Exploring COVID-19 causal genes through disease-specific Cis-eQTLs

Highlights

• •We systematically identified tens of thousands of COVID-19-specific cis-eQTLs, which were reported for the first time.
• •The individuals with TT genotype in rs1128320 were more susceptible to COVID-19 and had higher risks to develop severe COVID-19 based on the COVID-19-specific cis-eQTLs.
• •We excavated 48 causal genes by integrating the COVID-19-specific cis-eQTLs and GWAS data and validated them in siRNA-mediated depletion assay.

Abstract

Genome-wide association study (GWAS) analysis has exposed that genetic factors play important roles in COVID-19. Whereas a deeper understanding of the underlying mechanism of COVID-19 was hindered by the lack of expression of quantitative trait loci (eQTL) data specific for disease. To this end, we identified COVID-19-specific cis-eQTLs by integrating nucleotide sequence variations and RNA-Seq data from COVID-19 samples. These identified eQTLs have different regulatory effect on genes between patients and controls, indicating that SARS-CoV-2 infection may cause alterations in the human body's internal environment. Individuals with the TT genotype in the rs1128320 region seemed more susceptible to SARS-CoV-2 infection and developed into severe COVID-19 due to the abnormal expression of IFITM1. We subsequently discovered potential causal genes, of the result, a total of 48 genes from six tissues were identified. siRNA-mediated depletion assays in SARS-CoV-2 infection proved that 14 causal genes were directly associated with SARS-CoV-2 infection. These results enriched existing research on COVID-19 causal genes and provided a new sight in the mechanism exploration for COVID-19.

List of Abbreviations

genome-wide association study

(GWAS)

corona virus disease 2019

(COVID-19)

expression quantitative trait loci

(eQTL)

summary data-based Mendelian randomization

(SMR)

gene expression omnibus database

(GEO)

minor allele frequency

(MAF)

read depth

(DP)

mapping quality

(MQ)

Hardy-Weinberg equilibrium

(HWE)

differentially expressed genes

(DEGs)

unique molecular identifiers

(UMIs)

heterogeneity in dependent instruments

(HEIDI)

linkage disequilibrium

(LD)

Kyoto Encyclopedia of Genes and Genomes

(KEGG)

bovine serum albumin

(BSA)

1. Introduction

Genetic factors have been found to be associated with susceptibility to COVID-19. For example, SNPs at loci 3p21.31 and 9q34.2 were identified as genetic susceptibility loci, of which the latter was potentially involved in the ABO blood group system (Severe Covid et al., 2020). Moreover, a population with blood group A was reported to be associated with an increased SARS-CoV-2 infection risk, whereas group O was associated with a decreased risk (Zhao et al., 2021). Besides, some SNPs on ACE2 were also argued to be SARS-CoV-2 infection risk factors (Sienko et al., 2020). However, the mechanism that these SNPs influence COVID-19 remains obscure.

Recently, researchers found that variants of eQTLs in GTEx (https://www.gtexportal.org/home/) were associated with a poor prognosis of COVID-19 (Ramlall et al., 2020) and they believed that regulating gene expression might be one of the mechanisms by which SNPs involved in COVID-19 onset and progression. Currently, many databases, such as GTEx, offer available eQTL data. However, the samples used to analyze eQTLs were from diverse sources, meaning that disease status and other information were usually undistinguished. Thus, these eQTLs are not conditionally independent in certain respects. It is important to analyze conditional eQTLs in specific contexts as this could lead to improved fine mapping of GWAS associations (Dobbyn et al., 2018). Researches also indicated that in eQTL research, the objects should be refined and grouped as the characteristics of some subpopulations could be covered or mixed by others (van der Wijst et al., 2018), which is termed Simpson's paradox (Freitas, 2020).

One of the most significant applications of eQTLs is the discovery of causal genes, by integrating eQTLs with GWAS data. Liu et al. (2021) firstly reported potential causal genes for COVID-19, identified IFNAR2 and other potential genes which may be involved in the susceptibility or prognosis of COVID-19. Pairo-Castineira et al. (2021) demonstrated that low IFNAR2 or high TYK2 expression led to a high COVID-19 risk by integrating GWAS and eQTL data. However, the findings of previous studies might be interfered with the lack of disease specificity in the eQTL data they used, which did not strictly distinguish the disease from controls (Pairo-Castineira et al., 2021). Therefore, the COVID-19-specific eQTL data analysis has become a critical and essential task to explore COVID-19 mechanism. Similarly, researches on eQTL heterogeneity across different tissues may shed light on the understanding of COVID-19.

In recent years, analysis processes for identifying SNPs using RNA-Seq data have been vastly introduced and used to acquire eQTLs. For example, Vigorito et al. (2021) collected RNA-Seq data to obtain eQTLs. Subsequently, they identified a psoriasis-specific eQTL for GSTP1. Han et al. (2020) utilized RNA-Seq data to identify SNP-mediated lncRNAs related to multiple sclerosis mechanism. In this study, we combined COVID-19-specific eQTL data obtained from RNA-Seq data of SARS-CoV-2 infected samples and GWAS data to mine COVID-19 causal genes and genetic mechanisms to evaluate the similarities and differences among tissues.

2. Materials and methods

2.1. Data collection

RNA-Seq data of 575 SRR files (343 cases, 232 healthy controls) from six tissues (airway, blood, brain, heart, intestine, and lung) were collected from the Gene Expression Omnibus database (GEO), as depicted in Table 1. The cases were infected with SARS-CoV-2 and the controls were not. GWAS summary data was obtained from the COVID-19 Host Genetics Initiative (https://www.covid19hg.org/), which comprises 112,612 COVID-19 samples and 2474,079 controls (file name: C2_ALL_leave_23andme). The single-cell dataset of lung tissues was obtained from GEO (accession number: GSE171524).

Table 1.

Datasets used for cis-eQTL analysis.

Tissue	GEO Accession Number	COVID-19 SRR file number	Control SRR file number	Total
Airway	GSE150819; GSE153970; GSE156020; GSE175779	41	28	69
Blood	GSE152418; GSE166253; GSE179627	99	65	164
Brain	GSE157852; GSE164332; GSE179923	27	22	49
Heart	GSE150392; GSE151879; GSE156754; GSE162736; GSE169241	68	44	112
Intestine	GSE148696; GSE149312; GSE157059; GSE159201	16	15	31
Lung	GSE148697; GSE152586; GSE153218; GSE153277; GSE155241; GSE155518; GSE157057; GSE160435; GSE163547; GSE163959	92	58	150
Total		343	232	575

Note: Source files that were excluded are depicted in Supplementary Table S22.

2.2. Detection of SNPs and process of gene expression profiles

For bulk RNA-Seq data, quality control and sequence alignment were firstly conducted using fastp (Chen et al., 2018) and BWA software (Li, 2014), followed by variant calling utilizing bcftools (Narasimhan et al., 2016). The current human reference genome (GRCh38) was used. The SNPs were detected with the aid of the nucleotide sequence variations of the dbSNP database (https://ftp.ncbi.nih.gov/snp/organisms/human_9606/VCF/00-All.vcf.gz). Insertion and deletion variants (indels) were dropped and SNPs that did not meet the following criteria were filtered out: i) Minor Allele Frequency (MAF) > 1 %; ii) read depth (DP) > 10; iii) Mapping Quality (MQ) > 10; iv) meet the Hardy-Weinberg Equilibrium (HWE) (P_HWE > 5 × 10⁻⁵, R package ‘genetics’); v). were detected in > 10 % of the samples. SNP genotypes were converted to the standard forms (represented by 0, 1, and 2). Gene expression was quantified using feature counts after sequence alignment utilizing HISAT2 software (Kim et al., 2015), and the batch effect was eliminated using the R package 'limma' (Ritchie et al., 2015). Genes with zero values in > 10 % of the samples were excluded. Differentially expressed genes (DEGs) were identified by R package 'DESeq2′ (Love et al., 2014).

For single-cell RNA-Seq data, using the R package 'Seurat' (v4.0.5), data quality control was conducted to eliminate low-quality cells. There must be 300–3000 genes, 400–1000 unique molecular identifiers (UMIs), and less than 10 % of mitochondrial reads for each cell. The data was then scaled and normalized by the 'SCTransform' function, then the batch effect was removed by the 'RunHarmony' function, and the dimension reduction and clustering were conducted using 'RunUMAP'. The cell clusters were manually annotated according to the differentially expressed genes identified by function 'FindMarkers'. Annotation information was obtained from original dataset publications (Melms et al., 2021). We then conducted a differential analysis between COVID-19 patients and healthy controls for the proportion of cell types using Wilcoxon rank sum test.

2.3. Mapping analysis of cis-eQTLs

After preparing SNP genotype profiles, gene expression profiles (genes with zero values and genotypes with NA presented in more than 10 % of samples were excluded, and only the SNPs and genes in chromosomes 1–22 and X were included), and their corresponding annotation files (representing the position information of SNPs and genes), we split each file into COVID-19 and normal populations. R package 'matrixEQTL' (Shabalin, 2012) was applied to conduct the cis-eQTL mapping analysis (the distance between gene and SNP was set to < 1000 kb, P < 0.05).

2.4. Integrated cis-eQTL and GWAS analysis

Summary data-based Mendelian randomization (SMR) (Zhu et al., 2016) was used to identify COVID-19 causal genes. The method uses summary data from GWAS and expression quantitative trait locus studies to test for pleiotropic associations between gene expression levels and complex traits of interest. ‘–heidi-mtd’ was set to 1, ‘–diff-freq-prop’ was 0.2, ‘–diff-freq’ was 0.05, and ‘–peqtl-smr’ was 0.005. Subsequently, heterogeneity in dependent instruments (HEIDI) analysis was conducted to distinguish pleiotropy from linkage (43), utilizing the linkage disequilibrium (LD) reference provided by the 1000 Genome Project (https://data.broadinstitute.org/alkesgroup/FUSION/LDREF.tar.bz2). The SMR&HEIDI methods can be interpreted as an analysis to test whether the size of the SNP effect on the phenotype is mediated by gene expression (Zhu et al., 2016). Thus, the tool can be used to prioritize genes hit by GWAS for subsequent functional studies.

2.5. Functional and statistical analysis

The functions of genes regulated by cis-eQTLs in COVID-19 were analyzed using the R package 'clusterProfiler'. Differential expression evaluation was conducted using the Wilcoxon and one-way ANOVA test. Correlations between genes were evaluated using Pearson's correlation coefficient. The enrichment of specific genotypes in the populations was assessed using Fisher's exact test. The cluster analysis of multiple biological pathways is done by R package ‘GOSemSim’. The significance threshold was set at P < 0.05.

2.6. siRNA transfection assay

Lipofectamine RNAiMAX was incorporated to 10 μl OPTI-MEM using a Combi reagent dispenser to achieve a final dilution of 1:100. Then, 3000 Caco2 cells were incorporated to 40 μl complete medium. At 48 h post-transfection, cells were challenged with SARS-CoV-2 at 1.25 MOI. Forty-eight hours post-infection, plates were fixed using 4 % PFA in PBS for 4 h at 25 °C and then permeabilized with 0.4 % TritonX in PBS for 15 min at 25 °C. Subsequently, the plates were blocked with 10 % goat serum in 3 % bovine serum albumin (BSA) in PBS for 30 min at 25 °C, followed by incubation with primary antibody against SARS-CoV-2 Nucleocapsid Polyclonal (ThermoFisher Catalog # PA5–114,448) at 1:3000 in 3 % BSA in PBS at 4 °C overnight. The primary antibody inoculum was removed, and plates were washed thrice with PBS using a plate washer, and then incubated with anti-rabbit Alexa Fluor 488 (Invitrogen) at 1:3000 in PBS for 1 h at 25 °C. Then, the secondary antibody inoculum was removed, the plates were washed thrice with PBS using a plate washer, and DAPI was incorporated to the PBS. Finally, the plates were sealed and imaged using Opera Phenix. Two parallel experiments were conducted on the case and control groups. The SARS-CoV-2, B.1.1.7 (UK) strain used in this study (GenBank: MZ344997, GISAID virus name: hCoV-19/Hong Kong/HKU-210,318–001/2020, Accession no.: EPI_ISL_1,273,444) was isolated from the respiratory tract specimens of COVID-19 patients in Hong Kong and stored at the Physical Containment Level 3 Laboratory, Department of Microbiology, the University of Hong Kong.

3. Results

3.1. Identification of condition-dependent cis-eQTLs based on different groups

The workflow is illustrated in Fig. 1. We processed the RNA-Seq data and conducted cis-eQTL analysis. As a result, we obtained thousands of cis-eQTLs in the mixed samples (COVID-19 samples and healthy controls) of lung tissue (Fig. 2A), and the numbers of cis-eQTLs in other tissues are depicted in Supplementary Fig. S1A. These tissues were chosen based on the current common complications, such as encephalitis, arrhythmias, and diarrhea. Subsequently, we compared the cis-eQTLs obtained from the mixed samples with those from the GTEx database to test the repeatability of results. The results demonstrated that a small proportion (1.51 %) of these cis-eQTLs was previously reported in GTEx; however, 72.47 % of these had identical allelic directions (Fig. 2B), indicating that these cis-eQTLs reflected similar regulatory effects on genes and supporting the reliability of our approach to some degree. The existence of a few discordant cis-eQTLs may be due to differences in the sample composition and technical processes.

Fig. 1.

The overall workflow of our integrative approach.

Fig. 2.

Discovery of COVID-19-related cis-eQTLs. (A) Cis-eQTL numbers of lung tissue in different groups. Colorful bars on the left represent summarized cis-eQTL numbers, and gray bars on the top represent the numbers of cis-eQTL only shared by the groups marked with orange points. (B) Reproducibility of cis-eQTLs of lung tissue from mixed samples. Cyan areas include cis-eQTLs existed in GTEx, and brown and blue points represent the cis-eQTLs served as positive and negative regulators on their targeted genes, respectively. (C-D) COVID-19-related cis-eQTLs which regulatory direction remains unmasked in mixed population. (E-G) The regulatory effect of COVID-19-related cis-eQTLs upon unstratified and population stratification analysis. Cis-eOTL effect masks by the discordance in controls (E-F) or opposite allelic effects between case and control groups (G). (H) Results of IFITM1 siRNA-mediated depletion assay in SARS-CoV-2 infection. (I) Genotype-phenotype correlation analysis of rs1128320 in an independent cohort (GSE152418). (J) The differential IFITM1 expression levels in diverse severity groups.

The cis-eQTLs identified in the COVID-19 and healthy control groups demonstrated a wide range of differences in number and composition (Supplementary Fig. S1A). We take lung tissue, which has the most obvious COVID-19 clinical manifestations, as an example to illustrate these differences. Using mixed samples, 57,663 cis-eQTLs were obtained (Fig. 2A). However, stratified analysis by SAR-CoV-2 status found that 65,623 and 91,852 cis-eQTLs were detected in COVID-19 and healthy control samples, respectively (Fig. 2A). Among these, only 4993 cis-eQTLs retained the same regulatory effect on genes irrespective of whether the samples were mixed, which is a small proportion (8.56 %), considering that the results were obtained using the same mining process. When the regulatory effects of cis-eQTL in the COVID-19 was in the same direction as healthy control groups or was strong enough in the COVID-19 group, we could capture the cis-eQTLs associated with COVID-19 even with a mixed sample set. However, some peculiar eQTLs still differed between disease group and healthy controls. After sorting by P-values, we selected the top eQTLs that had all three genotypes (homozygous wild, heterozygous, and homozygous mutant type) as examples to elucidate the details. For RPL1AP10-rs796613144, the regulatory relationship was only significant in COVID-19 (Fig. 2C), and for COG3-rs41289553, the regulatory relationship was the opposite in COVID-19 to that of the healthy control group (Fig. 2D). Moreover, we found that some eQTLs that were undetected in the mixed samples had significant regulatory relationships in the COVID-19 group. For instance, in H2BC6-rs2494704, IFITM1-rs1128320, and TNFAIP3-rs8085, a regulatory relationship could only be observed after the subgroups were distinguished. The former two were only significant in COVID-19, and the latter was reversed in COVID-19 relative to the healthy controls (Fig. 2E–G). The other eQTLs demonstrating opposite effects are provided in Supplementary Table S1-S4.

3.2. The important role of IFITM1-rs1128320 in COVID-19 pathogenesis

In the above examples, we found that IFITM1, which encodes a protein that can restrict cellular entry by diverse viral pathogens, including influenza A, Ebola, and SARS-CoV-2 virus (Buchrieser et al., 2020), was strongly associated with COVID-19. IFITM1 can inhibit SARS-CoV-2 infection, as validated by single-cell transcriptome analysis (Shaath et al., 2020). We conducted siRNA-mediated depletion assay in the context of SARS-CoV-2 infection to verify the effect of IFITM1. The qRT-PCR results demonstrated that IFIMT1 depletion resulted in enhanced SARS-CoV-2 infection in Caco2 cells, indicating that IFITM1 is a strong restriction factor that inhibits SARS-CoV-2 infection (one-way ANOVA test, P = 1.82×10⁻⁵, Fig. 2H). In our study, rs1128320 T allele negatively regulated IFITM1 expression (Fig. 2F). Thus, we speculated that individuals with the T allele in rs1128320 might be more likely to be infected with SARS-CoV-2 and progress into severe COVID-19 than those with the C allele. According to the investigation, rs1128320 is located on the gene PSMD13, and its position on the genome is chr11:244,167 (GRCh38.p14), which mainly occurs variation of C>G / C>T, and will cause a missense mutation. However, it is not currently reported in ClinVar database. For further verification, we compared the frequency of diverse rs1128320 genotypes in an independent COVID-19 peripheral blood mononuclear cell dataset containing disease severity information (GSE152418, Fig. 2I). Of these samples, there were only two rs1128320 genotypes (CT and TT). As expected, the TT genotype proportion in the population that had been infected with COVID-19, including convalescent individuals, was 82.35 % (28/34), which was significantly higher than the 52.94 % (18/34) in the healthy control group (Fisher's exact test, P = 0.0186). This suggests that individuals with the TT genotype at the rs1128320 site may be more susceptible to COVID-19.

We subsequently compared IFITM1 expression levels in patients grouped by disease severity and in healthy controls (Fig. 2J). The results demonstrated that IFITM1 expression was significantly higher in patients with moderate symptoms than in healthy controls (Wilcoxon test, P = 1.7 × 10⁻⁸). However, with the aggravation of the condition of patients, IFITM1 expression gradually declined, and there was no significant difference compared to the healthy controls when the patients were admitted to the ICU (P = 0.35). When patients recovered from COVID-19, IFITM1 levels returned to levels similarly to those in healthy controls (P = 0.16). Based on these results, we inferred that, upon SARS-CoV-2 infection, the antiviral capacity of the human immune system is rapidly initiated. However, the T allele of rs1128320 could inhibit IFITM1 expression, which limits the resistance of human body to SARS-CoV-2, worsening the patient's condition. Consistently, 83.33 % (20/24) of COVID-19 patients with severe symptoms or in the ICU had the TT genotype, which was higher than 75.00 % (6/8) of patients with moderate symptoms. Given the IFITM1 alterations in COVID-19 patients, it is reasonable to assume that patients with the T allele in rs1128320 are more likely to develop critical illness.

In addition to IFITM1, we analyzed the expression of its family genes, IFITM2 and IFITM3. Unsurprisingly, we observed significant increase in the expression of these two genes in the COVID-19 group (P < 0.05), which is consistent with their antiviral ability. However, unlike IFITM1, we did not observe any differences in IFITM2/3 expression levels among COVID-19 patients with different severity (Supplementary Fig. S1B-C), suggesting that IFITM1, but not IFITM2 or IFITM3, might play a unique and critical role in COVID-19 progression.

Therefore, in-depth analysis of IFITM1 mechanism in COVID-19 is required, which may help in the developing specific drugs for critically ill patients.

3.3. Molecular mechanisms of COVID-19-specific cis-eQTLs

To learn about the functions of the cis-eQTLs in COVID-19 group, we need to analyze all cis-eQTLs obtained from disease group in the subsequent study, we termed the cis-eQTLs identified from COVID-19 samples as COVID-19-specific cis-eQTLs. It is important to note that we did not exclude cis-eQTL shared in health groups, the reasons were as follows: (i) the overlap of cis-eQTLs between disease and control group is rare; (ii) there is no clear indication that cis-eQTLs, shared with healthy controls, does not play a corresponding role in the COVID-19 disease process; (iii) the regulatory relationship between genes is usually complex, and there may be cooperative regulation between different genes so that the exclusion of some cis-eQTLs may affect the identification of the regulatory relationship between genes; (iv) even if cis-eQTL is shared between the two groups, there may be changes in effect, which means that even if a gene is regulated by the same eQTL in both groups, the expression effect may be different. The genes regulated by COVID-19-specific cis-eQTLs were presumed to be involved in COVID-19 pathogenic molecular mechanisms. Functional analysis revealed that these genes regulate the expression of viral genes and the life process of the virus (Fig. 3A). Especially, 'coronavirus disease-COVID-19′ pathway was directly found significantly enriched in the lung (P = 0.0072), airway (P = 0.0255), blood (P = 0.0011), and intestinal tissues (P = 0.0137). Interestingly, these genes were also involved in diverse nervous system disease and bacterial infection pathways in the Kyoto Encyclopedia of Genes and Genomes (KEGG) (Fig. 3B, Supplementary Fig. S1D-H). Consistent with previous studies, COVID-19 has been demonstrated to attack the nervous system, leading to diverse neurological complications including cerebral infarction, cerebral hemorrhage, and altered mental status (Aghagoli et al., 2021; Schulz et al., 2021; Vogrig et al., 2021). Additionally, studies have suggested that severe COVID-19 patients are more likely to have bacterial infections and, therefore, need to be treated with antibiotics (Ginsburg and Klugman, 2020), suggesting that the microbiome plays a key role in COVID-19 occurrence and development. Moreover, two ncRNAs, NEAT1 and MALAT1, which play critical roles in lung diseases (Jen et al., 2017) and were reported to be co-expressed in COVID-19 patients (Rodrigues et al., 2021), were also highly correlated in lung tissue in our study (P = 3.94×10⁻¹⁵, R = 0.71). Furthermore, ABL2 (Marchetti, 2020), ACADM (Huang et al., 2021; Liu et al., 2020), and BMP2 (Liu et al., 2020) were also shown to be abnormally expressed in the COVID-19 context.

Fig. 3.

Analysis of COVID-19-specific cis-eQTLs. (A) Gene Ontology biological processes, wherein COVID-19-specific eGenes are enriched. (B) KEGG pathways, wherein these COVID-19-specific eGenes are involved. Pink and blue modules represent a correlation with bacterial infections and neurological diseases, respectively. (C) KEGG pathways which high-frequent eSNPs are enriched in. eGenes involved in ‘Coronavirus Disease - COVID-19′ pathway are highlighted. (D) Manhattan plot of eSNPs. Red dots indicate highly significant (P ≤ 5 × 10⁻⁸) and clustered (< 1 Mb) eSNPs. (E) Susceptible eSNPs idendified in GWAS data. Of these, rs7144 and rs3781620 are associated with viral immunity or lung disease. (F) Circos plot summarizing genes that can be regulated by the susceptible eSNPs. Green lines in the outer circle represent differentially expressed genes between COVID-19 and control groups. Line height of the inside circle represents the numbers of regulatory SNPs.

We found that eGene (the gene regulated by eQTL) could be regulated by multiple eSNPs (single-nucleotide polymorphisms which were regarded as eQTL), and eSNP could regulate multiple eGenes, which means that most of them participated in multiple mechanisms simultaneously and that the higher the frequency, the more significant they are likely to be. Thus, we calculated eGenes and eSNPs frequencies (Supplementary Fig. S2A-B). According to the frequency distribution, eGenes with a frequency greater than 20 and eSNPs with a frequency greater than 10 were defined as high-frequency eGenes and eSNPs (Supplementary Table S5–6). Functional analysis of high-frequency eGenes demonstrated that they were predominantly concentrated in some cell life activities (Supplementary Fig. S2C-D), which is consistent with our expectation that these highly active genes may be essential. Meanwhile, high-frequency eSNPs were directly enriched in the 'coronavirus disease-COVID-19′ pathway (Fig. 3C).

We subsequently screened the top eSNPs (the most significant eSNP when there are multiple corresponding eSNPs of an eGene gene) according to the set threshold (P < 5 × 10⁻⁸), and their locations in the genome are depicted in Fig. 3D. We found that these top eSNPs formed 12 clusters (with at least three SNPs within 1 Mb, Supplementary Table S7). Functional and pathway analysis demonstrated that one of these clusters, comprising rs113009700, rs772408778, and rs750336994, was closely related to COVID-19, which was predominantly enriched in virus life activities and host defense mechanism, e.g., 'xenobiotic metabolic process', 'cellular response to xenobiotic stimulus', 'negative regulation of defense response to Virus by host', 'negative regulation of type I interferon-mediated signaling pathway', and 'Drug metabolism other enzymes' (Supplementary Fig. S2E-F). All three eSNPs were intron variants located in MICOS10. However, their specific mechanisms in COVID-19 require further verification.

We also conducted a conformance analysis with susceptible eSNPs found in GWAS (Supplementary Table S8-S13) and obtained 841 overlapping eSNPs in lung tissue (Fig. 3E, Supplementary Table S14), indicating that they may be highly susceptible eSNPs to COVID-19. Among these, the eSNP rs7144 was reported to be associated with measles virus-specific IgG levels and demonstrated significant differences between rs7144 genotypes and CD46 protein expression in T cells (Clifford et al., 2012). Additionally, evidence demonstrated that the eSNP rs3781620 was in nearly complete linkage disequilibrium with rs830083, which is related to an increased risk of lung cancer (Hu et al., 2006) (Fig. 3E). The top eGenes regulated by these overlapping susceptible eSNPs are depicted in Fig. 3F (Supplementary Table S14-S19). Among them, MAFF was previously demonstrated to be significantly dysregulated in COVID-19 patients (Ibrahim and Ellakwa, 2021) and was also an antiviral host factor that could suppress Hepatitis B virus core promoter transcription (Ibrahim et al., 2021). Additionally, RPL22L1, CGAS (Domizio et al., 2022), TRAF3 (Fu et al., 2021; Zheng et al., 2020), STAT2 (Miorin et al., 2020), RPL5, EIF2AK2 (Zheng et al., 2021), HBEGF, and C1S (Loganathan et al., 2020) are genes involved in the 'coronavirus disease-COVID-19′ pathway in KEGG. Although the roles of other eGenes in COVID-19 are not yet particularly clear, they provide a new direction for further research.

3.4. Discovery and verification of COVID-19 causal genes combined with GWAS data

Mining causal genes is critical for COVID-19. We obtained a set of GWAS data that were analyzed using large-scale COVID-19 samples. Using the SMR approach to combine cis-eQTLs detected in COVID-19 groups and GWAS data, we identified several COVID-19 causal genes (P_SMR < 0.05). The HEIDI test was subsequently conducted to distinguish between pleiotropy and linkage. Forty-eight genes passed the HEIDI test (P_HEIDI > 0.01, Fig. 4A), among which PTPRN2 was previously reported to be highly expressed in COVID-19 (Sharif-Askari et al., 2021). Additionally, ABL2 may cause COVID-19-driven endothelial damage and may be a potential target for therapeutic development (Marchetti, 2020). Several T cell receptor alpha variables (TRAV) were identified as COVID-19 causal genes in the blood in our study, which are pivotal components of T cell receptors (Liu et al., 2014) and are crucial in T cell-mediated virus clearance. For example, TRAV12–2 is one of the most frequently used gene segments in COVID-19 patients than in normal individuals (Wang et al., 2021). TRAV30, TRAV13–2, and TRAV20 are also associated with multiple diseases involving immune-related pathways (Zhou et al., 2020; Hong et al., 2020; Petersen et al., 2016).

Fig. 4.

The heterogeneity of causal genes and cis-eQTLs among different tissues. (A) Causal genes identified by SMR (P_SMR < 0.05) and which passed the HEIDI test (P_HEIDI > 0.01). Square color and size represent P_SMR value. The last row titled 'Susceptibility' represents whether the top eSNP of each causal gene is statistical significance in GWAS. (B) Results of siRNA-mediated depletion assay for the evaluation of SARS-CoV-2 infection. P-value above error bar reflects statistical effect. Only P < 0.05 is plotted. (C) Protein interaction network of COVID-19 key factors and causal proteins. (D) Correlation matrix of COVID-19 key factors and causal genes. Pie area and color represent the correlation coefficient. Insignificant correlations (P > 0.05) are represented as blank. (E) Co-expression relationship between FURIN and KDELR1 in COVID-19.

To further verify the roles that causal genes play in SARS-CoV-2 infection, we conducted the siRNA-mediated depletion assay targeting the protein-coding causal genes to investigate the roles of these genes in SARS-CoV-2 infection. As depicted in Fig. 4B, knockdown of ACE2, the critical receptor essential for SARS-CoV-2 infection, led to decreased infection and was used as a positive control. Depleting 14 of 35 protein-coding genes had a significant effect on the ability of SARS-CoV-2 to infect cells (Fig. 4B and Supplementary Table S20), which were considered as proviral (JPH3) or antiviral (CARMIL1, CTF1, DENND5A, DNMBP, FBXL19, GOLGA3, KDELR1, MEIOC, MEOX1, PIH1D1, REC8, SART3, and SH3PXD2A) host factors (one-way ANOVA test, P < 0.05). Utilizing a list of key SARS-CoV-2 infection associated factors and incorporating them into a protein-protein interaction network (string-db.org) together with these causal genes, KDELR1 was found to directly interact with FURIN and was indirectly linked with other key factors such as ACE2 (Fig. 4C). In terms of blood, five causal genes were validated, including KDELR1, and their co-expression relationships with key factors were investigated and many potential interactions were found. For example, KDELR1 had a positive co-expression relationship with most key factors, especially FURIN (Fig. 4D-E). Similar results were observed in other tissues (Supplementary Fig. S3A-C). Among these co-expression relationships, some were common in the COVID-19 and healthy control groups, while others were specific to the subgroups, including LINC02709 and DPP4, which were positively correlated in COVID-19 (Supplementary Fig. S3D) but negatively correlated in the healthy control group (Supplementary Fig. S3E), indicating a change in regulatory relationships in patients.

In conclusion, these results suggest that in COVID-19, causal genes may interact with key COVID-19 factors to play pathogenic roles, but the underlying pathways still need to be further explored.

3.5. Tissue-specificity of cis-eQTLs and causal genes

We also evaluated the composition of cis-eQTLs in multiple organizations. The number of cis-eQTLs identified in each tissue using the COVID-19 samples is depicted in Supplementary Fig. S4A. Even though the numbers of cis-eQTLs in each tissue are relatively high, tissue-shared cis-eQTLs are extremely rare, indicating that even under the same disease condition, the eSNP regulation on eGenes is different due to tissue specificity. Similarly, the eSNPs and eGenes involved in various tissues were significantly different (Supplementary Fig. S4B-C). The top 50 most significant biological processes of each tissue were collected and clustered according to semantic similarity and were divided into different functional modules (Supplementary Fig. S4D), among which cluster 3 was closely related to the life process of the virus. We found that the proportion of biological processes in the six tissues of each functional module varied to different degrees (Supplementary Fig. S4E). This further highlights the tissue specificity and necessity of differentiating tissues when studying COVID-19.

4. Conclusions and discussion

We identified COVID-19-specific cis-eQTLs and evaluated the regulatory changes between cis-eQTLs and genes according to SARS-CoV-2 infection status by strictly differentiating COVID-19 patients from healthy controls. These results suggest that when SARS-CoV-2 invades the human body, some molecular mechanisms may be modified, and the alterations might contribute to COVID-19 onset and progression, specifically observed in patients. And eQTLs with more biological significance could be found by targeting populations with specific phenotypes rather than by mixing all samples. Therefore, we believe that it is necessary to distinguish disease samples and healthy controls when identifying cis-eQTLs, which helps mine biological information diluted or neutralized by irrelevant subsets.

The subsequent results demonstrated that rs1128320 might be a COVID-19 susceptibility locus that exerts its influence by regulating IFITM1 expression. The analysis suggested that the role of IFITM1 in COVID-19 was peculiar, not the same as that of IFITM2/3, and was closely associated with severe/ICU COVID-19 patients. We also found that COVID-19-specific cis-eQTLs might be involved in multiple neuropathic pathways, which is consistent with the results of previous studies. Additionally, bacterial infection pathways were also observed, which provided new insights into COVID-19 infection mechanism and treatment, suggesting that microorganisms might play a role in COVID-19.

By combining the GWAS data, we identified 48 causal genes that may be related to COVID-19 occurrence and development. Further siRNA-mediated depletion experiments demonstrated the critical roles of 14 protein-coding genes including KDELR1 in SARS-CoV-2 infection, highlighting the potential for further mechanistic studies and drug development experiments. Although there is no direct evidence to prove the relationship between other genes and COVID-19, they still have great potential and warrant further study.

Cis-eQTLs also demonstrated great differences among various tissues, apart from disease specificity. It is well known that each human organ has a specific function, which is why previous studies were based on specific tissues to comprehend the generation or development mechanism of diseases. Therefore, it is essential to distinguish between tissues when studying COVID-19 mechanisms or treating its complications.

Previous studies have also verified that the number of certain cell types changes dramatically in COVID-19 and that such changes are tissue-specific (Melms et al., 2021; Wilk et al., 2020). For example, the number of macrophages in the lung tissue increased after SARS-CoV-2 infection (Ren et al., 2021). Therefore, we annotated the relative abundance of immune cells using the CIBERSORTx online tool (https://cibersortx.stanford.edu/) based on the lung tissue gene expression profile. We found that M0 macrophages were significantly increased in COVID-19 patients, as expected (P = 0.0068, Supplementary Fig. S5A). Additionally, resting memory CD4⁺ T (P = 0.0450) and regulatory T cells (P = 0.0063) also demonstrated significant differences between COVID-19 patients and healthy controls (Supplementary Fig. S5A). In another independent single-cell RNA dataset, GSE171524, comprising ten COVID-19 samples and seven healthy controls, we concluded that the number of macrophages increased in COVID-19 samples (32.6% vs. 13.9 %, Supplementary Fig. S5B), and the proportion of macrophages in each COVID-19 sample also increased (P = 1.6 × 10⁻⁵, Supplementary Fig. S5C). In Scovid (Qi et al., 2022), a database that provides a comprehensive resource of single-cell data for exposing COVID-19 molecular characteristics across multiple tissues, multiple causal genes were specifically expressed in some cell types (Supplementary Table S21). Thus, we hypothesized that some of these specific cell type changes are related to pathogenic genes, and eQTL may be not only disease- and tissue-specific but also cell type-specific (Perez et al., 2022; Yazar et al., 2022), which needs to be evaluated more carefully.

However, there are some limitations to our study. For example, although RNA-seq-based cis-eQTL identification overcame the insufficient data volume problem, there was still a challenge. The number of SNPs we identified was significantly less than that of traditional techniques, even if we relax the restrictions in the SNP calling process, resulting in a relatively low significance level in GWAS data colocalization (P < 0.05, Fig. 3E). Insufficient SNPs were available for our subsequent HEIDI analysis due to this deficiency; therefore, more causal genes could not be identified. And since we used RNA-seq data, we were only able to identify variants located in the exon region of the gene. And we didn't use GATK for identifying mutation sites, because it's too time-consuming compared with samtools, so we chose samtools instead due to the speed limit of our computer. Furthermore, the regulatory relationship between rs1128320 and IFITM1 has not been validated, and the relationship between IFITM1 and severe/ICU COVID-19 patients needs to be further explored. Most importantly, although we have directly observed that 14 causal genes play roles in SARS-CoV-2 infection, either inhibiting or promoting, further mouse model experiments are required to explore their specific mechanisms in depth. Relevant experiments are underway and the results will be published as soon as possible.

In conclusion, we elucidated the disease and tissue specificity of cis-eQTLs, verified the association between COVID-19 and genetic characteristics, and identified some COVID-19 susceptibility loci and causal genes. We provided a skewed method and perspective for research on the COVID-19 mechanism and emphasized that the research data should be more precise to avoid the interference caused by subgroups with different characteristics in future studies.

Data statement

RNA-Seq, scRNA-Seq, and GWAS data were downloaded from public databases. cis-eQTLs results identified from bulk RNA-Seq data are deposited in Github (https://github.com/liangcheng-hrbmu/COVID19-bulk-eQTL). Supplementary Tables 1 to 22 could be downloaded from this link: https://github.com/liangcheng-hrbmu/COVID19-bulk-eQTL/blob/master/Supplementary%20Tables.xls.

CRediT authorship contribution statement

Sainan Zhang: Writing – review & editing, Writing – original draft, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Ping Wang: Software, Methodology, Formal analysis. Lei Shi: Writing – review & editing. Chao Wang: Formal analysis. Zijun Zhu: Visualization. Changlu Qi: Visualization. Yubin Xie: Validation. Shuofeng Yuan: Validation. Liang Cheng: Writing – review & editing, Supervision, Conceptualization. Xin Yin: Validation. Xue Zhang: Conceptualization.

Declaration of competing interest

The authors declare no competing financial interests.

Data availability

• Data will be made available on request.