Sample metadata tables were downloaded from The Cancer Genome Atlas (TCGA) (Genomic Data Commons (GDC) portal: https://portal.gdc.cancer.gov). Out of 11,093 aliquots total, 10,732 were selected to keep one unique aliquot per patient per sample type, as illustrated in online supplementary figure 1. The final cohort consists of 9,657 primary tumors, 367 metastatic samples (all from skin cutaneous melanoma), and 708 normal tissues from 10,038 patients across 34 tumor types (33 primary and one metastatic) (online supplementary tables 1 and 2). Out of 708 normal tissues, 14 tumor types have 15 or more normal samples available and hence were included for statistical comparisons when applicable. The standardized, upper-quartile normalized, batch-corrected, and platform-corrected RNAseq expression of 20,531 genes in RNA-Seq by Expectation Maximization (RSEM)-quantified read count estimates were downloaded from Pan-Cancer Atlas consortium studies (https://gdc.cancer.gov/about-data/publications/pancanatlas) and log2-transformed for further analysis. Fragments Per Kilobase of transcript per Million mapped reads (FPKM) quantification of gene expression and whole-exome sequencing (WES) Binary Alignment Map (BAM) files were downloaded from GDC.33 Demographic and clinical information were retrieved from the TCGA Pan-Cancer Clinical Data Resource.34 BMI and smoking history were retrieved from legacy clinical files. To the authors’ knowledge, only two tumor types had diabetes status information (pancreatic adenocarcinoma (PAAD) and uterine corpus endometrial carcinoma (UCEC)), which was retrieved from clinical XML files on GDC. Eighty-two percent of patients with PAAD and 7% of patients with UCEC have diabetes status recorded. Hence, only PAAD was included in analysis. Commensal microbiota abundance and viral presence of TCGA samples were downloaded from published studies.35 36 Single-cell RNAseq (scRNAseq) gene expression in malignant cells, immune cells, and normal cells were retrieved from nine studies consisting of patients diagnosed with cancer and healthy donors. Links to data files, single-cell cohorts, and bioinformatics software are provided in online supplementary table 3. Data generated in this study are accessible on GitHub repository https://github.com/riyuebao/ACE2_TMPRSS2_multicorrelates.

The gene expression of ACE2 (Entrez Gene ID 59272) and TMPRSS2 (Entrez Gene ID 7113) was retrieved from the RSEM-quantified RNAseq data and used for all analyses described in this study. Spearman’s correlation was calculated between the expression of the two genes in tumor (n=10,024) and normal (n=708) samples across all tumor types and within individual tumor types. The expression percentile was calculated separately within each of the four analysis sets (ACE2 in normal, ACE2 in tumor, TMPRSS2 in normal, TMPRSS2 in tumor) following two steps. First, the median expression of ACE2 or TMPRSS2 was calculated within individual tumor types. Next, tumor types were ranked by the median expression of each gene from higher to lower, and the position of each tumor type in the ranked list was scaled to 0–100, hereafter regarded as “expression percentile” per tumor type, with smaller values indicating top-ranked tumor types. The same process was repeated for each gene in tumor samples (34 tumor types) and normal tissues (14 tumor types).

The expression of ACE2 and TMPRSS2 was compared between designated clinical groups, split by age (younger (<65 years)/older (≥65 years) in tumor or normal), gender (female/male in tumor or normal), race (African American (AA)/Asian/White in tumor or normal), menopause (not post/post in tumor or normal), BMI (level 1 (<25)/level 2 (25–30)/level 3 (30–35)/level 4 (>35)37 in tumor or normal), smoking history (never/light/heavy38 in tumor or normal), tumor stage (I/II/III/IV in tumor), tumor grade (G1G2/G3G4 in tumor). For tumor grade, G1 and G2 were collapsed to indicate low-grade to mid-grade (G1G2), and G3 and G4 were collapsed to indicate high-grade (G3G4). Within each clinical factor, sub-groups of <15 samples were excluded. For each clinical factor, comparisons were performed across all tumors and within individual tumor types. For all tumor types, first, data were fitted into a two-way analysis of variance (ANOVA) model with tumor type and clinical group as variables plus the interaction between the two. Second, if more than two clinical groups exist, Tukey’s honest significance test (HSD) was used with the fitted ANOVA model for pairwise comparisons while controlling for type I errors. Within each tumor type, Tukey’s HSD was used with one-way ANOVA models when more than two groups are present; otherwise Welch Two Sample t-test was used. The list of clinical groups and statistical results are provided in online supplementary tables 4–7.

HLA-A, HLA-B, and HLA-C genotypes were identified for 9,559 patients from TCGA across 34 tumor types using OptiType (version 1.3.2) with WES BAM files. Considering each patient carries two copies of HLA-A, B, or C alleles, both copies were counted toward the total number (19,118) of A, B, or C alleles in the entire cohort. The calculation of HLA prevalence was performed at the allele level. The comparison of ACE2 and TMPRSS2 gene expression between HLA genotypes was performed at patient level, given that gene expression was estimated per sample. First, for each allele, labels 0, 1, or 2 were assigned per patient (0=this patient does not carry this allele; 1=this patient carries one copy of this allele; 2=this patient carries two copies of this allele). Then, gene expression differences were tested using two-way ANOVA models with HLA genotype (0, 1, 2) and tumor type as variables plus the interaction between the two. For each test, at least 15 samples carrying the allele are required. The test was repeated for every HLA-A, B, and C allele with gene expression in tumor or normal samples, followed by Benjamini and Hochberg-False Discovery Rate (BH-FDR) correction for multiple comparisons. Same analysis was repeated with age, gender, and race as covariates in the model.

Five immune responsive and suppressive signatures (interferon-stimulated genes (ISG), T cell-inflamed (Tinfl), myeloid, angiogenesis (angio), and transforming growth factor-β (TGF-β)) (online supplementary table 8) were correlated with ACE2 and TMPRSS2 gene expression in tumor and normal tissues. The expression level of a signature was computed as the average expression of all genes in this signature after centering and scaling using R function scale with parameters scale=TRUE and center=TRUE. Spearman’s correlation was calculated between each signature and ACE2 or TMPRSS2. The full correlation metrics are provided in online supplementary table 9.

FPKM estimates of RNAseq gene expression was used for quantifying enrichment of 64 tumor and stroma cell types using xCell (version 1.1.0). xCell converts gene expression into rank-based metrics within each sample; hence, normalization and batch correction were not required. To make data comparable across samples, xCell was run once using all samples (n=10,732). Spearman’s correlation was computed between the enrichment score of each cell population and ACE2 or TMPRSS2 expression. The full correlation metrics are provided in online supplementary table 10.

The human papillomavirus (HPV), Epstein-Barr virus (EBV), and hepatitis B virus (HBV) were selected for this study as those are the three most prevalent cancer-associated viruses in the cohort. Other viruses detected were excluded. Samples were set to HPV positive/negative by at least 10 normalized reads per million (NRPM) mapped to the HPV genomes, EBV positive/negative by at least 5 NRPM mapped to the EBV genomes, and HBV positive/negative by at least 5 NRPM mapped to the HBV genomes given previously recommended thresholds.36 Stomach adenocarcinoma (STAD), esophageal carcinoma (ESCA), acute myeloid leukemia (LAML), and ovarian serous cystadenocarcinoma (OV) were reset to “negative” for HPV, EBV, and HBV after a manual inspection, which revealed no strong clinical support for viral presence in those tumor types. Within each tumor type, ACE2 and TMPRSS2 gene expression was compared between viral positive and negative tumors using Welch Two Sample t-test. Across all tumor types, two-way ANOVA was used to compare gene expression between viral positive and negative tumors with tumor type and viral group as variables plus the interaction between the two.

The abundance of 1,093 genus-level microbial taxa was quantified from tissue RNAseq data after rigorous quality control, batch correction, and contamination filtering, and normalized to 1 million reads to make data comparable across samples.35 Seven hundred and six normal tissues and 9,801 tumor samples were included in the analysis where data were available. Taxa were filtered to keep bacteria in analysis; viruses and archaea were excluded. Nine hundred and fifty taxa present in at least 20% of samples were kept for statistical testing. Within each tumor type, Spearman’s correlation was computed between each bacteria taxon and ACE2 or TMPRSS2 gene expression in tumor and normal tissues. For each test, at least 15 samples with taxon abundance ≥1 were required. Seventy-five taxa passed FDR-adjusted p<0.05 and Spearman’s ρ>0.5 or <−0.5 in at least one pairwise correlation (online supplementary table 11).

Least Absolute Shrinkage and Selection Operator (LASSO) regression models of ACE2 or TMPRSS2 gene expression were built in tumor (n=10,024) and normal (n=708) samples separately using 90 features consisting of tissue type as well as clinical (age, gender, race), immune signatures (ISG, T cell-inflamed, myeloid, angiogenesis, TGF-β), immune cell subsets (macrophage M1, macrophage M2, CD8 T cell, CD4 T cell), stroma cell subset (epithelial cell), and 75 microbiota features from microbiota correlation analysis (online supplementary figure 2). Macrophage M1/M2 and T cells were included in the model based on emerging evidence suggesting an important role of macrophage and proinflammatory phenotype39 in COVID-19 disease. Thirty-four tumor types were collapsed into 15 tissue types based on categorizations from The Human Protein Atlas to reduce complexity. Categorical variables were converted to dummy variables using R function dummyVars with parameter fullRank set to TRUE. Data were preprocessed to remove features that have near-zero variance, high correlation (Spearman’s ρ>0.75), or high collinearity. Each feature was scaled and centered. Given the purpose of this analysis was to evaluate the relative importance of features rather than training and validating a predictive model, we did not split samples into training and test sets. Instead, we used all samples with 10-fold cross-validation. Variable importance was reported as raw values and as scaled values to 0–100 (online supplementary table 12). R package caret (version 6.0–84) was used for analysis.

In all analyses, a minimal sample size of 15 per group was required for statistical testing. For group-wise comparisons, two-way ANOVA was used with tumor type and group of interest as variables plus the interaction. When more than two groups are present, Tukey’s HSD test was used for pairwise comparisons. Within each tumor type, two-sided Welch Two Sample t-test was used to compare log2-transformed gene expression between groups; for paired samples, two-sided paired t-test was used. Spearman’s correlation was used to determine the relationship between two continuous variables. For multiple comparisons, p values were corrected using BH-FDR method. LASSO regression models were used to evaluate variable importance with 10-fold cross-validation. Those analyses were performed using R (version 3.6.1) and Bioconductor (release 3.10). FDR-adjusted p values less than 0.05 were considered statistically significant.

Given the association between COVID-19 disease and various organ-specific symptoms, we investigated the distribution of ACE2 and TMPRSS2 expression in tumor and normal tissues across 34 tumor types consisting of 15 tissue types (online supplementary tables 1–3). None of the LAML samples express TMPRSS2; hence, 33 tumor types were used for the analysis of TMPRSS2.

When ranked by expression percentile in each tumor type, cholangiocarcinoma (CHOL), colon adenocarcinoma (COAD), PAAD, rectum adenocarcinoma (READ), and STAD are among the top 25% percentile for both genes, all of which are digestive organs (figure 1A). We acknowledge that TMPRSS2 is highly expressed in prostate adenocarcinoma (PRAD) likely due to known TMPRSS2:ERG gene fusion overexpression.40 ACE2 and TMPRSS2 expression showed weak positive correlation in normal (Spearman’s ρ=0.28, FDR-adjusted p<0.0001) and in tumor (ρ=0.30, FDR-adjusted p<0.0001) (figure 1B). Correlation within each tumor type showed a consistent pattern and can be further explored in external data files available on GitHub (https://github.com/riyuebao/ACE2_TMPRSS2_multicorrelates).

In 14 tumor types where 15 or more matched normal tissues are available (online supplementary table 1), we compared gene expression between tumor and matched normal from the same patients. The expression of ACE2 and TMPRSS2 was significantly higher in normal tissues relative to tumors (figure 1C). Within individual tumor types, this pattern was significant for ACE2 in breast invasive carcinoma (BRCA), COAD, chromophobe kidney cancer (KICH), liver hepatocellular carcinoma (LIHC), PRAD, thyroid cancer (THCA), and for TMPRSS2 in eight tumor types (FDR-adjusted p<0.05 and higher in normal). Three tumor types showed elevated levels for both genes in normal tissues compared with tumors (BRCA, COAD, and LIHC) (figure 1D).

Several clinical observations on COVID-19 indicated clinical factors such as BMI might be associated with severity.5 28 We sought to investigate the association between clinical variables (age, gender, race, tumor stage, tumor grade, menopause, BMI, smoking history) and ACE2 or TMPRSS2 gene expression in tumor or normal tissues from 10,038 patients. For each clinical factor, data were fitted into a two-way ANOVA model with tumor type and clinical factor as variables plus the interaction between the two. Overall, we did not observe significant differences in expression for either gene when comparing designated clinical groups within tumor or normal tissues (figure 2A–F) (online supplementary tables 4 and 5). We observed sporadic clinical groups within individual tumor types that do show significance (eg, age in ESCA for TMPRSS2) (figure 2G) though are of unclear clinical relevance at this time (online supplementary tables 6 and 7). The results suggested those clinical variables are not strongly associated with ACE2 or TMPRSS2 expression. In addition, we investigated the association between the presence of diabetes and gene expression in PAAD, where data were available. Similarly, no significant differences in gene expression were detected in tumor or normal tissues from patients with or without diabetes.

Two HLA genotypes (B*46:01, B*54:01) have been reported to be associated with severe clinical outcomes by other groups.41 We investigated the prevalence of the two alleles and identified low prevalence across all tumor types (0.6% and 0.2%, respectively, out of 19,118 HLA-B alleles from 9,559 patients). When looking into individual tumor types, both alleles were found to be significantly enriched in LIHC (FDR-adjusted p<0.0001), which was likely due to the enrichment of Asian populations in this cohort (43%). Comparison of ACE2 or TMPRSS2 gene expression from tumor or normal samples between HLA genotypes showed no significant differences after adjusting for tumor-type specific gene expression, and can be further explored in external data files available on GitHub.

To understand potential associations of ACE2 and TMPRSS2 in tissues relevant to patients being treated with cancer immunotherapy, we investigated the correlation between ACE2 or TMPRSS2 and immune gene expression signatures known to be relevant in immuno-oncology (online supplementary tables 8 and 9). ACE2 was positively correlated with ISG signature in tumor samples from 24/34 tumor types (71%, 14 reached FDR-adjusted p<0.05) and normal tissues from 10/14 tumor types (71%, 4 reached FDR-adjusted p<0.05) and negatively correlated with angiogenesis and TGF-β signatures (figure 3A). TMPRSS2 showed a mixed pattern of correlations with those immune responsive or suppressive signatures (figure 3B).

We did not find a consistent pattern of ACE2 with cell subsets across all tumor types with the exception of a high positive correlation with macrophage M2 in normal tissues from kidney cancers (kidney renal clear cell carcinoma (KIRC), KICH, kidney renal papillary cell carcinoma (KIRP)) (Spearman’s ρ=0.84, 0.82, 0.79, FDR-adjusted p<0.0001) and STAD (ρ=0.67, FDR-adjusted p<0.001) (figure 3C) (online supplementary table 10). TMPRSS2 was positively correlated with epithelial cell abundance in tumor samples from 29/33 tumor types (88%, 17 reached FDR-adjusted p<0.05) and normal tissues from 14/14 tumor types (100%, 9 reached FDR-adjusted p<0.05) (figure 3D).

With the observation of correlation with specific immune signatures, we sought to investigate whether ACE2 or TMPRSS2 were expressed directly by immune cells or other specific cell types in nine independent single-cell RNAseq studies. This includes tumor and immune cells from patients with cancer diagnosed with glioblastoma or melanoma (Single Cell Portal, Broad Institute), or head and neck cancer.42 In the cohorts of patients with cancer, less than 1% of the malignant cells express ACE2 and/or TMPRSS2, while few to none of the immune cells express either gene. An exploration of three studies focusing on immune cells from healthy donors (13,316 peripheral blood mononuclear cells (Immune Cell Atlas), 39,563 ileum lamina propria immunocytes (Immune Cell Atlas), 594,857 immune cells (Census of Immune Cells, EBI)) further confirmed the lack of ACE2 and TMPRSS2 expression in immune cell populations. Realizing that other studies had reported the genes as highly expressed in lung, heart, brain, and colon, we investigated published large-scale profiling of 702,968 single cells from patients with non-cancer or healthy donors (Human Cell Landscape).43 We found TMPRSS2 was expressed in stomach, colon, kidney, prostate, intestine, jejunum, pancreatic, esophagus, and bladder tissues, while ACE2 was only expressed in jejunum and fetal intestine. Last, we examined a public scRNAseq cohort from patients with COVID-19 infection (https://doi.org/10.5281/zenodo.3747336) and confirmed that ACE2 was not expressed in immune cells, and TMPRSS2 was present in 6 out of 140,956 cells total by one read count. Therefore, we concluded that ACE2 or TMPRSS2 is not expressed in immune cell populations, at least in the cohorts investigated. The expression of both genes in bulk RNAseq data was likely to be derived from non-immune cells, such as epithelial cells in the tissues.

Given associations between strong antitumor immune responses due to the presence of tumor-related virus and particular commensal microbiota, we sought to investigate associations between ACE2 and TMPRSS2 gene expression with the presence of virus or tissue microbiota. Across known viral positive and negative tumor types, we found an inconsistent pattern relative to viral associated versus non-viral associated tumors (HPV, EBV, HBV). We then correlated 1,093 commensal microbiota identified in tumor and normal tissue RNAseq data, as published in Ref. 35 with gene expression. We identified 75 taxa that showed significant and strong positive or negative correlation (Spearman’s ρ>0.5 or <−0.5) with either gene in at least one pairwise correlation (FDR-adjusted p<0.05) (figure 4A and B) (online supplementary table 11). COAD and STAD were the two tumor types demonstrating the strongest and most prevalent positive correlation of ACE2 and TMPRSS2 gene expression with abundance of specific bacteria taxa, respectively. Kidney cancers, including KIRP, KICH, and KIRC, also showed positive correlations between ACE2 and microbiota if not as prevalent as colon or stomach cancers. If ranked by Spearman’s correlation coefficient, Chlamydia was the top microbiota positively correlated with ACE2 in COAD (ρ=0.81, FDR-adjusted p<0.0001) (online supplementary table 11). For both genes, those patterns were less prominent in tumor samples, which could be due to high heterogeneity in tumors. Overall, we observed approximately a 2.6:1 ratio of commensal microbiota for gram-negative to positive groups in the 75 taxa (50 negatives, 19 positives, others undetermined) (online supplementary table 11).

To integrate all correlates and evaluate their relative importance in determining the gene expression of ACE2 and TMPRSS2, we built LASSO regression models in tumor and normal tissues separately using features from the clinical, immune, and microbial domains (online supplementary figure 2). Clinical features included were age, gender, and race, while menopause, BMI, and smoking history were excluded because >50% of the samples were missing information. HLA genotype was not included because of many categories and/or levels, which may lead to overfitting. Immune gene expression signatures included ISG, T cell-inflamed, myeloid, angiogenesis, and TGF-β. Immune cell type features included macrophage M1/M2, CD8, and CD4 T cells, and non-immune cell type features included epithelial cells. Microbe features included the 75 bacteria taxa from the analysis above. In addition, we collapsed 34 tumor types into 15 tissue types and included these in the model to account for tissue-specific gene expression variations.

We calculated the importance of each feature in the models with 10-fold cross-validation (online supplementary table 12). After quality control and filtering, among the features kept in each model, immune and epithelial cells were the top-ranked features that predict ACE2 expression in normal tissues and tumors. Microbiota were observed to be important features for ACE2 in normal tissues but not in tumors (figure 5A and B). For TMPRSS2 expression, epithelial cell abundance is the most important predictor in both normal and tumor samples (figure 5C and D). Taken together, these results suggested that immune signatures, epithelial cells, and commensal microbiota were important predictors for ACE2 expression, while TMPRSS2 expression was primarily determined by epithelial cells.