Single-cell analysis identified POSTN⁺ cells associated with the aggressive phenotype and risk of esophageal squamous cell carcinoma

Summary

Tumors are intricate and heterogeneous systems characterized by mosaic cancer cell populations with diverse expression profiles. Leveraging single-cell technologies, we employed the Scissor algorithm to delineate an epithelial subpopulation associated with the aggressive phenotype in esophageal squamous cell carcinoma (ESCC). This identified subpopulation exhibited elevated expression of genes involved in critical pathways, such as epithelial-mesenchymal transition and PI3K-Akt. Key signature genes within this subpopulation, namely CAV1, COL3A1, COL6A1, POSTN, and TAGLN, demonstrated significant upregulation concomitant with both tumorigenesis and tumor progression across independent single-cell datasets. Furthermore, we selected 1,450 expression quantitative trait loci of the top 62 signature genes of this cell subpopulation to investigate their potential in predicting ESCC risk. The results showed that the POSTN loci were predominantly associated with ESCC susceptibility. Through functional annotation and replication analyses, we identified that the rs1028728 in the POSTN promoter was significantly associated with increased ESCC risk in 7,049 ESCC cases and 8,063 controls (odds ratio = 1.29, 95% confidence interval: 1.18–1.42, p = 4.03 × 10⁻⁸). Subsequent biochemical experiments showed that the rs1028728[T] allele enhanced POSTN expression by affecting the binding of PRRX1 in the POSTN promoter. In summary, our meticulous single-cell analysis delineates an invasive epithelial subpopulation in ESCC, with POSTN emerging as an important marker for the aggressive phenotype. These findings offer more insights into potential strategies for the prevention and intervention of ESCC, enriching our understanding of this complex cancer landscape.

Graphical abstract

Integrating single-cell and bulk RNA sequencing data from various esophageal squamous cell carcinoma (ESCC) cohorts, we pinpointed that POSTN emerges as a significant marker linked to the aggressive phenotypes of epithelial cells and the susceptibility to ESCC. These findings provide a more profound comprehension of the ESCC development.

Main text

Esophageal squamous cell carcinoma (ESCC [MIM: 133239]) poses a significant health challenge, characterized as a highly lethal form of cancer with a mere 20–30% 5-year survival rate. It primarily manifests in the middle or upper section of the esophagus, constituting over 90% of all esophageal cancer cases in China.^1,2 Typically originating from the lining of esophageal squamous epithelium, ESCC development is triggered by exposure to carcinogens or mechanical damage, instigating abnormal proliferation of epithelial cells and ultimately progression to invasive cancer.³ Understanding the molecular mechanisms underpinning epithelial cell progression is crucial for identifying molecular markers and intervention targets in ESCC.

Traditionally, the search for cancer biomarkers has predominantly relied on bulk RNA sequencing (RNA-seq).⁴ However, the limitation of this approach in capturing signals from rare cell populations or specific cell types driving tumorigenesis and progression has prompted the adoption of single-cell RNA-seq (scRNA-seq),⁴ which is widely applied in investigating tumor pathogenesis, including ESCC.^2,5 It offers exceptional sensitivity and specificity⁶ and the ability to discern diverse cell types, states, and cell subpopulations within heterogeneous tissue.^7,8,9 Multiple single-cell algorithms have been developed to identify precancerous cells¹⁰ and those associated with specific cancer phenotypes, aiming to explore yet-undetermined genetic determinants and identify cell types and processes influenced by genetic and gene variations. Especially, an algorithm known as Scissor leverages bulk data and phenotype information to steer single-cell analysis and identifies subsets of cells that significantly contribute to phenotypic disparities and disease relevance.¹¹ Additionally, a previous study defined a confused cell identity (CCI) in ESCC at single-cell resolution, establishing its potential as a malignant cellular characteristic and an independent prognostic marker.¹²

The genetic etiology of ESCC remains largely unknown. Previous studies have extensively investigated the genomic characteristics and alterations in ESCC and normal esophageal tissues and across various stages of normal esophagus, early and late precancerous lesions, and esophageal cancer.^13,14 The major genomic changes observed in cancer genomes include more frequent chromosomal instability and copy-number alterations (CNAs) in ESCC compared to normal esophageal tissues, with the burden of somatic mutations and CNAs increasing with the progression of the disease. Several genome-wide association studies (GWASs) of germline variants in ESCC have been published. GWASs have discovered over two dozen risk loci for ESCC.¹⁵ Integrating scRNA-seq data with GWAS data can unveil pivotal genes vital for disease onset and progression.

Understanding the gene expression pattern and expression programs in invasive epithelial cells is crucial for unraveling the mechanisms of invasive carcinogenesis. In this study, we identified an aggressive tumor phenotype-related cell subpopulation and validated its heightened malignancy across multiple datasets and methods. Through case-control studies and multi-stage ESCC scRNA-seq data, we identified the signature gene POSTN (MIM: 608777) as a crucial regulator of ESCC susceptibility and its invasive phenotype. Notably, the rs1028728 variant within the POSTN promoter region emerged as a potential mechanism driving heightened ESCC invasion and susceptibility. These findings constitute a valuable resource for a deeper understanding of the intricate progression of epithelial cell invasion in ESCC.

To investigate the cellular subpopulations and biomarkers associated with the aggressive phenotype of ESCC, we integrated a bulk expression matrix and single-cell expression matrix using Scissor. The flowchart of this study can be found in Figure S1. The RNA-seq data for 174 ESCC tumors and 105 corresponding nontumor tissues were sourced from our prior study¹⁶ and The Cancer Genome Atlas (TCGA) database. To integrate these two datasets into a unified expression matrix, we utilized the R package “sva” to mitigate batch effects in the combined samples (Figure S2). scRNA-seq data of 44,122 epithelial cells from 60 ESCC tumors were obtained from the GEO database² (Figure S3A). To stratify cell subpopulations associated with aggressive tumor phenotypes, we incorporated single-cell expression matrices into the Scissor algorithm. Concurrently, bulk RNA expression data from ESCC and matched normal samples were utilized to establish benchmarks for the aggressive tumor and normal phenotypes.

Using Scissor with various alpha parameter choices, we identified varying proportions of aggressive tumor phenotype-related epithelial cells（defined as POSTN⁺ cells）, all exhibiting consistent reliability and significance (Figure 1A). We then introduced a novel feature, the CCI score, as reported in a previous study,¹² to validate the malignancy of the identified cell subpopulation of Scissor. We observed that the CCI scores of POSTN⁺ cells were significantly higher than those of other cells under alpha = 0.001 (Figures 1B and S3B), confirming that the selected invasive subpopulation exhibited a higher degree of malignancy. A previous scRNA-seq analysis identified three principal cell populations constituting the normal esophageal squamous epithelium: postmitotic differentiated keratinocytes (DKs), transient proliferating basal keratinocytes (BKs), and basal stem cells (BSs).^17,18,19 We showed a ternary plot positioning BSs, BKs, and DKs at the vertices, allowing for the representation of POSTN⁺ cell distribution in relation to the established populations of normal esophageal squamous epithelial cells. Consistent with previous results, this identified aggressive subpopulation almost fell in the center of the triangle, suggesting they may represent tumor cells acquiring a new identity with partial similarities to these three normal cell types¹² (Figure 1C). As a result, we identified a cell subpopulation associated with aggressive tumor phenotype, denoted as the POSTN⁺ subpopulation, comprising 15,493 epithelial cells. This subpopulation passed a highly reliable significance test with a test statistic of 0.980 and a p value of less than 0.001, accounting for 35.11% of the total epithelial cells (Figures 1A, 1D, and 1E). Prior scRNA-seq analysis of 60 patients with ESCC identified eight expression programs, each associated with distinct cellular functions and states. Cells that express at least 70% of a program’s gene set are deemed “activated” within that program.² To investigate the primary roles of POSTN⁺ cells, we consequently computed the percentage of POSTN⁺ cells within these program cells. As seen in Figure 1E, the POSTN⁺ subpopulation predominantly exhibited features of the epithelial differentiation 1 (Epi1), the cell-cycle (Cycling) program features, and the mesenchymal stromal cell-like qualities (Mes). These findings deepen our understanding of the intricate landscape of tumor cells in ESCC.

Figure 1.

Identification of cell subpopulation associated with aggressive phenotype in ESCC

(A) The bar chart and line graph, respectively, present the proportions of POSTN⁺ cells selected by Scissor, as well as the values of significance test statistics, under different alpha parameters from 0.0001 to 0.1.

(B) CCI scores of all epithelial cells at 0.001 alpha parameters. The figure on the bottom compares the CCI scores of the POSTN⁺ cells with the rest of the cells by the Mann-Whitney test. ∗∗∗∗p < 0.0001.

(C) The ternary diagram shows the similarity and variation in scRNA-seq data of POSTN⁺ cells, with the basal stem cells (BSs), transient proliferating basal keratinocytes (BKs), and postmitotic differentiated keratinocytes (DK) cells. The color represents the density levels of distribution.

(D) The uniform manifold approximation and projection (UMAP) visualization of the Scissor-selected cells. The red dots are cells associated with the aggressive phenotypes (alpha = 0.001).

(E) The percentage of POSTN⁺ cells in the eight different programs. The percentages of invasive phenotype cells in each program and within epithelial cells were compared using a chi-squared test, ∗∗∗∗p < 0.0001.

(F) GO, KEGG, and HALLMARK enrichment analyses were used to elucidate the potential functions of 62 signature genes in POSTN⁺ cells.

(G) The UMAP plot maps the aggressive phenotype epithelial score for each cell.

(H) Violin plots of the aggressive phenotype epithelial scores in different stages are plotted. ∗∗∗∗p < 0.0001 was calculated by the Mann-Whitney test.

(I) Violin plots of aggressive phenotypic epithelial scores for single-cell validation data of ESCC progression. The aggressive phenotype epithelial scores of cells in the HGIN and ESCC stages are significantly higher than those in the normal (NOR) stage. ∗∗p < 0.01 and ∗∗∗p < 0.001 were calculated by independent t test.

We then sought to determine whether signature genes within the POSTN⁺ subpopulation could serve as molecular markers for the occurrence and progression of ESCC. Through differential gene expression analysis between the POSTN⁺ subpopulation and other cells across eight expression programs, we identified 62 signature gene candidates and aggregated them into a gene set for computing the epithelial score associated with an aggressive phenotype for validation (Table S1). Notably, these genes were enriched in crucial cancer-associated pathways including epithelial-mesenchymal transition (19 genes), PI3K-Akt (7 genes), and proteoglycans in cancer (6 genes) (Figure 1F; Table S2). The aggressive phenotype epithelial score can characterize the features of POSTN⁺ cells, with its distribution consistent with the clustering distribution of POSTN⁺ cells, exhibiting significant differences between POSTN⁺ cells and other cells (Figures 1G and S3C). The aggressive phenotype epithelial scores also exhibited a gradual increase along with the tumor stage (Figures 1H and S4). To validate transcriptional changes in POSTN⁺ cells at various disease stages, we obtained a single-cell expression matrix from a validation cohort.⁵ Our analysis revealed that the expression levels of the part of these 62 signature genes exhibited significant elevation in tumor cells or at late precancerous stage (high-grade intraepithelial neoplasia [HGIN]), with no corresponding increase in normal or early precancerous stage (low-grade intraepithelial neoplasia) (Figure S5). Specifically, the aggressive phenotype epithelial cell score exhibited a notable increase with tumor advancement in both ESCC and HGIN stages (Figure 1I). Moreover, the expression of these genes demonstrated a discernible upregulation concomitant with tumor progression, especially for CAV1 (MIM: 601047), POSTN, COL3A1 (MIM: 120180), COL6A3 (MIM: 120250), and TAGLN (MIM: 600818) (Figures S4 and S5). Several identified signature gene candidates, such as TNC,²⁰ COL3A1,²¹ and COL6A3,²² have been widely implicated in cancer invasion. Caveolin-1 (CAV1), recognized as an oncogenic protein, is linked to tumor progression, invasive behavior, metastasis, and treatment resistance, including in ESCC.²³ We found that most of the signature genes associated with the aggressive phenotype were those traditionally thought to characterize fibroblasts. These genes exhibited high expression levels in invasive subpopulations of epithelial cells and either exclusively expressed or had elevated expression in HGIN and ESCC during the invasive progression of ESCC.

Genes identified by scRNA-seq data as pivotal in disease onset and progression may also contribute to disease susceptibility.²⁴ Consequently, we assessed whether these identified 62 signature genes in the POSTN⁺ subpopulation could confer risk of ESCC. We utilized two GWAS datasets in Chinese populations, consisting of 3,929 individuals with ESCC and 4,144 controls,^25,26 with basic demographic information presented in Table S3. A total of 1,450 expression quantitative trait loci of the 62 signature genes were genotyped, revealing that the POSTN loci stood out as the most significant signal with a false discovery rate <0.05 (Figure 2A). POSTN was predominantly expressed in the Mes program, constituting 64.6% of the POSTN⁺ subpopulation, significantly higher than the proportion of the malignant invasive subpopulation in all epithelial cells (p < 0.0001) (Figure 1E). The Mes program comprised a total of 2,743 epithelial cells featuring genes like VIM (MIM: 193060) and SPARC (MIM: 182120) and displayed activation of the epithelial-mesenchymal transition and angiogenic pathways.² Clustering of the POSTN⁺ subpopulation with cells in the Mes program is depicted in Figure 2B, and the difference in cell identity between the CCI scores of the POSTN⁺ subpopulation and those of the other cells confounded within the Mes program is highlighted in Figure 2C. Moreover, the POSTN⁺ subpopulations in Mes program cells are located in the middle of the ternary phase diagram (Figure 2D), indicating a potential malignant profile for these cells.

Figure 2.

POSTN plays a vital role as a regulatory factor in ESCC susceptibility and its aggressive phenotype

(A) Genome-wide results for the relationship between 62 signature gene candidate expression quantitative trait locus (eQTL) SNPs and ESCC. Manhattan plot shows that POSTN was the susceptibility gene with the highest association with ESCC. p < 5 × 10⁻⁴ (false discovery rate [FDR] < 0.05) was considered statistically significant, and rs1028728 was annotated.

(B) The UMAP visualization of the Mes cells.

(C) CCI scores of all cells of Mes at 0.001 alpha parameters tested by the Mann-Whitney test. ∗∗∗∗p < 0.0001.

(D) The ternary diagram shows the similarity and variation in scRNA-seq data of POSTN⁺ cells within Mes.

(E) Violin plots of mean POSTN expression in POSTN⁺ cells within Mes, other cells within Mes, and Mes are plotted for different stages. ∗∗p < 0.01 compared to controls by Mann-Whitney test.

(F) POSTN is significantly overexpressed in tumor tissues compared to normal tissues from multiple independent databases, including Chang et al.¹⁶, GEO: GEO53625, and TCGA ESCC samples. ∗∗∗∗p < 0.0001 was calculated by a Wilcoxon test in Chang et al. ESCC tissues and GEO: GEO53625. ∗∗p < 0.01 was calculated by Student’s t test in TCGA.

POSTN also emerged as a potential key gene in the progression of ESCC, with significantly increased expression levels across disease stages, especially in the POSTN⁺ subpopulation (Figure 2E). The expression of POSTN was also significantly upregulated in ESCC tumors in three independent bulk RNA-seq datasets compared to normal samples (Figure 2F). The dynamic interplay between the epithelium and mesenchymal stroma is crucial in triggering an invasive tumor cell phenotype, activating genes linked to cell proliferation, dedifferentiation, migration, and invasion.^27,28 POSTN, functioning as a nonstructural extracellular matrix protein, facilitates cell-matrix interactions.²⁹ During tumor progression, ECM proteins within the tumor stroma can establish a microenvironment conducive to growth and dissemination.³⁰ Previous research has identified POSTN as a crucial mediator of ESCC tumor invasion in invading EPC2-hTERT-EGFR-p53R175H cells.³¹ Functional validation in TE11 and HCE4 ESCC cell lines confirmed POSTN’s role in promoting tumor growth and invasion in ESCC.³²

Due to the commonality in the development process of SCC, there are also similarities in gene expression and mutation profiles. We further evaluated the expression of POSTN in head and neck squamous carcinoma (HNSC [MIM: 275355]), lung squamous carcinoma (LUSC [MIM: 211980]), and cervical SCC (MIM: 603956) in TCGA database. The results showed that the POSTN exhibited elevated expression levels in HNSC and LUSC, with expression levels increasing as the cancer progressed in HNSC (Figures S6A–S6C), suggesting that POSTN may be an invasive characteristic for SCCs. SCCs’ single-cell data in HNSC, LUSC, and oropharyngeal SCC further reveal that, aside from fibroblasts, POSTN is expressed in tumor cells as well (Figures S6D–S6F). This aligns with Chao Liu et al.’s investigation into the onset and progression of human cervical SCC through single-cell sequencing.³³ Their study revealed heightened expression of POSTN within the Epi1 epithelial cell subset.³³ In particular, Luhai Wang’s team in 2022 found that POSTN plays a vital role in ovarian tumor metastasis by enriching M2 macrophages and cancer associated fibroblasts (CAFs) through an autocrine effect in cancer cells.³⁴ These findings suggest that the POSTN expressed in epithelial cells may be a crucial molecular marker for ESCC tumorigenesis and progression.

We further conducted a fine-mapping analysis to identify functional variants influencing POSTN expression within this locus. Through bioinformatics annotation, we identified the rs1028728 in the POSTN promoter, exhibiting significant enrichment of H3K4me1 modifications and displaying the most promising potential to affect POSTN expression (Figure 3A). Subsequently, a significant correlation was observed between this variant and POSTN expression (p < 0.05) in 497 normal esophageal mucosa tissues from the GTEx cohort (Figure 3B). In the discovery stage, individuals carrying the risk allele of rs1028728 exhibited a 29% increased risk of ESCC (95% confidence interval [CI] 1.14–1.47, p = 8.64 × 10⁻⁵), and the association was successfully replicated in 3,120 ESCC cases and 3,919 controls³⁵ (odds ratio [OR] = 1.26, 95% CI 1.10–1.45, p = 9.88 × 10⁻⁴) (Table 1). In the combined samples, the association between rs1028728 and ESCC risk reached the genome-wide association significance level (OR = 1.29, 95% CI = 1.18–1.42, p = 4.03 × 10⁻⁸) (Table 1).

Figure 3.

The rs1028728 variant exhibited an allele-specific effect on POSTN expression through binding with PRRX1

(A) Integrative analysis of the potential function of rs1028728 by querying UCSC. The SNP rs1028728 was in the POSTN promoter region.

(B) The violin plots of expression correlation between rs1028728 and POSTN through the GTEx v.8 eQTL Calculator.

(C) Relative reporter gene activity of plasmids containing the rs1028728[A] or -[T] allele in KYSE30 and KYSE150 cells.

(D) The rs1028728[T] allele resides within a PRRX1 binding motif.

(E) Detection of PRRX1 mRNA by RT-qPCR in KYSE30 and KYSE150 cells, respectively. After cell seeding in a 6-well plate, transfections were done with small interfering RNA (siRNA) targeting PRRX1 or siControl. ∗∗∗∗p < 0.0001 was calculated using a Mann-Whitney test in KYSE30, whereas it was calculated using a two-tailed Student’s t test in KYSE150.

(F) The effect of PRRX1 knockdown on the relative luciferase activity of constructs containing the rs1028728[A] or rs1028728[T] allele in KYSE30 and KYSE150 cells. ∗∗∗∗p < 0.0001 values were calculated using a two-tailed Student’s t test.

(G) The correlation between POSTN expression and PRRX1 expression was calculated in normal esophageal mucosa from GTEx, GEO: GSE53625 dataset, Chang et al.¹⁶, ESCC cohorts, and TCGA cohorts. All p and r values were using Pearson’s correlation analysis.

(H) Correlation analysis of POSTN and PRRX1 expression in Mes POSTN⁺ cells was conducted, and the scatterplot was generated with corresponding r and p values, following the aforementioned methodology. UMAP plots were employed to visually represent the clustering of Mes cells and the expression patterns of POSTN and PRRX1.

Table 1.

The association between rs1028728 and the susceptibility to esophageal squamous carcinoma in Chinese populations

Stage	rs1028728	Cases (%)	Controls (%)	OR (95% CI)a	p
Discovery	AA	3,376 (85.9)	3,678 (89.0)	1.00 (Ref)	N/A
Discovery	AT	537 (13.7)	443 (10.7)	1.32 (1.15–1.51)	6.41 × 10−5
Discovery	TT	16 (0.4)	14 (0.3)	1.23 (0.60–2.55)	0.5795
Discovery	dominant	N/A	N/A	1.31 (1.15–1.50)	5.79 × 10−5
Discovery	recessive	N/A	N/A	1.19 (0.58–2.47)	0.6407
Discovery	additive	N/A	N/A	1.29 (1.14–1.47)	8.64 × 10−5
Replication	AA	2,667 (85.5)	3,461 (88.3)	1.00 (Ref)	N/A
Replication	AT	432 (13.8)	446 (11.4)	1.24 (1.07–1.44)	4.25 × 10−3
Replication	TT	21 (0.7)	12 (0.3)	1.95 (0.95–4.22)	0.0756
Replication	dominant	N/A	N/A	1.26 (1.09–1.46)	1.78 × 10−3
Replication	recessive	N/A	N/A	1.91 (0.93–4.11)	0.0865
Replication	additive	N/A	N/A	1.26 (1.10–1.45)	9.88 × 10−4
Combined samples	AA	6,043 (85.7)	7,148 (88.7)	1.00 (Refref)	N/A
Combined samples	AT	969 (13.8)	889 (11.0)	1.30 (1.16–1.43)	2.24 × 10−7
Combined samples	TT	37 (0.5)	26 (0.3)	1.65 (1.00–2.77)	0.0513
Combined samples	dominant	N/A	N/A	1.31 (1.19–1.44)	6.14 × 10−8
Combined samples	recessive	N/A	N/A	1.60 (0.97–2.68)	0.0681
Combined samples	additive	N/A	N/A	1.29 (1.18–1.42)	4.03 × 10−8

OR, odds ratio; CI, confidence interval; Ref, reference; N/A, not applicable.

^aCalculated by unconditional logistic regression model after adjusting for gender and age.

To test whether the rs1028728 variant influences the promoter activity of POSTN, we conducted luciferase reporter gene assays on KYSE30 and KYSE150 cells. The results revealed that the construct harboring the POSTN promoter with the rs1028728[A] allele displayed markedly lower luciferase expression than the construct containing the rs10287228[T] allele (Figure 3C). We further explored whether the rs1028728 variant affected POSTN expression by altering the binding of a transcription factor. Through motif enrichment analysis, we identified that the rs1028728[T] allele may preferentially bind to parental homology box1 (PRRX1 [MIM: 167420]), a key inducer of epithelial-mesenchymal transition^36,37 (Figure 3D). We then knocked down the PRRX1 expression by using two different small interfering RNAs (Figure 3E) and found a significant decrease in luciferase expression of the construct harboring the POSTN promotor containing the rs10287228[T] allele compared to the control vector (Figure 3F). Conversely, the luciferase expression of the construct harboring the POSTN promotor containing the rs10287228[A] allele remained largely unaffected upon PRRX1 knockdown (Figure 3F). Moreover, significant positive correlations were observed between the expression of PRRX1 and POSTN in both bulk RNA-seq and scRNA-seq datasets (Figures 3G and 3H). Collectively, these findings indicate that the risk allele T of rs1028728 enhances the binding capability of PRRX1, thereby regulating the high expression of POSTN. The transcription factor PRRX1 may be involved in promoting epithelial-mesenchymal transition through regulating POSTN.

This study employed single-cell integrated analysis to identify epithelial cells and characteristic genes associated with the invasive phenotype of ESCC, overcoming the limitations of bulk RNA-seq. The findings were robust and reliable, validated through multiple databases and diverse methodologies. Additionally, we conducted population and mechanistic studies, leading to the discovery of novel invasive biomarkers. Our study possesses certain limitations. The designation of rs1028728 as an ESCC risk SNP in populations beyond the Chinese demographic necessitates additional confirmation. Furthermore, the mechanism of action of POSTN and PRRX1 in epithelial cells and its role in the interaction between epithelial cells and CAFs warrant further investigation. In conclusion, through single-cell sequencing analysis, bulk RNA-seq analysis, epidemiological studies, and a series of biochemical experiments, our findings advance our comprehension of the molecular mechanisms underpinning POSTN and offer valuable insights for future research endeavors.

Data and code availability

The published article includes all datasets generated or analyzed during this study.

Web resources

• Gene Expression Omnibus, https://www.ncbi.nlm.nih.gov/geo/.

• Genotype-Tissue Expression, http://gtexportal.org.

• GSA-Human, https://bigd.big.ac.cn/gsa-human.

• HaploReg, v.4.1, http://compbio.mit.edu/HaploReg.

• Jasper website prediction, https://jaspar.genereg.net/.

• RegulomeDB, v.2.2, https://regulomedb.org/regulome-search/.

• TCGA database, http://cancergenome.nih.gov.

• Tumor Immune Single-cell Hub 2 (TISCH2), http://tisch.comp-genomics.org/home/.

• UCSC Genome Browser, http://genome.ucsc.edu/.