed hemostatic process terms. These include regulation of body fluid levels (adj.p = 2.05E-04), platelet activation (adj.p = 4.89E-04), and hemostasis (adj.p = 3.59E-07). Results from KEGG analysis specific to mononuclear cells from AC to CR displayed infectious disease and signaling terms. These include platelet activation (adj.p = 2.52E-04), TNF signaling pathway (adj.p = 1.15E-03), and human cytomegalovirus infection (adj.p = 7.57E-04). In the detailed comparison of granulocytes from AC to CR, the most significantly dysregulated genes were MTCO2P12 (log2FC = 4.83; adj.p = 2.27E-30), MTND6P4 (log2FC = 5.15; adj.p = 1.02E-27), MTCO3P12 (log2FC = 4.69; adj.p = 6.14E-21), MTND5P11 (log2FC = 3.36; adj.p = 2.88E-16), and MTURN (log2FC = 1.44; adj.p = 1.31E-14) (Supplemental Fig. 6).

3.12. Comparison of differential expression analyses

Across LC mononuclear comparisons (NC vs LC, AC vs LC, and CR vs LC), there were 11 common differentially expressed genes (Fig. 5). These were unique in that their expression patterns, whether up-regulated or down-regulated, were consistent across all comparisons. Across all CR mononuclear comparisons (NC vs CR, AC vs CR, and LC vs CR), there are 0 common differentially expressed genes.

Fig. 5. Common Genes Across Long COVID Comparisons.

This figure displays the genes that were differentially expressed (log2FC) across all comparisons in mononuclear cells for Long COVID. Colors reflect log2 fold changes (log2FC).

3.13. Gene expression differences between Long COVID and Recovered groups are independent of age

None of the differentially expressed genes from the mononuclear analysis or the PBMC sub-analysis showed a significant association with age after correction (Supplementary Fig. 7). This indicates that the differences in gene expression observed between LC and CR are not caused by age differences, but instead result from disease-related gene expression changes.

3.14. Viral transcriptome

Using only the cells assigned to a study condition (NC, AC, CR, and LC), there are 342 granulocyte samples and 930 mononuclear cell samples. Evaluating these cells for SARS-CoV-2 viral RNA using the modified RNA-seq analysis pipeline described above, we detect a significant amount of SARS-CoV-2 in mononuclear cells compared to granulocytes (p = 4.40E-11) (Fig. 6A). Among conditions, there was significantly lower SARS-CoV-2 RNA in LC than both AC (adj.p = 9.13E-04) and NC (adj.p = 3.56E-05). Further, the amount of detectable RNA transcripts aligning to SARS-CoV-2 compared to total RNA transcripts per sample was lowest among LC (μ = 1.32E-06), followed by CR (μ = 9.37E-06). The highest amount of viral RNA transcripts was in AC (2.68E-05) (Fig. 6B). Among cell types, the mean of monocytes and PBMC was 3 logs higher than all other cell types (Fig. 6C).

Fig. 6. SARS-COV-2 RNA-seq.

This figure illustrates the RNA-seq alignment to the SARS-CoV-2 genome, normalized by the ratio of SARS-CoV-2 aligned reads to total reads. A) T-test comparing the ratio of SARS-CoV-2 aligned reads to total reads between granulocytes (n = 342) and mononuclear cells (n = 930) (p = 4.40E-11). B) Welch ANOVA comparing conditions in mononuclear cells: acute COVID (AC) (n = 466), COVID recovered (n = 81), Long COVID (n = 23), and no known COVID (NC) (n = 360) (p = 2.11E-08). C) Welch ANOVA in mononuclear cells comparing cell types: CD4/CD8/B (n = 18), leukocytes (n = 252), memory CD4+ (n = 9), monocytes (n = 73), PBMC (n = 493), and Treg (n = 86) (p = 2.39E-16).

4. Discussion

In this meta-analysis of 21 studies involving a total of 1,272 participants, we demonstrate considerable transcriptome differences between people with NC, AC, CR, and LC in aggregated mononuclear cells, PBMC, and granulocytes. Overall, in support of our hypotheses, we observed mitochondrial dysfunction in mononuclear cells and granulocytes in people experiencing LC. Whereas the direct cause of immune cellular/metabolic function is not known, our data suggest that SARS-CoV-2 may have a long-term impact on a variety of different cell types, either due to direct or indirect effects of viral infection and potential viral persistence.

In comparing the mononuclear cell transcriptomes, the comparison between NC and AC showed expected viral infection markers, including genes and ontologies related to the activation of innate immune response, GTPase, and human papillomavirus infection. (Hou, 2022) In all the remaining comparisons, looking at the changes from NC and AC to after infection (CR and LC), there were consistent changes in mitochondrial-related genes and ontologies. LC had specific top ontologies of mitochondrial gene expression and mitochondrial protein-containing complex compared to NC. However, the variability of differential expression in the PBMC-only comparison, suggests that these mitochondrial signals may be specific to certain mononuclear cell subtypes. These findings emphasize the importance of cell-type-specific resolution and adequate sample size when investigating the immune transcriptome in Long COVID.

Consistent with prior limited RNA-seq studies on LC, we found only a small set of differentially expressed genes (n = 13 mononuclear and n = 65 PBMC) between CR and LC. (Yin, 2024) In that study by Yin et al., using the Long-term Impact of Infection with Novel Coronavirus (LIINC) cohort, (Peluso, 2022) with a much smaller sample size and utilizing the GRCh37 genome reference, the main findings were two differentially expressed OR7D2 (Olfactory Receptor Family 7 Subfamily D Member 2) and ALAS2 (5′-Aminolevulinate Synthase 2). While neither of these genes was found to be differentially expressed herein, it is noteworthy that a mitochondrially related gene was differentially expressed in the LIINC study within a geographically related cohort (ALAS2). In this study comparing CR to LC with a larger sample size and using the updated GRCh38 human genome reference, we found 13 DEGs in aggregated mononuclear cells and 65 DEGs in PBMC, with eight genes related to mitochondria – five nuclear-encoded and three mitochondrial-encoded (one from the mononuclear comparison and seven from the PBMC comparison) (MT-CYB, MT-CO3, MT-TV, MTND2P28, MTND6P3, MTCO1P40, MTCO2P22, and MTCO2P12). MT-CYB (Mitochondrially Encoded Cytochrome B) and MT-CO3 (Mitochondrially Encoded Cytochrome C Oxidase Subunit III) are directly involved in ATP production and were seen with considerable upregulation in LC. This finding is consistent with our previous ex vivo studies, conducted in a completely different cohort, which demonstrated increased ATP production in LC PBMC.(Dirajlal-Fargo, 2024).

Comparing NC with AC between PBMC and all mononuclear cells, our results show statistically significant overlap in DEGs identified in mononuclear cells and those identified solely in PBMC analyses comparing NC and AC groups. This notable overlap highlights a fundamental transcriptional response shared across multiple immune cell subsets. Nonetheless, many DEGs remained unique to either PBMC or aggregated mononuclear analyses, likely reflecting biological heterogeneity among cell subsets and differences in statistical power due to varying sample sizes. These findings underscore the importance of interpreting mixed-cell population studies with caution due to inherent biological and statistical limitations. Future investigations using larger, cell-type-specific datasets or single-cell RNA sequencing are essential to accurately identify contributions of distinct immune subsets in COVID-19-associated transcriptional responses.

In the granulocyte cell transcriptome evaluation, the NC to AC also showed what would be expected for a coronavirus infection. Notably, response to virus and viral processes were identified in to GO analyses of NC to AC DEGs. The NC to CR comparison, showed gene ontologies related to mitochondrial and immune processes. The AC to CR comparison did not show mitochondrial genes, and instead involved physiological hemostasis. These differences between mononuclear cells and granulocytes may reflect the method of granulocyte uptake of circulating antigens, which can persist in plasma for up to 14 months post-infection, and other body sites for up to 22 months post-infection. (Peluso, 2024; Proal, 2023; Swank, 2023) The absence of mitochondrial changes in AC to CR granulocytes could also reflect granulocyte’s inherently low mitochondrial number and predominantly glycolytic metabolism. (Vorobjeva et al., 2023).

There are several possible explanations for these findings of mitochondrial transcriptome dysregulation in mononuclear cells before COVID, after acute COVID-19, and proceeding either to a normal recovery or LC. The first involves persistent viral RNA or antigen being taken up by peripheral immune cells, leading to prolonged activation of the intracellular mitochondrial sensing virus. (Maison et al., 2023) Immune cells take up viral RNA or protein once they enter circulation, triggering an immune response. This activation by viral RNA predominantly occurs via the mitochondrial protein MAVS (mitochondrial antiviral signaling), and viral RNA can further cause damage to mitochondria. (Maison et al., 2023) The activation of an immune response would increase ATP production and oxygen consumption by immune cells, resulting in LC symptoms, as described previously. (Dirajlal-Fargo, 2024) Further, this is supported by SARS-CoV-2 antigens persisting in plasma as discussed above. (Peluso, 2024; Proal, 2023; Swank, 2023) While we detected low intracellular viral RNA in circulating cells, persistence may occur in other forms, such as antigens, or in tissue reservoirs, such as the gut or bone marrow.

The second possible explanation for the mitochondrial expression is related to CH. CH mutations are an integral part of aging that confer fitness advantages. (Bernstein, 2024) These mutations are selection-based, meaning specific stem cells are selected to seed the body as a person ages. Variants that confer fitness advantages are selected and expand over time. (Bernstein, 2024; Watson, 2020) Interestingly, these seed cells contain ACE2 and TMPRSS2, unlike the cells that they differentiate into – peripheral immune cells. (Kazmierski, 2022; Ratajczak, 2021) This leads to the hypothesis that, during acute infection, selected seed stem cells are infected and destroyed by SARS-CoV-2. This would mean the host would need to revert to a previously unselected lineage, which does not contain the adaptive mutations that have occurred throughout that person’s lifetime and conferred fitness to that individual. This immature lineage, lacking a lifetime of accumulated advantageous mutations, may contain different copy numbers of mitochondrial DNA, be more mitochondrially active, or have adaptations specific to mitochondria, which could explain the mitochondrial expression findings herein. The degree of change and the severity of the loss of fitness would differentiate between those who recover and those who progress to LC. Consistent with this hypothesis and in the context of CH drivers, as defined by Bernstein et al. (2024), we found that the comparison of mononuclear AC to CR progression revealed six CH driver genes: three classical (PPM1D, BRCC3, and SRSF2) and three novel (CCL22, CCDC115, and ZNF234). In contrast, during the AC to LC progression mononuclear comparison, we identified only two novel drivers (CCL22 and CCDC115). In the granulocyte analysis, the AC to CR comparison contained the novel CH driver SRCAP, upregulated in CR. These findings suggest that recovery requires a robust clonal rebound across HSC clones. While Bernstein et al. identified CH drivers based on mutations that confer selective advantages at the DNA level, our findings highlight the differential RNA expression of these same genes, suggesting changes in selective pressures or altered functional states. While permanent DNA mutations in CH genes are known to increase lifelong infection risk, we hypothesize that a severe, bone-marrow-infiltrating SARS-CoV-2 infection can induce a similar state of immune impairment via the sustained dysregulation of these genes’ RNA expression. Another interpretation is that SARS-CoV-2 infection reshapes stem cell niches, introducing new selective pressures on HSC populations. However, as stated, this study measures RNA expression levels, not the underlying genetic mutations that define CH. Whole-genome sequencing of whole blood in LC and CR patients, compared to matched whole blood samples before infection (and the pandemic), and mitochondrial function, genetics, and microscopy, would be needed to explore this theory further.

Finally, the third hypothesis for mitochondrial expression is that peripheral immune cells following AC are apoptotic. Apoptosis is a primary form of cell death during AC. (Yuan, 2023) Mitochondria are linked to and regulate apoptotic protein levels. This is due to an increase in reactive oxygen species produced as mitochondrial levels increase. Mitochondria are involved in both the intrinsic and extrinsic pathways of apoptosis. Additionally, caspase-8 (CASP8), seen over-expressed in LC in the PBMC-only CR vs LC herein, is involved in the extrinsic apoptosis pathway. (Yuan, 2023; Tummers and Green, 2017; Márquez-Jurado, 2018) Further, MTCO2P12, which contributes to cytochrome c (a component of the intrinsic apoptotic pathway) and caspase-9 activation, is over-expressed in CR in the mononuclear CR vs LC herein (Wang and Youle, 2009), suggesting different apoptotic pathways between CR and LC. Although apoptosis primarily occurs during acute infection, persistent apoptotic signaling might sustain mitochondrial dysregulation. Previously, we showed no difference in cell viability between CR and LC. (Dirajlal-Fargo, 2024) As such, this is likely not a measure of active apoptosis, and further confirmation or rejection of apoptosis could be obtained by looking at peripheral immune cells of LC patients using flow cytometry with Annexin V and propidium Iodide staining. Other researchers have shown increased annexin V and apoptosis in T4 cells during the acute phase for patients who eventually developed Long COVID. (Cezar, 2023) Apoptosis of immune cells after SARS-CoV-2 infection may create opportunities for other viruses like EBV, CMV, and HIV to cause damage. All of these have been shown to influence the likelihood of LC and are related to symptomatology. (Peluso, 2023) The idea that SARS-CoV-2 has a widespread effect on these metabolic processes is supported by the prior discovery of similar dysregulation of the mitochondrial and autophagy pathways in human nasopharyngeal samples and lung epithelial cell lines infected with SARS-CoV-2. (Singh, 2021) By demonstrating persistent mitochondrial dysregulation in individuals with LC, our study extends these findings to peripheral immune cells and suggests that these pathways may contribute to immune dysfunction that persists after the acute infection.

The data herein, which shows detectable RNA in mononuclear cells – with a preference for PBMCs and monocytes – but not granulocytes, suggests that the seed hematopoietic stem cells, myeloid progenitor cells, and lymphoid progenitor cells do not harbor residual SARS-CoV-2. Monocytes likely uptake residual viral RNA while circulating, as they do with intact SARS-CoV-2. (Junqueira, 2022) Furthermore, the detectable RNA in the No COVID (NC) control group can likely be attributed to the heterogeneity of that cohort, as detailed in Supplementary Table 1, the varied inclusion criteria across studies allowed for the potential inclusion of individuals with prior known exposure. This suggests that viral RNA persistence in circulating immune cells may not fully explain the observed mitochondrial gene changes. The hypothesis of viral persistence is strongly supported as one component of LC, yet our findings suggest mitochondrial dysfunction may independently or synergistically contribute to persistent symptoms. It is essential to mention that SARS-CoV-2 has been detected in bone marrow post-mortem, suggesting that the virus does infiltrate the bone marrow, the HSC compartment, but may not persist in people who recover from COVID-19.(Jurek, 2022).

4.1. Limitations

Our study has several limitations. First, although we included the SRA study ID as a covariate to adjust for study-specific effects statistically and some cell-type differences, our meta-analysis encompasses a wide range of immune cell types, including leukocytes, monocytes, PBMCs, granulocytes, G-MDSCs, memory CD4 + T cells, buffy coat, and neutrophils. This biological heterogeneity cannot be fully accounted for through statistical adjustment, as it may confound some of the observed transcriptomic differences, given that these cells possess inherently distinct expression profiles. This heterogeneity may also mask condition-specific signals. Additionally, some mitochondrial gene changes may reflect residual differences in immune cell composition, general immune activation, medication use, or viral persistence rather than Long COVID-specific mechanisms alone. We cannot entirely exclude these alternative explanations.

Second, the PBMC-only sub-analysis of mononuclear cells, despite having a smaller sample size in the CR group, yielded substantially more differentially expressed genes than the broader mononuclear cell analysis and resulted in different mitochondrial genes than were identified in the whole mononuclear cell analysis. This discrepancy may reflect both biological effects specific to aggregated mononuclear cells compared to PBMC and reduced statistical power in the smaller PBMC-only subgroup (n = 28 for CR in PBMC compared to n = 81 for CR in mononuclear cells).

Lastly, biological variability across cohorts, differences throughout the span of the pandemic, and variability in SARS-CoV-2 exposure or re-exposure likely affected detectable viral RNA levels and may have influenced the findings.

5. Conclusions

This study demonstrated significant transcriptome differences in mononuclear cells and granulocytes between no known COVID, acute COVID, COVID recovered, and Long COVID. Our findings suggest that SARS-CoV-2 infection alters mitochondrial transcription in peripheral immune cells long-term. These findings provide insights into possible mechanisms underlying persistent symptoms and suggest avenues for further research. We suggest three possible explanations for this ongoing mitochondrial dysregulation: (1) prolonged immune activation by persistent viral RNA or antigens; (2) disruption of hematopoietic clonal dynamics as a result of SARS-CoV-2 infection; and (3) sustained apoptotic signaling within immune cells. Collectively, our results establish a basis for further research on the function of immunometabolism in post-viral syndromes and emphasize mitochondrial dysfunction as a recurrent aspect of Long COVID immune dysregulation.

Supplementary Material

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.mito.2025.102072.

Data availability

All differential expression results files, GO results files, KEGG results files, and code are available at github.com/davidpmaison/longcovid_metaanalysis. All data is publicly available through the Sequence Read Archive. This review was not registered with PRISMA.