Pathway analysis of renal cell carcinoma genome-wide association studies identifies novel associations

Abstract

Background:

Much of the heritable risk of renal cell carcinoma (RCC) associated with common genetic variation is unexplained. New analytic approaches have been developed to increase the discovery of risk variants in genome-wide association studies (GWAS), including multi-locus testing through pathway analysis.

Methods:

We conducted a pathway analysis using GWAS summary data from six previous scans (10,784 cases and 20,406 controls) and evaluated 3,678 pathways and gene sets drawn from the Molecular Signatures Database. To replicate findings, we analyzed GWAS summary data from the UK Biobank (903 cases and 451,361 controls) and the Genetic Epidemiology Research on Adult Health and Aging cohort (317 cases and 50,511 controls).

Results:

We identified 14 pathways / gene sets associated with RCC in both the discovery (P < 1.36 × 10⁻⁵, the Bonferroni correction threshold) and replication (P < 0.05) sets, 10 of which include components of the PI3K/AKT pathway. In tests across 2,035 genes in these pathways, associations (Bonferroni-corrected P < 2.46 × 10⁻⁵ in discovery and replication sets combined) were observed for CASP9, TIPIN and CDKN2C. The strongest SNP signal was for rs12124078 (P_Discovery = 2.6 × 10⁻⁵, P_Replication = 1.5 × 10⁻⁴, P_Combined = 6.9 × 10⁻⁸), a CASP9 expression quantitative trait locus.

Conclusions:

Our pathway analysis implicates genetic variation within the PI3K/AKT pathway as a source of RCC heritability and identifies several promising novel susceptibility genes, including CASP9, which warrant further investigation.

Impact:

Our findings illustrate the value of pathway analysis as a complementary approach to analyzing GWAS data.

INTRODUCTION

Kidney cancer is one of the ten most common cancers in the United States, with around 74,000 new cases and 14,000 related deaths in 2019 (1). Renal cell carcinoma (RCC) is the most common type of kidney cancer, accounting for over 90% of kidney cancer diagnoses. RCC has a heritable basis, with relatives of patients having a two-fold increased risk (2, 3). While a number of rare familial RCC syndromes caused by inheritance of high-impact mutations have been identified, even collectively, they only account for less than 5% of RCC (4). Evidence for the role of common low-impact genetic variants influencing RCC risk has been established in genome-wide association studies (GWAS), which have so far identified 13 susceptibility loci (5).

Since GWAS risk variants typically have a small effect size, they are difficult to identify in individual SNP-based GWAS after accounting for multiple testing, even with large sample numbers. Pathway-based analyses, involving joint testing of SNPs within gene sets defined by biological pathways, have the potential to empower the identification of new associations not captured by testing individual genetic variants (6, 7).

To gain further insight into the heritability of RCC, we evaluated 3,678 canonical pathways and gene sets using summary-level data from a GWAS meta-analysis. To validate our findings, we analyzed GWAS summary data from the UK Biobank and Kaiser Permanente Genetic Epidemiology Research on Adult Health and Aging (GERA) cohorts.

MATERIALS AND METHODS

Study Populations

The RCC GWAS meta-analysis, described previously (5), combined summary results from six independent GWAS totaling 10,784 RCC cases and 20,406 controls of European ancestry. Briefly, genotypes had been assayed across the scans using a combination of Illumina SNP arrays (Illumina Inc, San Diego, CA, USA). After performing imputation on all scans using 1,094 subjects from the 1000 Genomes Project (phase 1 release 3) as the reference panel, 7,437,091 SNPs were included in the meta-analysis. To facilitate the identification of novel genetic signals, we excluded from our pathway analysis 36,616 SNPs within 500kb of genetic variants previously reported to be associated with RCC at genome-wide significance (P < 5 × 10⁻⁸). All tests of statistical significance used in this analysis were two sided.

To replicate study findings we made use of summary-level association statistics for 30,798,054 SNPs from a GWAS of RCC conducted among UK Biobank participants (903 cases, 451,361 controls) downloaded from GeneATLAS (http://geneatlas.roslin.ed.ac.uk/) (8). The SNP beta coefficients and standard errors in GeneATLAS were computed using mixed linear models; we transformed these summary statistics to odds ratios (ORs) using LMOR (9). Standard errors were calculated from the reported P-value and estimated OR.

For replication of SNP-level analyses, we used summary results from UK Biobank and, for selected SNPs (n=10), a GWAS of kidney cancer conducted among persons of European ancestry in the GERA cohort (317 cases and 50,511 controls). Details of the GERA GWAS have been previously described (10).

Pathway analysis

We downloaded definitions for 3,762 human-derived pathways and gene sets (C2 gene set collection) from the Broad Institute Molecular Signatures Database (MSigDB) v6.1 (http://software.broadinstitute.org/gsea/msigdb/collections.jsp) for the pathway-level analysis. Genomic definitions for genes were downloaded from human genes NCBI36 and reference genome GRCh37.p13 using the Ensemble BioMart tool.

We conducted gene- and pathway-level meta-analyses using the summary statistics-based adaptive rank truncated product (sARTP) method (https://www.rdocumentation.org/packages/ARTP2/versions/0.9.45/topics/sARTP). sARTP combines SNP associations across variants 20kb upstream and downstream of a given gene with adjustment for the size of genes and pathways through a resampling procedure to evaluate the global testing P-value, with proper adjustment of multiple comparisons (11). A web-based sARTP application tool is available allowing users to submit pathway analysis jobs online and receive results computed using NCI computing resources (https://analysistools.nci.nih.gov/pathway/). Significance of gene- and pathway-level associations were estimated from the null distribution generated from 10 million resampling steps. A panel of 503 European subjects (population codes: CEU, TSI, FIN, GBR, IBS) in the 1000 Genomes Project (phase 3, v5) was used in sARTP to estimate the linkage disequilibrium between SNPs. To mitigate the impact of population stratification, we applied genomic control inflation factors to rescale the standard errors of the log odds ratios for SNPs in each GWAS (lambda values 1.009 – 1.058) and in the meta-analysis (lambda = 1.037).

We successfully analyzed 3,678 of the 3,762 pathways and gene sets downloaded from MSigDB; tests of 84 pathways failed because of a lack of SNP coverage. To control the family-wise error rate in our discovery pathway analysis, we considered a Bonferroni-corrected P-value of 1.36 × 10⁻⁵ as being statistically significant (i.e., 0.05/3,678). For our analysis of promising pathway-level results in replication datasets, we considered an alpha of 0.05 as being significant. A schematic summarizing our pathway analysis approach is provided in Supplementary Figure 1.

Gene-level and SNP-level analyses within selected pathways

We evaluated gene-level and SNP-level test results among all constituent genes of pathways found to be associated with RCC in the discovery and replication sets, and combined association statistics through meta-analysis using fixed-effects models.

Functional annotation of SNP associations

We queried public databases to explore the possible biologic effects of selected SNPs. We searched RegulomeDB (http://www.regulomedb.org) and HaploReg (https://pubs.broadinstitute.org/mammals/haploreg/haploreg.php) to assess the likelihood that SNPs map to regulatory elements, and the NHGRI-EBI GWAS Catalog (https://www.ebi.ac.uk/gwas/) to search for previously reported GWAS associations with other traits. We assessed potential SNP associations with gene expression in 527 TCGA (The Cancer Genome Atlas) renal cancer tumor cases (KIRC) using the PancanQTL database (http://bioinfo.life.hust.edu.cn/PancanQTL/) (12). We also explored eQTL kidney cortex expression data (n=73) using GTEx (https://gtexportal.org/home/) (13)

RESULTS

We identified 14 pathways and gene sets significantly associated with RCC in the discovery and replication sets, ranging in size from 12 to 732 genes and 402 to 31,986 SNPs (Table 1). Notably, 10 of the 14 pathways / gene sets include components of the Phosphatidylinositol-3-kinase (PI3K) /Akt signaling pathway, with AKT1 and CASP9 among the most significant gene-level signals for all 10 pathways. The four other gene sets were “Benporath cycling genes” (genes related to embryonic stem cell identity showing cell-cycle stage-specific expression), “Fortschegger PHF8 targets up” (genes upregulated in HeLa cells upon knockdown of PHF8 by RNAi), “West adrenocortical tumor dn” (down-regulated genes in pediatric adrenocortical tumors compared to the normal tissue) and “Pujana CHEK2 PCC network” (genes positively co-expressed with CHEK2).

Table 1.

Pathways and gene sets significantly associated with renal cell carcinoma in the GWAS meta-analysis (P < 1.36 × 10⁻⁵)^a and UK Biobank (P<0.05)

			Discovery Set (GWAS Meta-Analysis; 10,784 cases and 20,406 controls		Replication Set (UK Biobank; 903 cases and 451,361 controls)
MSigDB Pathway / Gene Set	NGenes	NSNPs	P	Most Significant Genesb	P	Most Significant Genesb
Benporath cycling genes	612	26858	2.00 × 10−7	TIPIN, CDKN2C, CENPQ	0.016	FANCD2,RAN,ZC3HC1
Reactome AKT phosphorylates targets in the cytosol	12	402	6.50 × 19−7	AKT1, CASP9, CDKN1B	0.00018	CASP9,AKT1S1,AKT2
BioCarta RAS pathway	21	957	7.50 × 10−7	AKT1, CASP9, MAP2K1	0.0034	CASP9
Reactome PI3K/AKT activation	34	1557	8.50 × 10−7	AKT1, CASP9, IRS2	0.0060	CASP9
Fortschegger PHF8 targets up	247	14870	1.30 × 10−6	TIPIN, INSR, GRB10	0.013	IGF1R,ATF4,CTSC
Reactome PIP3 activates AKT signaling	25	1095	1.45 × 10−6	AKT1, CASP9, CDKN1B	0.0045	CASP9
West adrenocortical tumor dn	491	31986	2.25 × 10−6	RBPMS, IRF5, ZFP36L2	0.042	CASP9,RPL38,KCNK3
Reactome signaling by SCF-KIT	72	3976	5.10 × 10−6	AKT1, CASP9, MAP2K1	0.0074	CASP9,CDK1,SOCS1
Reactome GAB1 signalosome	34	1832	6.15 × 10−6	AKT1, CASP9, CDKN1B	0.0060	CASP9
KEGG prostate cancer	82	5107	6.25 × 10−6	AKT1, CASP9, MAP2K1	0.00046	CASP9,IGF1R,ATF4
BioCarta HDAC pathway	29	2284	6.45 × 10−6	AKT1, INSR, MEF2D	0.034	IGF1R
Reactome PI3K events in ERBB4 signaling	34	2757	7.05 × 10−6	AKT1, CASP9, CDKN1B	0.0060	CASP9
PID PI3KCI AKT pathway	33	1451	9.00 × 10−6	AKT1, CASP9, CDKN1B	0.00058	CASP9
Pujana CHEK2 PCC network	732	26725	9.50 × 10−6	POT1, CDKN2C, TFDP2	0.016	RAN,MIEF1,CENPA

^aPathways/gene sets statistically significant at Bonferroni-adjusted α-level of 1.36 × 10⁻⁵ (0.05/3678).

^bTop 3 gene-level test results with a P-value <0.001

We also explored gene- and SNP-level signals within the 14 significant pathways and gene sets (Supplementary Table 1). In testing across the 2,035 constituent genes, associations with RCC (Bonferroni-corrected P < 2.46 × 10⁻⁵ in discovery and replication sets combined) were observed for CASP9 (P = 3.7 × 10⁻⁷), TIPIN (P = 8.2 × 10⁻⁶) and CDKN2C (P = 1.7 × 10⁻⁵). Promising gene signals in both the discovery and replication sets were also observed for AKT1, ARID1A, EP300, FANCD2, HIST1H4, KCNK3, MAP2K1, RBPMS and RPL4. We identified 4 highly promising SNPs with associations in both the discovery and replication sets at P < 0.0001 and P < 0.05, respectively (Table 2): rs12124078 (within the CASP9 region; P_Combined = 6.9 × 10⁻⁸), rs41324853 (CDKN2C; P_Combined = 2.4 × 10⁻⁷), rs61758464 (AKT1; P_Combined = 3.5 × 10⁻⁷) and rs2979488 (RBPMS; P_Combined = 4.4 × 10⁻⁷).

Table 2.

Summary of SNPs within significant pathways with RCC associations observed in the discovery (P < 1.0 × 10⁻⁴) and replication (P<0.05) sets.

				Discovery		Replicationb		Combinedc
SNP (Gene)	Chr	Positiona	A/a	OR (95% CI)	P	OR (95% CI)	P	OR (95% CI)	P
rs12124078 (CASP9)	1	15869899	A/G	0.92 (0.88 to 0.96)	2.6 × 10−5	0.84 (0.77 to 0.92)	1.5 × 10−4	0.91 (0.87 to 0.94)	6.9 × 10−8
rs41324853 (CDKN2C)	1	51449575	T/C	1.10 (1.05 to 1.14)	2.0 × 10−5	1.15 (1.05 to 1.25)	2.3 × 10−3	1.11 (1.06 to 1.15)	2.4 × 10−7
rs2979488 (RBPMS)	8	30280630	A/G	1.10 (1.05 to 1.14)	1.3 × 10−5	1.13 (1.03 to 1.25)	8.8 × 10−3	1.11 (1.06 to 1.15)	4.4 × 10−7
rs61758464 (AKT1)	14	10525780	G /A	0.87 (0.82 to 0.93)	5.2 × 10−6	0.86 (0.76 to 0.98)	0.022	0.87 (0.83 to 0.92)	3.5 × 10−7

^aGRCh37.p13

^bReplication includes results from UK Biobank (903 cases, 451 361 controls) and GERA Cohort (317 cases, 50 511 controls) combined by meta-analysis using fixed effects model.

^cSummary results from meta-analysis using fixed effects model.

We explored the potential functional impact of these SNPs by integration of publicly accessible resources (Table 3). The variant rs12124078 has a RegulomeDB score of 1f, being associated with CASP9 expression across several non-kidney tissues. We confirmed that this eQTL relationship extends to kidney tissue, with the higher-risk A allele associated with reduced CASP9 expression in both TCGA (P = 3.6 × 10⁻⁷) and GTEx (P = 0.0075) datasets, as well as the majority of other tissue sets in GTEx (Supplementary Figure 2). While the three other SNPs had weaker predicted functional relevance, rs2979488 and rs61758464 were associated with expression of RBPMS and ZBTB42 respectively in TCGA. In a search of the NHGRI-EBI GWAS Catalog for associations with other traits, rs12124078 and rs2979488 have previously been significantly associated with glomerular filtration rate and leukocyte count, respectively.

Table 3.

Exploration of SNP functional relevance: RegulomeDB score, expression quantitative trait locus (eQTL) analyses for selected SNP-gene pairs in tumor (TCGA-KIRC) and normal (GTEx) kidney tissue samples and previous genome-wide significant findings for other traits.

				eQTL Analyses
				TCGA, KIRC (N=527)		GTEx, Kidney Cortex (N=73)
SNP	A/a	RegulomeDB Score	Gene	B a	P	β	P	In NHGRI-EBI GWAS Catalog (Trait)
rs12124078	A/G	1f	CASP9	0.26	3.6 × 10−7	0.39	0.0075	Glomerular filtration rate
rs41324853	T/C	4	- b					- c
rs2979488	A/G	4	RBPMS	−0.27	2.6 × 10−6	0.02	0.89	Leukocyte count
rs61758464	G/A	5	ZBTB42	−0.28	1.2 × 10−4	−0.12	0.51	-

^aSNP-gene association at false discovery rate < 0.05 in database of TCGA eQTLs (12). β represents directional expression effect for rare allele.

^bNo eQTL identified.

^cNo entry in GWAS Catalog.

DISCUSSION

In this pathway-based meta-analysis of RCC GWAS summary results, we identified 14 pathways and gene sets associated with risk. In targeted SNP investigations across the 14 pathways in the discovery and replication sets, we observed an association approaching genome-wide significance overall for the variant rs12124078, which we found to be consistently associated with CASP9 expression in kidney tissue. We also observed promising associations with genetic variation in close proximity to AKT1, CDKN2C, TIPIN and RBPMS.

The majority of the pathway findings appear to be driven by nucleotide variation in components of the PI3K/AKT signaling network, an important regulator of cell growth, proliferation, metabolism, survival, and apoptosis (14). This is one of the most frequently dysregulated signal transduction pathways in human cancers, including kidney cancer, with genetic alterations in constituent genes present in 15% of RCC (15). PI3K/AKT signaling is particularly important in the pathogenesis of clear cell RCC; aberrant pathway activation leads to upregulation of mammalian target of rapamycin (mTOR) signaling, which in turn upregulates hypoxia-inducible factor-mediated expression of angiogenic factors (16). PI3K-inhibiting therapeutic agents are used in treating metastatic RCC.

We also observed replicable pathway-level signals for four gene sets that are related to cancer; two involve cell cycle regulation (“Benporath cycling genes” and “Pujana CHEK2 PCC network”) (17, 18) , a third captures genes downregulated in pediatric aderenocortical tumors compared to normal tissue (“West adrenocortical tumor dn”) (19), and the fourth lists genes up-regulated upon knockdown of PHF8, a histone lysine demethylase and suspected transcription activator overexpressed in several types of cancer (“Fortschegger PHF8 targets up”) (20, 21). As all four gene sets are comparatively large, involving between 247 and 732 genes, it is possible that the observed RCC associations are reflective of signals from a subset of genes, such as CDKN2C and RBPMS.

When we conducted gene- and SNP-level investigations within the 14 significant pathways, the strongest evidence of an association with RCC was with CASP9 and the nearby variant rs12124078, with the A allele associated with increased risk. CASP9 encodes caspase-9, a critical initiator of cell apoptosis that is regulated by PI3K/AKT signaling; Akt phosphorylation at serine-196 inhibits caspase-9 protease activity, decreasing apoptosis (22, 23). Interestingly, the rs12124078 A allele has also been associated with lower glomerular filtration rate and decreased CASP9 expression in peripheral blood monocytes (24). We have confirmed this eQTL, with the A allele being associated with reduced CASP9 expression in TCGA and GTEx kidney tissue and the majority of other GTEx tissue sets. Collectively, these findings are consistent with a reduction in caspase-9-mediated apoptotic activity potentially underlying the association between rs12124078 and RCC.

Genetic variation within the AKT1, RBPMS, TIPIN and CDKN2C gene regions also showed promising evidence of association with RCC. AKT1 is a key member of the PI3K/AKT pathway, encoding a serine/threonine kinase regulating numerous mechanisms affecting cell growth, metabolism and angiogenesis (25). We found the nearby variant rs61758464 to be associated with RCC and while rs61758464 was not related to AKT1 expression in renal tissue, it is notable that the risk allele was associated with reduced AKT1 expression (P = 2.2 ×10⁻⁸; FDR = 2.5 × 10⁻⁴) in blood eQTL data (26). We also observed in TCGA, but not GTEx, an association between this variant and expression of ZBTB42, which encodes a poorly characterized member of the C2H2 zinc finger protein family suspected to play a role in skeletal muscle development (27).

RBPMS encodes a member of the RNA recognition motif family of RNA-binding proteins. The function of RBPMS is poorly understood, although recent evidence suggests a role in mRNA transport and localization (28). The biologic basis for a role of RBPMS in RCC development remains to be established, although it has been shown to interact with VHL in cultured 786-O renal cancer cells (29). Intriguingly, RBPMS expression has been reported to be significantly elevated in tumor tissue of obese vs. non-obese clear cell RCC patients suggesting a possible link between BMI and renal cancer (30).

TIPIN, which was associated with RCC in gene-level testing, encodes a replisome-associated protein that contributes to genome maintenance by mediating Chk1 and Chk2 activation in response to DNA damage (31). TIPIN expression has been reported to be down-regulated in kidney tumor vs. matched normal tissue, possibly reflecting dysregulation of cell-cycle checkpoints (32).

Our findings for CDKN2C, involved in cell cycle regulation, and the nearby SNP rs41324853 likely reflects a previously reported GWAS risk locus. Although we filtered out GWAS results for SNPs within 500kb of previously identified GWAS hits prior to our analysis to prioritize the discovery of new loci, rs41324853 is 542kb from and moderately correlated with (r² = 0.33, D’ = 0.93) the known RCC GWAS risk marker rs4381241 (5). When we ran a logistic model including both SNPs within the discovery set, rs41324853 was no longer associated with RCC risk (P = 0.39). Our eQTL analyses do not offer any further insight into the causal pathway underlying this locus.

Strengths of our pathway-based GWAS meta-analysis, to our knowledge the first of its kind for RCC, include the large sample size and the use of an independent GWAS replication set for confirmation of pathway-, gene- and SNP-level findings. An additional strength is our use of the sARTP method for pathway analysis, which possesses many useful properties. Pathway analysis generally targets two types of null hypotheses: the competitive null hypothesis (33) (i.e., that genes in a candidate pathway are no more associated with the outcome than any other genes outside this pathway) and the self-contained null hypothesis (34) (i.e., that none of the genes in a pathway of interest is associated with the outcome). The sARTP procedure focuses on the self-contained null hypothesis, as the main goal of this project is to identify outcome-associated genes or loci. As pointed out by Goeman et al. (35), tests for the competitive null hypothesis often assume that genotype measured at different genes are independent when evaluating the association significance level. This assumption, which is generally invalid in practice, is not required by sARTP when testing the self-contained null hypothesis. Other strengths of sARTP are its use of summary data and that it can handle large pathways, which might consist of thousands of genes and tens of thousands of SNPs. Other advantages of sARTP include its use of summary data and its capability handling large pathways, which might consist of thousands of genes and tens of thousands of SNPs (11).

A limitation of the sARTP approach involves the inherent assumption of mapping SNPs 20 kb upstream and downstream of each gene to identify candidate SNPs that may play a regulatory role in gene expression. This distance has been used for annotating SNPs/genes in previous pathway analyses (11, 36), as studies have shown functional variants are located approximately 16–20kb within transcription start sites (37–39). However, it does not capture cis regulatory effects outside this window, nor trans mechanisms.

In summary, our pathway-based analysis of the RCC GWAS meta-analysis has provided new promising genetic susceptibility regions that merit further investigation and functional follow-up.

Abbreviations

CI

confidence interval

eQTL

expression quantitative trait locus

GERA

Genetic Epidemiology Research on Adult Health and Aging cohort

GWAS

genome-wide association study

KIRC

kidney renal cell carcinoma

OR

odds ratio

MSigDB

Molecular Signatures Database

RCC

renal cell carcinoma

SNP

single-nucleotide polymorphism

sARTP

summary statistics-based adaptive rank truncated product

TCGA

the Cancer Genome Atlas