Identification of Susceptibility Loci and Genes for Colorectal Cancer Risk

Abstract

Background & Aims

Known Genetic factors explain only a small fraction of genetic variation in colorectal cancer (CRC). We conducted a genome-wide association study (GWAS) to identify risk loci for CRC.

Methods

This discovery stage included 8027 cases and 22577 controls of East-Asian ancestry. Promising variants were evaluated in studies including as many as 11044 cases and 12047 controls. Tumor-adjacent normal tissues from 188 patients were analyzed to evaluate correlations of risk variants with expression levels of nearby genes. Potential functionality of risk variants were evaluated using public genomic and epigenomic databases.

Results

We identified 4 loci associated with CRC risk; P values for the most significant variant in each locus ranged from 3.92×10⁻⁸ to 1.24×10⁻¹²: 6p21.1 (rs4711689), 8q23.3 (rs2450115, rs6469656), 10q24.3 (rs4919687), and 12p13.3 (rs11064437). We also identified 2 risk variants at loci previously associated with CRC: 10q25.2 (rs10506868) and 20q13.3 (rs6061231). These risk variants, conferring an approximate 10%–18% increase in risk per allele, are located either inside or near protein-coding genes that include TFEB (lysosome biogenesis and autophagy), EIF3H (initiation of translation), CYP17A1 (steroidogenesis), SPSB2 (proteasome degradation), and RPS21 (ribosome biogenesis). Gene expression analyses showed a significant association (P <.05) for rs4711689 with TFEB, rs6469656 with EIF3H, rs11064437 with SPSB2, and rs6061231 with RPS21.

Conclusions

We identified susceptibility loci and genes associated with CRC risk, linking CRC predisposition to steroid hormone, protein synthesis and degradation, and autophagy pathways and providing added insight into the mechanism of CRC pathogenesis.

INTRODUCTION

Colorectal cancer (CRC) is one of the most frequently diagnosed malignancies in the world, including most Asian countries.¹ Genetic factors play an important role in the etiology of both familial and sporadic CRC.² Through family-based studies, multiple CRC susceptibility genes have been identified, including APC, MLH1, MSH2, MSH6, PMS2, BMPR1A, SMAD4, POLE, POLD1, GREM1, NTHL1, LKB1/STK11 and MUTYH.³ Deleterious germline mutations in these genes, however, account for less than 5% of CRC cases in the population. It is believed that many common, low-penetrance genetic risk variants exist for CRC and, collectively, these variants explain a substantial proportion of genetic variation for CRC. Genome-wide association studies (GWAS) have become a powerful tool to uncover genetic susceptibility factors for complex diseases. To date, more than 40 CRC GWAS risk loci have been identified,^4–21 which have expanded our understanding of the etiology of CRC. However, these common risk variants, along with known CRC susceptibility genes, explain only a small fraction of genetic variation for CRC.²²

It has been demonstrated that GWAS conducted in non-European populations are valuable in identifying genetic variants for complex traits.^{8, 11} Because of different genetic architectures and environmental exposures, studies in non-European populations can often discover important genetic risk variants that are otherwise difficult to identify.^{8, 11} Furthermore, these studies can uncover risk variants specific to the study population. To search for new CRC susceptibility loci, we conducted a large, multi-staged genetic study including 19,071 cases and 34,624 controls of East Asian ancestry. To investigate the generalizability of our findings to other populations, we evaluated the newly-identified risk variants in16,984 cases and 18,262 controls of European ancestry.

Materials and Methods

Study population

This study was conducted as part of the Asia Colorectal Cancer Consortium (ACCC), including a total of 19,071 cases and 34,624 controls of East Asian ancestry from 14 studies conducted in China, South Korea and Japan (Supplementary Table 1). The details of these studies are described in the supplementary materials. Compared to our previous GWAS studies,^{8, 11} we expanded the discovery effort by including the Japan Biobank (BBJ) study (2,814 cases/11,358 controls) and three additional studies (3,519 cases/5,989 controls) genotyped with a customized exome array, the details of which are described in the genotyping section. Specifically, the discovery stage consisted of 8,027 cases and 22,577 controls, including 4,888 cases and 10,243 controls from six studies conducted among Chinese in Shanghai and Guangzhou, 2,814 cases and 11,358 controls from a study conducted among Japanese in Japan and 325 cases and 976 controls from a study conducted among Koreans in South Korea. Replication stage 1 consisted of 6,532 cases and 8,140 controls, including 677 cases and 1,114 controls conducted among Chinese in Guangzhou, 236 cases and 472 controls from a study conducted among Japanese in Japan and 5,619 cases and 6,554 controls from three studies conducted among Koreans in South Korea. Replication stage 2 included 4,512 cases and 3,907 controls from a study conducted among Chinese in Beijing. All study protocols were approved by the relevant institutional review boards and all study participants provided written informed consent.

The studies in European descendants comprised 16,984 cases and 18,262 controls recruited in North America, Europe and Australia, which are included in three consortia; the Genetics and Epidemiology of Colorectal Cancer Consortium (GECCO), the Colorectal Transdisciplinary (CORECT) Study, and the Colon Cancer Family Registry (CCFR). The details of these studies have been described previously.¹⁰

Genotyping and imputation in the discovery stage

Genotyping and quality control (QC) for studies in the discovery stage have been described previously.^{8, 11} Briefly, genotyping for five of the eight studies included in the discovery stage were completed using Affymetrix or Illumina SNP arrays (Supplementary Table 2). In previous studies conducted in the ACCC, imputation was conducted using the CHB and JPT HapMap 2 panel as the reference.^{8, 11} To increase the coverage of genotypes in this study, we imputed genotypes for each of these five studies with data from the 1000 Genomes Project Phase I Release (Version 3) as the reference. Only SNPs with minor allele frequency (MAF) > 5% and a high imputation quality (R² > 0.70) were included. After QC filtering, approximately 4 million genotyped or imputed SNPs on 22 autosomes remained for analysis. Genotyping for the remaining three studies included in the discovery stage was completed using Illumina Infinium assays with 59,317 variants as the customer add-on contents onto the Illumina HumanExome Beadchip. Of the 302,218 variants included in this array, 44,428 have an MAF of 5% or higher. We did not impute genotypes in these three studies given the low-density coverage of the genome by these common variants. After QC filtering, 40,647 SNPs with an MAF of 5% or higher remained for the analysis. Little evidence of substantial population stratification was found in any of the eight studies included in this stage (the inflation factor for each study was < 1.05).

SNP selection and genotyping in the replication stages

No apparent heterogeneity across the eight studies included in the discovery stage was noted, and thus a fixed-effects meta-analysis was performed to obtain the overall summary statistics for the association of genetic variants with CRC risk. We selected SNPs for further evaluation according to the following criteria: (i) P ≤ 0.0001 in the discovery stage; (ii) consistent direction of association across all eight studies included in the discovery stage (P for heterogeneity > 0.05), (iii) in no or moderate linkage disequilibrium (LD) (r² < 0.30 in East Asians) with any known CRC risk variant or SNPs that had been evaluated in our previous replication stages.^{8, 11} A total of 38 independent association signals were identified. For each signal, we selected the SNP with the most significant P-value for replication. Of these 38 SNPs, 36 were successfully genotyped and evaluated in an independent sample of 6,532 cases and 8,140 controls (replication stage 1). SNPs that showed evidence of association (one-sided P < 0.05) in replication stage 1 and in the same direction of the association as seen in the discovery stage were further evaluated in another independent sample of 4,512 cases and 3,907 controls (replication stage 2) (Supplementary Table 3).

Using the iPLEX Sequenom MassARRAY platform, genotyping for replication stage 1 and 2 studies was conducted at the Vanderbilt Molecular Epidemiology Laboratory and the Fudan-VARI Center for Genetic Epidemiology Core lab, respectively. Four negative controls (water) and eight positive quality controls (HapMap or duplicate samples) were included in each 384-well plate. We filtered out SNPs with (i) genotype call rate < 95%, (ii) genotyping concordance rate < 95% in positive control samples, (iii) an unclear genotype call or (iv) P for Hardy-Weinberg equilibrium < 10⁻⁵ in controls. We calculated the mean concordance rate using data from positive quality control samples. The mean concordant rate was 99% with a median value of 100% in each of the five participating studies included in replication stage 1 and was 98% with a median value of 98% in the replication stage 2 study.

Genotyping of samples in GECCO, CORECT, and CCFR was conducted using Illumina and Affymetrix Arrays. The details of genotyping, QC and imputation have been reported previously.^{5, 10}

Gene expression and protein abundance evaluation of candidate genes

The gene closest to each of the risk variants identified in this GWAS was selected for evaluation of its expression level in correlation with the risk variant in adjacent normal tissues from 188 CRC patients of East-Asian descent. The details of quantitative real-time polymerase chain reaction analysis and quantitative western blotting are provided in Supplementary Methods.

Statistical analysis

Associations between SNPs and CRC risk were evaluated on the basis of the log-additive model using mach2dat,²³ PLINK version 1.0.7,²⁴ R version 3.0.2 (See URLs) and SAS version 9.4. Per-allele odds ratios (ORs) and 95% confidence intervals (CIs) were derived from logistic regression models, adjusting for age, sex and the first ten principal components when appropriate. Association analysis was conducted separately for each study. Because there was no apparent heterogeneity across participating studies, fixed-effects meta-analyses were used to obtain summary results for each stage and for all stages combined with the inverse-variance method using METAL.²⁵ SNPs associated with CRC risk at P < 5×10⁻⁸ in the combined analysis were considered genome-wide significant. Stratified analyses were performed to assess whether the associations differ by tumor sites (colon or rectum), ethnicity (Chinese, Korean or Japanese) and sex (men or women). We estimated heterogeneity across studies and subgroups with Cochran’s Q test. We estimated haplotype frequencies with the expectation maximum algorithms in the Haplo. stat package (see URLs) in R (version 3.0.2) and conducted haplotype analyses when two or more SNPs were identified at the same locus.

Details of the eQTL analysis using data from East-Asian patients and the analysis of the SPSB2 isoform using data from The Cancer Genome Atlas project are provided in Supplementary Methods.

Bioinformatics analyses

To identify putative functional variants for newly identified loci, all variants in LD (r² > 0.20) with the risk variant in each locus were identified using data from the 1000 Genomes Project (phase1 release v3.0). Non-coding variants were mapped to the 15 chromatin states across multiple normal colorectal mucosal tissues derived by the Roadmap Epigenomics (REMC) project ²⁶ and to the epigenomic maps of histone markers (H3K4Me1, H3K4Me3, and H3K27Ac) across CRC cell lines derived by the Encyclopedia of DNA elements (ENCODE) project.²⁷ We focused on variants in promoters or enhancers. We identified the corresponding genes as the target genes for variants in promoters. We identified the interacting genes for variants in enhancers using data of enhancer- transcription start site (TSS) association from the Functional ANnoTation Of the Mammalian Genome 5 (FANTOM5) project, which evaluated correlations of the transcriptional activity of the enhancer and putative target genes. ²⁸ Coding variants were evaluated using the SIFT²⁹ and PolyPhen-2 algorithms.³⁰ We proposed the most likely candidate gene at each locus according to the following criteria: 1) functional evidence; 2) distance to the risk variant; 3) results from eQTL analysis; 4) biologic functions and potential roles in cancer according to previous literature.

Results

Combined analyses of data from the discovery and replication stages revealed seven SNPs associated with CRC risk at the genome-wide significance level (P < 5×10⁻⁸), including rs4711689 at 6p21.1 (TFEB), rs4919687 at 10q24.3 (CYP17A1), rs11064437 at 12p13.3 (SPSB2), rs2450115 and rs6469656 at 8q23.3 (EIF3H), rs10506868 at 10q25.2 (VTI1A), and rs6061231 at 20q13.3 (RPS21) (Table 1). SNP rs10506868 at 10q25.2 was correlated with a risk variant (rs12241008, r² = 0.60, D′ <1 in East Asians) identified in a recent GWAS study,⁷ and thus this locus was not further evaluated in the subsequent analysis of this study. Interestingly, rs10506868 is located in intron 6 of the VTI1A gene, a fusion partner of a CRC susceptibility candidate gene, TCF7L2, identified in our previous GWAS study.⁸

Table 1.

Summary results for newly identified genetic risk variants associated with colorectal cancer risk at P < 5.0 ×10⁻⁸ East-Asian descendants

SNP	Locus	Genea	Positionb	Allelesc	EAFd	Discoverye		Replicatione		Combinede		Heterogeneity
						8,027 cases /22,577 controls		11,044 cases/12,047controls		19,071 cases /34,624 controls
						OR (95% CI)e	Pe	OR (95% CI)e	Pe	OR (95% CI)e	Pe	I2	Pf
rs4711689	6p21.1	TFEB	41692812	A/G	0.85	1.12 (1.06–1.18)	3.41×10−5	1.10 (1.05–1.16)	2.89×10−4	1.11 (1.07–1.15)	3.92×10−8	31	0.13
rs2450115	8q23.3	EIF3H	117624093	T/C	0.58	1.11 (1.06–1.17)	9.19×10−6	1.11 (1.07–1.16)	2.94×10–8	1.12 (1.09–1.15)	1.24×10−12	0	0.51
rs6469656	8q23.3	EIF3H	117647788	A/G	0.67	1.12 (1.07–1.16)	2.59×10−7	1.10 (1.06–1.15)	1.55×10−6	1.11 (1.08–1.14)	2.03×10−12	38	0.07
rs4919687	10q24.3	CYP17A1	104595248	G/A	0.80	1.16 (1.09–1.24)	7.52×10−6	1.13 (1.08–1.19)	1.91×10−7	1.14 (1.10–1.19)	7.82×10−12	0	0.48
rs10506868	10q25.2	VTI1A	114319380	T/C	0.28	1.09 (1.04–1.14)	1.00×10−4	1.10 (1.05–1.14)	9.25×10−6	1.10 (1.06–1.13)	3.98×10−9	0	0.84
rs11064437	12p13.3	SPSB2	6982162	C/T	0.75	1.14 (1.09–1.20)	2.80×10−7	1.10 (1.05–1.15)	2.11×10−5	1.12 (1.08–1.16)	4.48×10−11	52.5	0.03
rs6061231	20q13.3	RPS21	60956917	C/A	0.91	1.19 (1.12–1.27)	8.56×10−8	1.18 (1.07–1.28)	2.20×10−4	1.18 (1.13–1.25)	8.04×10−11	0	0.47

EAF, effect allele frequency; OR, odds ratio; CI, confidence interval.

^aThe closest gene.

^bThe chromosome position (bp) is based on the National Center for Biotechnology Information (NCBI) database, build 37.

^cEffect/reference alleles are based on forward allele coding in NCBI, build 37. OR and 95% CI were estimated based on the effect allele (bold).

^dEAF in controls from all studies combined.

^eSummary OR (95% CI) and P were obtained from a fixed-effects meta-analysis.

^fP for heterogeneity across all studies was calculated using Cochran’s Q test.

Associations with CRC risk in each of the six risk variants were, in general, consistent across all studies with little evidence of heterogeneity (Figure 1). Stratified analyses by sex and tumor site showed little evidence of heterogeneity (Supplementary Table 4). Except for rs11064437 at 12p13.3 (P = 0.0003), no apparent heterogeneity was found for the association of variants with CRC risk by ethnicity. The heterogeneity with rs11064437 was primarily driven by its null association in studies in South Korea (OR = 1.01, P = 0.73).

Figure 1. Forest plots for newly identified CRC risk variants (a) rs4711689, (b) rs2450115, (c) rs6469656, (d) rs4919687, (e) rs11064437, (f) rs6061231.

Per-allele OR estimates are presented as boxes with the area proportional to the inverse variance of the estimates. Horizontal lines represent the coverage of 95% confidence intervals (CIs).

We next evaluated the association of the six newly identified CRC risk variants in populations of European ancestry using data from 16,984 cases and 18,262 controls. With the exception of rs4711689 and rs11064437, all other SNPs were associated with CRC risk at P < 0.05 and in the same direction as observed in East Asian studies (Table 2). There was evidence of heterogeneity in the association of CRC risk with rs4919687, rs4711689 and rs6061231 between populations of European ancestry and East Asian ancestry (P < 0.01). The effect allele frequency (EAF) for all six risk variants differs substantially between European descendants and East Asians (Tables 1–2). For example, for rs11064437, the EAF is 0.75 for East Asians while it is 0.996 for European descendants, which may explain the non-significant finding for this SNP in populations of European ancestry.

Table 2.

Associations of newly identified risk variants with CRC risk in populations of European ancestry (16,984 cases and 18,262 controls)

Locus	SNP (alleles)a	Geneb	Positionc	EAFd	OR (95% CI)e	Pe	P. hetf
6p21.1	rs4711689 (A/G)	TFEB	41692812	0.54	1.01 (0.97–1.04)	0.69	8.94×10−5
8q23.3	rs2450115 (T/C)	EIF3H	117624093	0.81	1.08 (1.03–1.12)	0.0003	0.18
8q23.3	rs6469656 (A/G)	EIF3H	117647788	0.89	1.06 (1.01–1.11)	0.02	0.13
10q24.3	rs4919687 (G/A)	CYP17A1	104595248	0.70	1.05 (1.02–1.09)	0.002	0.002
12p13.3	rs11064437 (C/T)	SPSB2	6982162	0.996	1.17 (0.85–1.62)	0.33	0.77
20q13.3	rs6061231(C/A)	RPS21	60956917	0.71	1.07 (1.04–1.11)	8.90×10−5	0.002

EAF, effect allele frequency; OR, odds ratio; CI, confidence interval.

^aEffect/reference alleles are based on forward allele coding in NCBI, build 37.

^bthe closest gene.

^cChromosome position (bp) is based on NCBI Build 37.

^dEAF from 1000 Genomes Project October 2013 release (CEU+FIN+GBR+IBS+TSI).

^eThe summary OR (95% CI) and P value were obtained from a fixed-effects meta-analysis.

^fP for heterogeneity between East Asian and European populations was calculated using Cochran’s Q test.

SNPs rs2450115 and rs6469656 lie approximately 24kb apart at 8q23.3, and were correlated only moderately (r²= 0.20, D′ <1 in East Asians). The association with either of these SNPs remained statistically significant (P = 9.60×10⁻⁶ for rs2450115 and P = 8.30×10⁻⁴ for rs6469656) in the conditional analysis including both SNPs in the same model (Supplementary Table 5), suggesting that there may be two independent association signals. Furthermore, the haplotype (TA) comprising the risk allele of both SNPs was significantly associated with a 1.20-fold increased risk of CRC (P = 1.34 ×10⁻¹⁵) compared to the haplotype (CG) comprising the low-risk allele of both SNPs. These two SNPs lie close to rs16892766 (6.6kb centromeric for rs2450115 and 17 kb telomeric for rs6469656), a CRC-associated variant identified in a previous GWAS in European descendants.¹⁷ However, rs16892766 and any of its highly correlated SNPs are monomorphic in East Asian populations. Thus rs16892766 was not in LD with either of the two risk variants identified in our study. Even in European-ancestry populations, these two SNPs also are poor surrogates for rs16892766 (r² = 0.02, D′=1 for rs2450115 and r² = 0.01, D′<1 for rs6469656). Interestingly, neither rs2450115 nor rs6469656 was reported to be associated with CRC risk in two previous fine-mapping studies of 8q23.3 for CRC risk in European descendants.^{31, 32} In our study with a larger sample size, we showed that rs2450115 and rs6469656 were associated with CRC risk in European descendants (P = 0.0003 and 0.02, respectively, Table 2). These two SNPs are in LD with each other in European descendants (r² = 0.40, D′=1), and only SNP rs2450115 remained statistically significant in the conditional analysis including both SNPs in the model (P = 0.007 for rs2450115 and P = 0.78 for rs6469656).

We analyzed the expression levels of the nearest genes in adjacent normal tissues obtained from 188 CRC patients of East Asian ancestry and performed expression quantitative trait locus (eQTL) analysis. CYP17A1 expression is largely confined to adrenal and gonadal tissues in adults, and thus was not detected in our samples. The remaining four genes evaluated were all highly expressed in normal colorectal tissues. Of the two variants evaluated at 8q23.3, only rs6469656 was significantly correlated with EIF3H expression at P < 0.05 (Figure 2). For the remaining three loci, each of the risk variants was statistically significantly correlated with its nearest gene at P < 0.05 (Figure 2, Supplementary Table 6). At 8q23.3, we also evaluated the expression of the UTP23 gene, a candidate gene suggested by a previous fine-mapping study at this locus.³¹ However, we did not find any statistically significant correlation of risk variants at 8q23.3 with UTP23 expression (P = 0.90 and 0.74, respectively, for rs2450115 and rs6469656). To evaluate the correlation of mRNA and protein levels for the four genes whose expression levels were significantly associated with the risk variants, we analyzed the protein level for each of these genes in adjacent normal tissues obtained from 16 patients (Supplementary Figure 1). The null hypothesis was that there would be no correlation between mRNA and protein levels. The correlation (r) between mRNA expression and protein level was −0.76, 0.64, −0.42 and 0.58 for TFEB, EIF3H, SPSB2 and RPS21, respectively (P = 0.05, 0.12, 0.07 and 0.24, respectively).

Figure 2. Expression quantitative trait locus (eQTL) analyses in samples of normal colorectal mucosa adjacent to the tumor obtained from 188 CRC patients of East Asian ancestry.

a. TFEB relative expression and genotypes of rs4711689; b. EIF3H relative expression and genotypes of rs6469656; c. SPSB2 relative expression and genotypes of rs11064437; d. RPS21 relative expression and genotypes of rs6061231.

To evaluate whether the expressions of these genes were deregulated in tumor tissues, we evaluated the differences between tumor and normal tissues in the expression level of the EIF3H, UTP23, SPSB2, TFEB and RPS21 genes using data from the 188 patients mentioned above. With the exception of TFEB, expression levels of the other four genes were significantly higher in tumors (P < 0.005, Supplementary Table 6). For TFEB, the expression level in tumors was significantly lower than normal tissues (P = 9.26 ×10⁻¹¹, Supplementary Table 6).

We used publically available functional genomic data to further annotate the risk loci identified in our study. We observed an overlap of promoter and enhancer sequences at each locus and identified multiple potential target genes. Combining results from eQTL analyses, functional annotation, and literature review, as provided in detail in the discussion section, we proposed the likely candidate gene: TFEB at 6p21.1, EIH3H at 8q23.3, CYP17A1 at 10q24.3, SPSB2 at 12p13.3 and RPS21 at 20q13.3. The results of in silico functional annotation of the newly identified loci and literature review on potential roles of candidate genes are summarized in Table 3 and Supplementary Table 7. Regional association plots and functional genomic landscapes of these loci are presented in Figure 3 and Supplementary Figure 2.

Table 3.

Summary of risk variants, potential regulatory roles and the likely candidate gene in each new CRC risk locus identified in this study

SNP (Locus)	Genomic annotation	No. of genes in 1-M window	No. of correlated SNPs (r2 > 0.20)	Potential regulatory roles	The likely candidate gene	Expression difference in tumor vs. adjacent normal tissues*	Biological functions	Known roles in cancer
rs4711689 (6p21.1)	Intron 6 of TFEB	43	17	Promoter, enhancer	TFEB	Down-regulated	Lysosomal biogenesis and autophagy	Oncogenic transcriptional factor for some cancers
rs2450115 (8q23.3)	3′ of EIF3H	12	65	Unknown	-	-	-	-
rs6469656 (8q23.3)	3′ of EIF3H	13	125	Promoter, enhancer	EIF3H	Up-regulated	Translational initiation	Promotion of protein synthesis and cancer cell proliferation
rs4919687 (10q24.3)	Intron 1 of CYP17A1	40	412	Promoter, enhancer	CYP17A1	N/A	Steroid biogenesis	Modification of hormone level
rs11064437 (12p13.3)	Splice receptor at intron 1 of SPSB2	79	108	Promoter, enhancer	SPSB2	Up-regulated	Proteasomal degradation	Less known
rs6061231 (20q13.3)	5′ flanking of RPS21	52	126	Promoter, enhancer	RPS21	Up-regulated	Ribosome biogenesis	Potential cancer driver for CRC

^*The expression of each gene was evaluated in paired tumor and adjacent normal colorectal mucosal samples from 188 East Asian patients.

Figure 3. Regional association plots of the risk loci (a) rs4711689 at 6p21.1ⁱ;(b) rs2450115 at 8q23.3;(c) rs6469656 at 8q23.3; (d) rs4919687 at 10q24.3; (e) rs11064437 at 12p13.3; (f) rs6061231 at 20q13.3.

In each plot, the log₁₀ (P-values) (y axis) for the association of SNPs with CRC risk are shown according to their chromosomal positions (x axis) in NCBI Build 37. Blue lines represent the estimated recombination rates from the 1000 Genomes Project (Phase 3). Arrows indicate genomic locations of genes within the 1Mb regions centered on the newly-identified risk variants. The color of the SNP represents its LD (r²) with the index SNP at each locus. Only results from the GWAS studies (4,508 cases /16,588 controls in total) are shown. The plots were generated using the online tool LocusZoom.⁵⁹

We evaluated the association of all GWAS-identified CRC genetic susceptibility loci with data available in our study (Supplementary Table 9). We replicated all 10 risk loci initially identified in GWAS conducted in East Asians at P < 5 × 10⁻⁸. Of the 27 risk variants initially identified in European descendants, 21 were associated with CRC risk at P < 0.05 in East Asians in the same direction as reported previously. In particular, variants at 1q41, 8q24.21, 10p14 and 18q21.1 were associated with CRC risk at P < 5 × 10⁻⁸. SNPs rs6691170, rs16892766, rs35360328, rs3184504, rs73208120, rs72647484 and rs17094983 are monomorphic in East Asian descendants, and thus we were not able to evaluate their associations in our study. We also did not evaluate the X-chromosome SNP rs5934683, for which the genotype data were not available in our study.

Discussion

In this multiple-staged GWAS with a total of 19,071 cases and 34,624 controls, we identified four novel risk loci for CRC (6p21.1, 8q23.3, 10q24.3 and 12p13.3) and new variants in two previously GWAS-identified loci (10q25.2 and 20q13.3). Using gene expression data from 188 CRC patients of East Asian ancestry, we identified potential candidate genes including TFEB, EIF3H, SPSB2 and RPS21.

At 12p13.3, rs11064437 lies at a splice receptor site within intron 1 of SPSB2. The T allele of rs11064437 changes the splice site from TC (canonical site AG³³ on the reverse strand) to TT, and is predicted to disrupt the splicing and introduce a transcriptional isoform with a shortened untranslated region at exon 2 (Supplementary Figure 3). Using data in colon tumor tissues from TCGA, we found a suggestive association of the T allele of rs11064437 with the relative abundance of this isoform (P = 0.11, Supplementary Figure 4). In addition, some of the highly correlated variants were mapped to an active promoter region of SPSB2 in normal colorectal cells (Supplementary Figure 2). These findings, together with results from the eQTL analysis in CRC patients, suggested that variants at 12p13.3 might contribute to CRC risk by regulating the expression of SPSB2. Previous studies have shown that alternative splicing of untranslated regions of exons can affect the stability³⁴ and translation efficiency of pre-mRNAs.³⁵ Similarly, it is possible that rs11064437 may affect post-transcriptional regulation of SPSB2. Misregulated pre-mRNA splicing, which produces aberrant isoforms, has been found to contribute to tumorigenesis.³⁶ Therefore, rs11064437 may be one of the causal variants for the association observed at this locus. Future experiments evaluating SPSB2 splicing diversity and its relevance for the observed association are needed to elucidate the role of this variant. SPSB2 encodes a protein with a suppressor of cytokine signaling domain that functions as an E3 ubiquitin ligase and a SPRY domain that recognizes substrate for ubiquitination and degradation.³⁷ This protein is predicted to regulate the expression of interacting proteins by targeting them for ubiquitination and subsequent proteasomal degradation.³⁸ For example, SPSB2 specifically interacts with the inducible nitric oxide synthase (iNOS), promotes its degradation and decreases the NO production.³⁸ Selective inhibitors of iNOS have been proposed for chemoprevention of CRC.³⁹ Nonetheless, there were other potential target genes at this locus, such as ENO2, CDCA3 and USP5 (Supplementary Table 7).

At 6p21.1, rs4711689 lies in intron 6 of TFEB. This SNP and its highly correlated SNPs were mapped to a region with enhancer activities in normal colorectal tissues (Supplementary Figure 2). Using enhancer-promoter interaction data derived by the FANTOM5 project, we identified TFEB as a potential target gene (enhancer near rs4711689 (chr6:41672431-41703298) and TFEB expression pairwise correlation r =0.33, FDR < 2.2 ×10⁻¹⁶). Together with the eQTL analysis at this locus, we proposed that TFEB was a likely candidate gene at this locus. TFEB encodes a basic helix-loop helix leucine zipper transcription factor, which has been found to be oncogenic for melanoma,⁴⁰ renal cell carcinoma⁴¹ and some sarcomas.⁴² However, the role of TFEB in CRC is unknown. Recent studies have shown that this gene is a master regulator of the autophagy-lysosome pathway and of energy metabolism.⁴³ Deregulated autophagy has been linked to cancer initiation and progression.⁴⁴ These pieces of evidence further support our hypothesis that TFEB is the likely candidate gene at this locus. Notably, TFEB activation is regulated by mTORC1 signaling⁴⁵, which is thought to be one of the fundamental mechanisms for sustaining tumor growth in CRC.⁴⁶ It is possible that TFEB may contribute to CRC risk, at least in part, by mediating downstream effects of the mTORC1 pathways. Our results, along with those from previous studies, support a new role for TFEB and provide clues for future studies on the contribution of autophagy deregulation to colorectal etiology.

At 10q24.3, rs4919687 lies in intron 1 of the gene CYP17A1. This gene encodes a membrane-bound dual-function monooxygenase that catalyzes both 17α-hydroxylase and 17, 20-lyase activities to produce androgenic and estrogenic sex steroids, which is the key branch point in steroidogenesis. Epidemiological studies suggest that steroid hormones may be associated with risk of CRC and other cancers.⁴⁷ Thought to affect steroid hormone levels, CYP17A1 variants, particularly promoter variant rs743572, have been extensively studied in association with risk of hormone-related cancers; ^{48, 49} however, the results were very inconsistent. Interestingly, this promoter variant (rs743572) is correlated with the risk variant rs4919687 identified in our study (r² = 0.24, D′<1 in East Asians and r² = 0.67, D′=1 in European descendants). We herein provide, for the first time, convincing evidence for an association of CYP17A1 gene variants with CRC risk, suggesting that steroid hormones may indeed play a role in the etiology of CRC.

At 8q23.3, two risk variants were identified in our study. Consistent with results from our eQTL analyses, rs2450115 and its correlated variants were mapped to a region of a quiescent state in both normal and cancerous colorectal cells (Supplementary Figure 2), suggesting that these DNAs may be largely inactive, while some variants that are highly correlated with rs6469656 were mapped to active promoter regions of the gene EIF3H (Supplementary Figure 2). Thus, it is likely that some of these variants might be involved in the transcriptional regulation of EIF3H. Together with the evidence from the eQTL analysis, we identified EIF3H as a likely target gene for variants at this locus. EIF3H encodes a subunit of the eukaryotic translation initiation factor 3. An oncogenic role of this gene was demonstrated by its ability to transform NIH-3T3 cells into their malignant forms.⁵⁰ Recent studies have shown that EIF3H promoted cancer cell growth by stimulating protein synthesis.⁵¹ Another possible candidate gene at this locus is RAD21 (Supplementary Figure 2), which encodes a subunit of cohesin that involves chromosome segregation and DNA repair. A recent study showed that RAD21 was a pivotal mediator of APC heterozygous loss, the event initiating colorectal tumorigenesis.⁵²

At 20q13.3, rs6061231 lies 5.2kb upstream of the gene RPS21, and 36kb telomeric to rs4925386 (r² = 0.15, D′ <1 in East Asians and r² = 0.44, D′ <1 European descendants), a risk variant identified in a previous GWAS conducted in populations of European ancestry.²⁷ However, rs4925386 showed a much weaker association with CRC risk (OR = 1.05, P = 0.06) than rs6061231 (OR = 1.20, P = 2.06 ×10⁻⁶) in our analysis of 2,098 cases and 6,172 controls of East-Asian ancestry for which these two SNPs were genotyped (Supplementary Table 8). The association with rs6061231 remained statistically significant, while the association with rs4925386 diminished in the conditional analysis including both SNPs in the model (Supplementary Table 8), indicating that the association signal at 20q13.3 was better captured by rs6061231 in East Asians. SNP rs6061231 and some of its highly correlated SNPs were mapped to a 5′ flanking region of RPS21 (Supplementary Figure 2), suggesting that these variants might affect the transcriptional activities of RPS21. In line with evidence from eQTL analysis, these epigenomic data supported the hypothesis that RPS21 was the likely target gene. RPS21 encodes a ribosomal protein that is a component of the 40S subunit of ribosomes. However, its role in tumorigenesis remains unclear. A recent study that characterized differences in eQTLs between tumor and matched normal colon tissues showed that RPS21 had a significantly higher allelic-specific expression somatic event rate and suggested that this gene was a potential driver in CRC.⁵³ This finding further supported that RPS21 played a role in CRC etiology.

In the analysis comparing mRNA and protein levels in adjacent normal tissues, we observed a positive correlation of the mRNA level and protein abundance for the EIF3H and RPS21 genes. For the TFEB and SPSB2 genes, however, their mRNA and protein levels were inversely correlated, although the correlation for the SPSB2 gene was not statistically significant. Although we could not rule out the possibility of chance finding for the inverse correlation, TFEB and SPSB2 are regulatory proteins that are likely to be unstable but have stable mRNAs in steady-state cells. As suggested by previous studies,⁵⁴ such proteins are predisposed to rapid translational regulation, and thus may sometimes show an inverse correlation of their mRNAs. On the other hand, EIF3H and RPS21 are structural proteins with stable mRNAs and proteins, for which mRNA levels may explain a majority of the variation in protein levels, according to previous studies of proteomics in colon tumors cells and other mammalian cells. ⁵⁵

By comparing the expression level of candidate genes in tumors with adjacent normal tissues, we showed that these genes were deregulated in tumor tissues, suggesting a possible role of these genes in the development of CRC. For example, the down regulation of TFEB, the master regulator of autophagy, in tumor tissues compared to adjacent normal tissues suggests a decreased level of autophagy in the tumors, which was consistent with previous studies showing that autophagy was suppressed in certain cancer cells.⁵⁶ Furthermore, it has been proposed that TFEB is a druggable target for autophagy induction therapy⁵⁷ that has been suggested as a treatment for some cancers. ⁵⁸

We found that more than three quarters of the risk variants initially identified in GWAS conducted in European descendants were directly replicated in East Asians at P < 0.05, indicating that European and East Asian descendants share most of the genetic risk variants for CRC. However, the strength of the association for some of these risk variants was weaker in East Asians than European descendants. Furthermore, approximately 25% of the risk variants identified initially in studies of European descendants were not replicated in our study, including 7 SNPs that are monomorphic in Asians. Not replicating some of the known risk variants is expected given the difference in LD patterns between Asian and European-ancestry populations. It is likely that other variants in these loci may be associated with CRC risk, and fine-mapping studies are needed to identify these variants.

Our study has several limitations. Multiple participating studies did not collect information for family history of CRC, preventing us from evaluating the possible interaction between family history and these risk variants with adequate power, particularly since the prevalence of family history of CRC is low in the study populations. We evaluated eQTLs only for the genes closest to the risk variants identified in this study. However, it is possible that these risk variants might be correlated with other genes. Future studies are needed to fully uncover the biologic mechanism for the associations observed at these novel loci.

In conclusion, to our knowledge, this is the largest GWAS conducted to date to search for novel genetic susceptibility loci for CRC. Our study provides strong evidence for possible roles of SPSB2, TFEB, EIF3H, CYP17A1 and RPS21 in the etiology of CRC. These genes are involved in different aspects of cellular homeostasis, from translational initiation, protein synthesis, to proteasomal degradation and lysosomal biogenesis and autophagy, highlighting the connection of homeostatic regulation and CRC etiology. Aberrant regulation of intestinal epithelial homeostasis plays an important role in the initiation and progression of CRC. Our study has provided additional insights into the genetics and biology of CRC. Results from our study will be helpful for future studies to uncover molecular mechanisms of CRC tumorigenesis.

Abbreviations

ACCC

Asia Colorectal Cancer Consortium

CRC

colorectal cancer

CYP17A1

cytochrome P450, family 17, subfamily A, polypeptide 1

EIF3H

Eukaryotic Translation Initiation Factor 3, Subunit H

GWAS

genome-wide association study

RPS21

ribosomal protein S21

SNP

single nucleotide polymorphism

SPSB2

splA/ryanodine receptor domain and SOCS box containing 2

TFEB

transcription factor EB

VTI1A

vesicle transport through interaction with t-SNAREs 1A