Meta-analysis of genome-wide association data identifies four new susceptibility loci for colorectal cancer

Abstract

Genome-wide association (GWA) studies have identified multiple loci at which common variants modestly influence the risk of developing colorectal cancer (CRC). To enhance power to identify additional loci with similar effect sizes, we conducted a meta-analysis of two GWA studies, comprising 13,315 individuals genotyped for 38,710 common tagging SNPs. We undertook replication testing in up to eight independent case-control series comprising 27,418 subjects. We identified four previously unreported CRC risk loci at 14q22.2 (rs4444235, BMP4; P = 8.1 × 10⁻¹⁰), 16q22.1 (rs9929218, CDH1; P = 1.2 × 10⁻⁸), 19q13.1 (rs10411210, RHPN2; P = 4.6 × 10⁻⁹) and 20p12.3 (rs961253; P = 2.0 × 10⁻¹⁰). These findings underscore the value of large sample series for discovery and follow-up of genetic variants contributing to the etiology of CRC.

Whereas inherited susceptibility is responsible for ~35% of all CRC¹, high-risk germline mutations in APC, the mismatch repair (MMR) genes, MUTYH (MYH), SMAD4, BMPR1A and STK11/LKB1 account for <6% of all cases². Recent GWA studies have validated the hypothesis that part of the heritable risk is caused by common, low-risk variants, identifying CRC susceptibility loci mapping to 8q24 (rs6983267)^3,4, 8q23.3 (rs16892766, EIF3H)⁵, 10p14 (rs10795668)⁵, 11q23 (rs3802842)⁶, 15q13 (rs4779584)⁷ and 18q21 (rs4939827, SMAD7)^6,8.

GWA studies are not contingent on prior information concerning candidate genes or pathways, and thereby have the ability to identify important variants in hitherto unstudied genes. However, the effect sizes of individual variants, the need for stringent thresholds for establishing statistical significance, and financial constraints on numbers of variants that can be followed up inevitably constrain study power. We recently published two separate GWA studies for CRC. To augment the power to detect additional CRC risk loci, we have conducted a meta-analysis of data from these studies and followed up the best supported associations in large sample sets. This analysis, in conjunction with a replication study using eight independent case-control series, has enabled us to identify four new loci predisposing to CRC. This brings to ten the number of independent loci conclusively associated with CRC risk, and provides additional insight into the genetic architecture of inherited susceptibility to CRC.

RESULTS

Meta-analysis of genome-wide association scans

The GWA studies were both conducted by centers in London and Edinburgh, and were both based on designs involving two-phase strategies and using samples from UK populations (Table 1 and Supplementary Table 1 online). The London phase 1 was based on genotyping 940 cases with familial colorectal neoplasia and 965 controls ascertained through the Colorectal Tumour Gene Identification (CoRGI) consortium for 555,352 SNPs using the Illumina HumanHap550 BeadChip Array. Phase 1 in the Edinburgh study consisted of genotyping 1,012 early-onset (aged ≤55 years) Scottish CRC cases and 1,012 controls for 555,510 SNPs using the Illumina HumanHap300 and HumanHap240S arrays. After applying quality control filters, the following data were available: London phase 1, 547,487 SNP genotypes from 922 familial neoplasia cases (614 with CRC and 308 with high-risk colorectal adenomas) and 927 controls; Edinburgh phase 1, 548,586 SNP genotypes from 980 CRC cases and 1,002 controls.

Table 1.

Overview of study design

Study	Cases	Controls	Number of SNPs genotyped
GWA series
London phase 1	922	927	547,487
Edinburgh phase 1	980	1,002	548,586
London phase 2	2,854	2,822	38,715
Edinburgh phase 2	2,024	2,092	38,710
Replication series
London replication	3,286	3,017	9
Edinburgh replication	676	842	8
VCQ58	1,543	2,236	9
FCCPS	962	846	7
Canada	1,175	1,184	3
DACHS	1,373	1,480	3
Kiel	2,169	2,145	3
SEARCH	2,222	2,262	3
Total	20,186	20,855

London phase 2 was based on genotyping 2,873 CRC cases and 2,871 controls ascertained through the National Study of Colorectal Cancer Genetics (NSCCG), whereas Edinburgh phase 2 was based on genotyping 2,057 cases and 2,111 controls. For phase 2, the London and Edinburgh samples were genotyped for a common set of SNPs: the 14,982 SNPs most strongly associated with colorectal neoplasia from London phase 1; the 14,972 most strongly associated SNPs from Edinburgh phase 1 (432 of these SNPs were common to both the London and Edinburgh lists of most strongly associated SNPs); and 13,186 SNPs showing the strongest association with CRC risk from a joint analysis of all CRC cases and controls from both phase 1 data sets (that were not already included in any of the preceding categories). Therefore, phase 2 was based on genotyping 42,708 SNPs in total. After applying quality control filters, the following data were available: London phase 2, 38,715 polymorphic SNPs in 2,854 cases and 2,822 controls; Edinburgh phase 2, 38,710 polymorphic SNPs in 2,024 cases and 2,092 controls. Overall, there were 38,710 polymorphic SNPs common to all four data sets (phases 1 and 2 in London and Edinburgh).

Prior to undertaking the meta-analysis of phases 1 and 2, we searched for potential errors and biases in the four case-control series. Comparison of the observed and expected distributions showed little evidence for an inflation of the test statistics in any of the data sets (inflation factor λ = 1.02 and 1.05 for London phases 1 and 2, and 1.02 and 1.08 for Edinburgh phases 1 and 2, based on the 90% least significant SNPs; Supplementary Fig. 1 online), thereby excluding the possibility of significant hidden population substructure, cryptic relatedness among subjects or differential genotype calling. Using data on all CRC cases and controls from the four series, we derived joint odds ratios (ORs) and confidence intervals (CIs) under a fixed-effects model for each SNP, and associated P values from the standard normal distribution. The distribution of the association P values was significantly skewed from the null distribution: 76 of the SNPs had a P value < 10⁻⁴, greater than the 54 conservatively expected under the null hypothesis (P = 2 × 10⁻³, binomial test).

Of the 23 SNPs associated with CRC risk at P < 10⁻⁵, 14 map to regions that have been the subject of previous fast-tracking replication analyses: 8q24 (rs6983267, rs7014346, rs7837328, rs10808555)^3,4, 11q23 (rs3802842, rs11213809, rs10749971)⁶, 18q21 (rs12953717, rs4939827)⁸, 8q23 (rs16892766, rs11986063, rs6983626)⁵ and 15q13 (rs4779584, rs10318)⁹ (Supplementary Table 2 online). We therefore focused on the nine remaining SNPs from five distinct genomic loci that were associated with CRC risk at P < 10⁻⁵ (Table 2), and that potentially represented previously unreported disease-associated loci. This threshold for follow-up did not exclude the possibility that other SNPs represented genuine association signals, but was simply a pragmatic strategy for prioritizing replication.

Table 2.

Summary of results for nine SNPs selected for replication

							Phases 1 and 2 meta		Replicationb		All phases
SNP	Gene	Chr.	Position (bp)	Major allele	Minor allele	MAFa	OR (95% CI)	P	OR (95% CI)	P	OR (95% CI)	P	Phet
rs10411210	RHPN2	19	38,224,140	C	T	0.10	0.79	4.9 × 10−8	0.90	7.0 × 10−4	0.87	4.6 × 10−9	0.02
							(0.72–0.86)		(0.85–0.96)		(0.83–0.91)
rs961253		20	6,352,281	C	A	0.36	1.13	8.9 × 10−7	1.11	4.6 × 10−5	1.12	2.0 × 10−10	0.70
							(1.08–1.19)		(1.06–1.17)		(1.08–1.16)
rs355527		20	6,336,068	G	A	0.33	1.13	1.2 × 10−6	1.11	3.7 × 10−5	1.12	2.1 × 10−10	0.95
							(1.08–1.19)		(1.06–1.17)		(1.08–1.17)
rs9929218	CDH1	16	67,378,447	G	A	0.29	0.88	1.4 × 10−6	0.93	3.5 × 10−4	0.91	1.2 × 10−8	0.56
							(0.83–0.92)		(0.90–0.97)		(0.89–0.94)
rs4444235	BMP4	14	53,480,669	T	C	0.46	1.12	1.8 × 10−6	1.10	9.4 × 10−5	1.11	8.1 × 10−10	0.64
							(1.07–1.18)		(1.05–1.16)		(1.08–1.15)
rs1862748	CDH1	16	67,390,444	C	T	0.31	0.88	2.6 × 10−6	0.93	7.1 × 10−4	0.91	2.9 × 10−8	0.71
							(0.84–0.93)		(0.90–0.97)		(0.88–0.94)
rs4951291		1	202,273,161	C	T	0.14	0.85	5.3 × 10−6	1.02	0.6	0.93	4.1 × 10−3	0.01
							(0.79–0.91)		(0.95–1.09)		(0.88–0.98)
rs7259371	RHPN2	19	38,226,481	G	A	0.18	0.86	5.7 × 10−6	0.91	5.2 × 10−3	0.89	2.2 × 10−7	0.03
							(0.81–0.92)		(0.86–0.97)		(0.85–0.93)
rs4951039		1	202,273,220	A	G	0.14	0.85	5.8 × 10−6	1.09	0.05	0.94	2.4 × 10−2
													1.8 × 10−3
							(0.79–0.91)		(1.00–1.19)		(0.89–0.99)

^aMAF in controls in weighted average of MAFs from each replication series

^bBased on replication data from the following series: for rs10411210 and rs9929218, London replication, Edinburgh replication, VCQ58, FCCPS, Canada, DACHS, Kiel and SEARCH; for rs1862748, London replication, Edinburgh replication, VCQ58, Canada, DACHS, Kiel and SEARCH; for rs4951039, London replication and CVQ58; for all other SNPs, London replication, Edinburgh replication, VCQ58 and FCCPS.

Replication analyses

To identify the true risk alleles among these nine SNPs, we conducted a replication study initially based on four independent case-control series (London replication, Edinburgh replication, VCQ58 and Finnish Colorectal Cancer Predisposition Study) involving a total of 13,408 individuals (Table 1, Supplementary Table 1 and Supplementary Table 3 online). On the basis of the combined analyses, we found that signals from seven SNPs, representing four loci, reached strong levels of evidence (i.e., P < 5.0 × 10⁻⁷) for an association with CRC risk (Table 2 and Fig. 1), with four SNPs (three genomic regions) satisfying the proposed threshold for genome-wide statistical significance (i.e., P < 10⁻⁸).

Figure 1.

Regional plots of the four confirmed associations (14q22.2, 16q22.1, 19q13.1 and 20p12.3). Each panel shows single-marker association statistics (as −log₁₀ P) from the combined analysis of phase 1 (red), phases 1 and 2 (blue) and all phases (yellow) as a function of genomic position (NCBI build 36.1). Recombination rate across each region in HapMap CEU shown in black (right y axis). Also shown are the relative position of genes mapping to each region of association.

The strongest statistical evidence for a new CRC risk locus was provided by two SNPs: rs961253 (combined OR = 1.12, 95% CI 1.08–1.16, P = 2.0 × 10⁻¹⁰, P_het = 0.70, I² = 0%; Fig. 2) and rs355527 (combined OR = 1.12, 95% CI 1.08–1.17, P = 2.1 × 10⁻¹⁰, P_het = 0.95, I² = 0%; Table 2). The corresponding statistics excluding the discovery phase 1 of the GWA studies were OR = 1.12, 95% CI 1.08–1.16; P = 4.1 × 10⁻⁹ and OR = 1.13, 95% CI 1.08–1.17; P = 2.8 × 10⁻⁹, respectively. These two SNPs are in strong linkage disequilibrium (LD) (r² = 0.82) and map to a 38-kilobase (kb) region of LD at 20p12.3 (6,316,089–6,354,440 base pairs (bp)), a region bereft of genes or predicted protein-encoding transcripts (Fig. 1 and Supplementary Fig. 2 online). Furthermore, there are no predicted genes or micro-RNAs in the vicinity of the markers (i.e., 250-kb flanking sequence). However, BMP2 maps 342 kb telomeric to the locus, and this may have relevance due to homology and functional similarity to another associated locus (see below).

Figure 2.

Forest plot of effect size and direction for the four SNPs associated with CRC. (a) rs961253. (b) rs4444235. (c) rs10411210. (d) rs9929218. Boxes denote allelic OR point estimates, their areas being proportional to the inverse variance weight of the estimate. Horizontal lines represent 95% CIs. The diamond (and broken line) represents the summary OR computed under a fixed-effects model, with the 95% CI given by its width. The unbroken vertical line is at the null value (OR = 1.0).

The second strongest statistical evidence for an association was for rs4444235, which maps to a 16-kb region of LD at 14q22.2 (53,477,192–53,494,200 bp; combined OR = 1.11, 95% CI 1.08–1.15, P = 8.1 × 10⁻¹⁰, P_het = 0.64, I² = 0%; Table 2 and Fig. 2; excluding phase 1, OR = 1.10, 95% CI 1.06–1.14, P = 2.0 × 10⁻⁷). This SNP is 9.4 kb from the transcription start site of the gene encoding bone morphogenetic protein 4 preproprotein (BMP4; Fig. 1 and Supplementary Fig. 2). Like BMP2, BMP4 is a member of the transforming growth factor-β family of signal transduction molecules that play an important role in CRC¹⁰. BMP signaling inhibits intestinal stem cell self-renewal through suppression of Wnt-β-catenin signaling¹¹. Intriguingly, inactivating mutations in the BMP receptor subunit BMPR1A are one cause of the rare juvenile polyposis syndrome^12,13, which carries a very high risk of CRC, and we have previously found evidence that SNPs close to the BMP antagonist GREM1 are associated with CRC risk⁹.

The third strongest evidence for an association was shown by rs10411210 and rs7259371, two SNPs in moderate LD (r² = 0.54), which map to a 96-kb block of LD (38,203,614–38,300,573 bp) on 19q13.1 encompassing Rho GTPase binding protein 2 (RHPN2) (Fig. 1 and Supplementary Fig. 2). rs10411210 showed the strongest evidence of association in this region, with an OR of 0.83 (95% CI 0.78–0.88; P = 1.1 × 10⁻⁹; P_het = 0.05, I² = 50%; excluding phase 1, OR = 0.85, 95% CI 0.80–0.91; P = 1.0 × 10⁻⁶), whereas for rs7259371 the OR was 0.89 (95% CI 0.85–0.93, P = 2.2 × 10⁻⁷, P_het = 0.03, I² = 55%; excluding phase 1, OR = 0.90, 95% CI 0.86–0.95, P = 2.8 × 10⁻⁵). RHPN2 encodes a Rho GTPase involved in the regulation of the actin cytoskeleton and cell motility¹⁴. Expression of RhoA has been implicated in the biology of several cancers, including CRC, promoting invasiveness¹⁵ through downregulation of adherens junction formation¹⁶.

The fourth locus displaying strong statistical evidence for association was rs9929218 (combined OR = 0.90, 95% CI 0.87–0.94, P = 1.2 × 10⁻⁷, P_het = 0.79, I² = 0%), which localizes to intron 1 of the gene encoding cadherin 1 (CDH1, also known as E-cadherin) and maps to a 110-kb region of LD at 16q22.1 (67,286,613–67,396,803 bp) (Fig. 1 and Supplementary Fig. 2). Another SNP in CDH1, rs1862748 (r² = 0.91 with rs9929218) also showed strong evidence for association (P = 4.9 × 10⁻⁷). CDH1 has an established role in CRC: aberrant activation of Wnt–β-catenin signaling is an initiating event in the development of CRC¹⁷, and somatic inactivation of CDH1 by mutation or promoter methylation occurs frequently¹⁸, leading to increased activity of the β-catenin–TCF transcription factor. Furthermore, germline mutations in CDH1 can cause familial diffuse-type gastric cancer¹⁹, occasionally in association with early-onset CRC^20,21.

Because the association between rs10411210 (RHPN2) and CRC risk showed evidence of between-study heterogeneity, and the association between rs9929218 (CDH1) was of borderline genome-wide significance, we genotyped both SNPs in four additional independent replication case-control series from populations of European ancestry (Studies of Epidemiology and Risk Factors in Cancer Heredity (SEARCH), Canada, Darmkrebs: Chancen der Verhütung durch Screening (DACHS) and Kiel). After pooling data from these studies with the previously genotyped data sets, the association for rs9929218 attained genome-wide significance (OR = 0.91, 95% C: 0.89–0.94, P = 1.2 × 10⁻⁸, P_het = 0.56, I² = 0%; excluding phase 1, P = 4.8 × 10⁻⁷; Table 2 and Fig. 2). The association for rs10411210 was still highly significant, but there remained evidence of between-study heterogeneity, albeit nonsignificant after correction for multiple testing (OR = 0.87, 95% CI: 0.83–0.91, P = 4.6 × 10⁻⁹, P_het = 0.02; I² = 52%; excluding phase 1, P = 1.4 × 10⁻⁶; Table 2 and Fig. 2).

We assessed in more detail the pattern of the risks associated with each of the four newly identified loci, as defined by rs961253, rs4444235, rs10411210 and rs9929218. For rs961253, the minor allele was associated with an increased risk of CRC in a dose-dependent manner, with a higher risk in homozygous than heterozygous carriers (OR_het = 1.14, 95% CI 1.08–1.20 and OR_hom = 1.24, 95% CI 1.15–1.34), supporting a multiplicative model (Supplementary Table 4 online). A similar risk pattern was observed for rs4444235 (OR_het = 1.13, 95% CI 1.07–1.20 and OR_hom = 1.23, 95% CI 1.15–1.32). For rs10411210, the minor allele was associated with a decreased risk of CRC in a dose-dependent manner (OR_het = 0.87, 95% CI 0.83–0.92 and OR_hom = 0.72, 95% CI 0.59–0.89). There was little difference in the fit provided by multiplicative and dominant models due to the very low frequency of rare homozygotes, although a recessive model could be excluded (Supplementary Table 4). For rs9929218, genotype-specific ORs were most compatible with a multiplicative model (OR_het = 0.92, 95% CI 0.89–0.96 and OR_hom = 0.82, 95% CI 0.76–0.89; Supplementary Table 4).

Genotype-phenotype correlations

We assessed associations between clinicopathological variables and genotypes through case-only logistic regression. The association between rs4444235 (BMP4) and CRC was significantly stronger in cases with microsatellite stable tumors compared with microsatellite instability (MSI) cases (P = 2.1 × 10⁻³; based on 2,074 cases from London phase 2 and London replication series; Supplementary Table 5 online). There was also some evidence of an association with gender for rs9929218 (P = 0.02, based on 17,152 cases from all series bar VCQ58 and FCCPS), with the susceptibility allele more common in females than males (Supplementary Table 5). We did not find any other significant associations between SNP genotype and clinicopathological data (specifically, age at diagnosis, site of tumor or family history of CRC; Supplementary Table 5).

Architecture of genetic susceptibility to colorectal cancer

We estimate the contribution of each of the ten loci identified to date to the familial risk of CRC to be <1%. As epistasis could affect the overall contribution of the loci to the genetic susceptibility of CRC, we examined for pairwise interactions. Because the six previously validated SNPs had not been genotyped in all replication series from the current study, the number of individual genotypes available for each calculation varied. However, all comparisons were based on at least 13,000 individuals. This analysis had >80% power to detect an interaction between the two SNPs that have the lowest minor allele frequency (MAF) and smallest effect size, namely rs16892766 and rs10411210. We found no evidence of interactive effects between any of the CRC disease loci identified thus far (Supplementary Table 6 online), suggesting an independent role for each locus in CRC development. Counting two for a homozygote, the risk of CRC increased with possession of an increasing numbers of risk alleles for the ten loci (P_trend = 5.4 × 10⁻⁴⁴; based on 9,692 cases and controls from London and Edinburgh phase 2; Supplementary Table 7 online). On the basis of an additive model, the ten loci so far identified are likely to collectively account for ~6% of the excess familial risk.

To gain insight into the basis of the residual familial risk of CRC, we estimated the power of our meta-analysis to identify each of the ten disease-associated loci on the basis of their MAF and associated genotypic risk (Fig. 3). Although our combined analysis was well powered to identify common variants (MAF > 0.3) such as those at 8q24, the power to identify those loci with lower MAF or smaller genotypic risk was limited.

Figure 3.

Power of the meta-analysis to identify each of the ten colorectal cancer susceptibility alleles, stipulating a P value of 10⁻⁵. Genotypic risks associated with each locus derived from replication data.

Although fine-mapping and resequencing are required to identify the specific variants underlying each of the ten associations so far identified, we conducted two analyses to investigate the basis of causality. Accepting the caveat that HapMap is not comprehensive, we interrogated HapMap to identify nonsynonymous SNPs highly correlated (i.e., r² > 0.8) with the most strongly associated SNP within each region of association. The only nonsynonymous SNPs showing strong LD was rs17563 (BMP4 A152V), which is correlated with rs4444235 (D′ = 0.96, r² = 0.81; Supplementary Table 8 online). However, genotyping both phase 2 case-control series for rs17563 provided no support for A152V as the basis for the 14q22.2 association (P = 0.027 compared with P = 5.90 × 10⁻⁴ for rs4444235). These data suggest many of the associations identified so far are mediated through LD with sequence changes that influence gene expression rather than protein sequence, or through LD with low frequency variants (i.e., variants with MAF of 0.001–0.01) that are not cataloged by HapMap. The C-160A promoter variant (rs16260), which is in strong LD with rs9929218 (D′ = 0.95, r² = 0.92), has been previously documented to influence CDH1 transcription²². We have previously reported allele-specific expression associated with SMAD7 variants⁸. To explore whether any of the other eight associations reflect similar cis-acting regulatory effects on a nearby gene, we searched for genotype–expression correlations in 90 lymphoblastoid cell lines using previously described data^23,24. We did not, however, find any significant relationship between SNP genotype and gene expression (Supplementary Fig. 3 online).

DISCUSSION

By pooling data from two GWA studies and conducting replication analyses, we have identified four previously unreported variants influencing CRC susceptibility, in addition to the six variants we have previously shown to be associated with CRC risk. The new loci identified in the current study are of modest effect size, which is unsurprising given that those with a larger impact on CRC were discovered in previously reported fast-tracking analyses^3,4,6,8,9. Collectively, these data are thus consistent with a model in which the loci readily detectable through current GWA studies of 1,000–2,000 CRC cases and controls are associated with modest effects (i.e., those with MAF > 0.2 and conferring genotypic risks of ~1.2).

Our estimate of the contribution of the ten loci to excess familial risk of CRC is likely to be conservative as the effect of the causal variant will typically be larger than the association detected through a tag SNP. In addition, multiple causal variants may exist at each locus, including low-frequency variants with significantly larger effects on risk. Indeed, our analysis provides evidence that CDH1 may be a locus where a spectrum of common and rare alleles contributes to CRC risk, as reported in other complex traits.

Our GWA data and the results from similar gene discovery efforts in other tumors are proving to be highly informative regarding the allelic architecture of cancer susceptibility in general. First, the number of common variants explaining more than 1% of inherited risk is very low, and only a small proportion of the heritability of any cancer can be explained by the currently identified loci. Second, the genetic landscape defined by the common risk loci identified to date does not feature significant epistatic effects. Third, few of the observed associations seem to be due to correlation with common coding variants, and furthermore, many of the loci map to regions bereft of genes or protein-coding transcripts. It is therefore likely that much of the common variation in cancer risk is mediated through sequence changes influencing gene expression, perhaps in a subtle fashion. Therefore, in addition to searching for cis-regulatory correlations, expression quantitative trait locus analyses may aid elucidation of causality. Whereas examination of genotypic effects on expression in lymphoblastoid cell lines can be informative if genes are ubiquitously expressed, it is likely to be limited by tissue-specific effects. Furthermore, the target tissue need not even be the organ or cell type from which cancer develops. Hence, although we failed to demonstrate effects of cis-regulatory polymorphisms in LD with the variants identified, lymphocytes show either low levels of gene expression (CDH1, CDH3, RHPN2) or absent expression (BMP4, BMP2) of target genes, making analyses based on colorectal or other tissues desirable.

Many mendelian cancer predisposition genes influence the risk of more than one tumor type. Pleiotropic effects are also a feature of 8q24 variants such as rs6983267, which affects the risk of CRC, prostate and ovarian cancer^25,26. To examine whether any of the four variants we have identified influences the risk of breast or prostate cancer, we interrogated GWA data from the Cancer Genetic Markers of Susceptibility (CGEMS) breast and prostate cancer study. Although requiring replication, there was evidence from CGEMS that variation at 20p12 (rs355527) influenced the risk of developing breast cancer.

The power of our study to detect the major common loci conferring risks of 1.2 or greater (such as the 8q24 variant) was high. Hence, we consider that there are unlikely to be many additional CRC SNPs with similar effects for alleles with frequencies >0.2 in populations of European ancestry. However, it is notable that the 10p14 variant rs10795668, which we have previously robustly shown⁵ to be associated with CRC risk (P = 6.9 × 10⁻¹²), was not captured by our meta-analysis, which stipulated a threshold of <10⁻⁵ (P value in meta-analysis 3.1 × 10⁻⁴). In contrast, we had low power to detect alleles with smaller effects and/or MAFs < 0.1. By implication, variants with such profiles are likely to represent a much larger class of susceptibility loci for CRC, whether because of truly small effect sizes or submaximal LD with tagging SNPs. In addition to sample size considerations, the 550K tagging SNPs we used for the GWA studies capture on average ~80% of common SNPs in the European population (i.e., r² > 0.8), but only ~12% of SNPs with MAFs of 5–10% are tagged at this level, limiting power to detect this class of susceptibility allele. Furthermore, GWA-based strategies are not optimally configured to identify low-frequency variants with potentially stronger effects. In addition, these arrays are not ideally formatted to capture copy number variants, which may also affect CRC risk. Thus, it is highly likely that a large number of low-penetrance variants remain to be discovered. This assertion is supported by the continued excess of associations observed over those expected, in addition to the regions studied herein. Further efforts to expand the scale of GWA meta-analyses, in terms of both sample size and SNP coverage, and to increase the number of SNPs taken forward to large-scale replication, may identify additional variants for CRC.

Irrespective of the nature of the causal variants, a high proportion of the population carries at-risk genotypes. Whereas individual alleles exert only small effects, much larger risks are seen in carriers of multiple risk alleles. Therefore, the ten SNPs identified thus far potentially have public health relevance as further susceptibility loci are identified. Additional studies are required to characterize genetic variation at these loci and determine their relationship to the functional consequences that lead to CRC. Furthermore, it will also be intriguing to examine how our findings translate to non-European populations, some of which are typified by a considerably lower prevalence of CRC.

METHODS

Study participants

Supplementary Table 1 provides a summary of all cases and controls in the study.

London genome-wide association study

Phase 1 comprised 940 cases with colorectal neoplasia (443 males, 497 females) ascertained through the CoRGI consortium. All had at least one first-degree relative affected by CRC and one or more of the following phenotypes: CRC at age 75 or less; any colorectal adenoma at age 45 or less; three or more colorectal adenomas at age 75 or less; or a large (>1 cm diameter) or aggressive (villous and/or severely dysplastic) adenoma at age 75 or less, Controls (n = 965; 439 males, 526 females) were spouses or partners unaffected by cancer and without a personal family history (to second-degree relative level) of colorectal neoplasia. All cases and controls were of European ancestry and from the UK. Phase 2 consisted of 2,873 CRC cases (1,199 males, 1,674 females; mean age at diagnosis 59.3 years; s.d. ± 8.7) ascertained through two ongoing initiatives at the Institute of Cancer Research/Royal Marsden Hospital NHS Trust (from 1999 onwards: the NSCCG²⁷ and the Royal Marsden Hospital NHS Trust/Institute of Cancer Research Family History and DNA Registry. A total of 2,871 healthy individuals were recruited as part of ongoing National Cancer Research Network genetic epidemiological studies, NSCCG (n = 1,235), the Genetic Lung Cancer Predisposition Study²⁸ (1999–2004; n = 917) and the Royal Marsden Hospital NHS Trust/Institute of Cancer Research Family History and DNA Registry (1999–2004; n = 719), These controls (1,164 males, 1,707 females; mean age 59.8 years; s.d. ± 10.8) were the spouses or unrelated friends of individuals with malignancies. None had a personal history of malignancy at time of ascertainment. All cases and controls were British of recent European ancestry, and there were no obvious differences in the demography of cases and controls in terms of place of residence within the UK. Genotyping data for these samples have been presented previously⁵.

Edinburgh genome-wide association study

Phase 1 included 1,012 CRC cases (518 males, 494 females; mean age at diagnosis 49.6 years; s.d. ± 6.1) and 1,012 age- and gender-matched cancer-free population controls (518 males, 494 females; mean age 51.0 years; s.d. ± 5.9). Cases were enriched for genetic etiology by early age at onset (age ≤ 55 years). Known dominant polyposis syndromes, hereditary nonpolyposis colrectal carcinoma/Lynch syndrome or bi-allelic MUTYH mutation carriers were excluded. Control subjects were population controls, matched by age (± 5 years), gender and area of residence within Scotland. Phase 2 comprised 2,057 CRC cases (1,249 males, 808 females; mean age at diagnosis 65.8 years; s.d. ± 8.4) and 2,111 population controls (1,257 males, 854 females; mean age 67.9 years; s.d. ± 9.0) ascertained in Scotland. Cases were taken from an independent, prospective, incident CRC case series and aged <80 years at diagnosis. Control subjects were population controls matched by age (± 5 years), gender and area of residence within Scotland. Genotyping data for phase 1 and for 15,008 SNPs in phase 2 have been presented previously⁶.

Replication series

London replication: 3,286 CRC cases (2,158 males, 1,128 females; mean age at diagnosis 59.1 years; s.d. ± 8.1) and 3,017 controls (1,212 males, 1,805 females; mean age 61.7 years; s.d. ± 11.4) ascertained through NSCCG post 2005. Edinburgh replication: 676 CRC cases (335 males, 341 females; mean age at diagnosis 53.2 years; s.d. ± 15.4) and 842 cancer-free population controls (394 males, 448 females; mean age 51.8 years; s.d. ± 11.5). Controls were recruited as part of the Generation Scotland study²⁹. VCQ58: 1,543 CRC cases (925 males, 618 females, mean age of diagnosis 62.4 years; s.d. ± 10.7), consisting of 1,234 (2 SNPS typed for 1,310) samples from the VIOXX in Colorectal Cancer Therapy: Definition of Optimal Regime (VICTOR)/QUASAR2 trials and 309 CRC cases collected through the CoRGI study. There were 2,236 controls (1,007 males, 1,226 females, three unknown) consisting of 1,438 population controls from the Wellcome Trust Case Control Consortium 1958 birth cohort (58C, also known as the National Child Development Study, which included all births in England, Wales and Scotland during a single week in 1958 (ref. ³⁰), 418 cancer-free controls collected through the CoRGI study and 380 European Collection of Cell Cultures samples. All cases and controls were of European origin and from the UK. FCCPS: 962 CRC cases (509 males, 452 females, one unknown; mean age at diagnosis 66.9 years; s.d. ± 12.2) and 846 controls (randomly selected anonymous Finnish blood donors) ascertained in southeastern Finland. SEARCH: 2,222 CRC cases (1,278 males, 944 females; mean age at diagnosis 59.2 years; s.d. ± 8.1) and 2,262 controls (949 males, 1,313 females; mean age 57,6 years; s.d. ± 15.1. Samples were ascertained through the SEARCH study based in Cambridge, UK. Recruitment of CRC started in 2000; initial patient contact was though the general practitioner. Control samples were collected post 2003. Eligible individuals were sex and frequency matched in 5-year age bands to cases. DACHS: 1,373 CRC cases (790 males, 583 females; mean age at diagnosis 68.1 years; s.d. ± 10.4) and 1,480 controls (719 males, 761 females; mean age 68.0 years; s.d. ± 9.9) ascertained through DACHS, a population-based case-control study of incident CRC in the Rhine-Neckar-Odenwald region around Heidelberg between 2003 and 2006. Kiel: 2,169 CRC cases (1,089 males, 1,080 females; mean age at diagnosis 60.9 years; s.d. ± 8.8) and 2,145 controls (1,059 males, 1,086 females; mean age 64.7 years; s.d. ± 10.0) ascertained through the PopGen and SHIP population-based biobank projects based in Kiel and Greifswald, Germany. Canada: 1,175 CRC cases (503 males, 672 females; mean age at diagnosis 60.3 years; s.d. ± 8.7) and 1,184 controls (667 males, 517 females; mean age 60.9 years; s.d. ± 8.1) ascertained through the Ontario Familial Colorectal Cancer Registry.

In all cases, CRC was defined according to the ninth revision of the International Classification of Diseases³¹ by codes 153–154, and all cases had pathologically proven adenocarcinoma. Collection of blood samples and clinicopathological information from cases and controls was undertaken with informed consent and ethical review board approval in accordance with the tenets of the Declaration of Helsinki.

Genotyping

DNA was extracted from samples using conventional methodologies and quantified using PicoGreen (Invitrogen). The London phase 1 GWA study was conducted using the Illumina HumanHap550 Bead Arrays, and the Edinburgh phase 1 GWA study was conducted using Illumina HumanHap300 and HumanHap240S according to the manufacturer’s protocols (Supplementary Methods online). DNA samples with GenCall scores <0.25 at any locus were considered ‘no calls’. In London and Edinburgh phase 2, genotyping was conducted using Illumina Infinium custom arrays according to the manufacturer’s protocols. For both phases 1 and 2, a SNP was deemed to have failed if fewer than 95% of DNA samples generated a genotype at the locus. To ensure quality of genotyping, a series of duplicate samples were genotyped, and cases and controls were genotyped in the same batches in both phases 1 and 2.

Phase 3 genotyping was conducted using either competitive allele-specific PCR KASPar chemistry (KBiosciences) or Taqman (Applied Biosystems), except for Canadian samples, which were genotyped using single-base primer extension chemistry matrix-assisted laser desorption ionization time of flight mass spectrometry (MALDI-TOF MS) detection (Sequenom). All primers and probes used are available on request. Genotyping quality control was tested using duplicate DNA samples within studies and SNP assays, together with direct sequencing of subsets of samples to confirm genotyping accuracy. For all SNPs, >99.9% concordant results were obtained.

MSI in CRCs was determined using the following methodology: 10 µm sections were cut from formalin-fixed paraffin-embedded tumors, lightly stained with toluidine blue, and regions containing at least 60% tumor microdissected. Tumor DNA was extracted using the QIAamp DNA Mini kit (Qiagen) according to the manufacturer’s instructions and genotyped for the mononucleotide microsatellite loci BAT25 and BAT26, which are highly sensitive markers of MSI³². Samples showing novel alleles at either BAT26 or BAT25 or both markers were assigned as MSI (corresponding to a high level of instability³³).

Statistical analysis

Genotype data were used to search for duplicates and closely related individuals among all samples in phases 1 and 2. Identity-by-state values were calculated for each pair of individuals; for any pair with allele sharing >80%, the sample generating the lowest call rate was removed from further analysis.

In London phase 1, genotyped samples were excluded from analyses for the following reasons: carriers of another susceptibility allele (five cases); first-degree relative with CRC (11 controls); duplicated (two cases, seven controls); relatedness (one case, 15 controls). In London phase 2, genotyped samples were excluded from analyses for the following reasons: duplicated (eight cases, two controls); relatedness (two cases, 18 controls); gender discrepancies (13 controls). In Edinburgh phases 1 and 2, genotyped samples were excluded from analyses for the following reasons: identified as of non-European descent (seven in phase 1, three in phase 2); previously unrecognized carriers of known monogenic hereditary CRC syndromes (five DNA MMR gene mutation carriers in phase 2); gender discrepancies between records and genotype (14 in phase 1, 22 in phase 2); hidden relatedness (five in phase 1).

The adequacy of the case-control matching and possibility of differential genotyping of cases and controls was formally evaluated using Q-Q plots of test statistics. The inflation factor λ was calculated by dividing the mean of the lower 90% of the test statistics by the mean of the lower 90% of the expected values from a χ² distribution with 1 d.f. Deviation of the genotype frequencies in the controls from those expected under Hardy–Weinberg equilibrium was assessed by χ² test (1 d.f.), or Fisher’s exact test where an expected cell count was <5.

Association between SNP genotype and disease status was primarily assessed using the allelic 1 d.f. test or Fisher’s exact test where an expected cell count was <5. The risks associated with each SNP were estimated by allelic, heterozygous and homozygous ORs using unconditional logistic regression, and associated 95% CIs were calculated in each case.

Joint analysis of data generated from multiple phases was conducted using standard methods for combining raw data based on the Mantel-Haenszel method³⁴. Joint ORs and 95% CIs were calculated assuming a fixed-effects model, and tests of significance of the pooled effect sizes were calculated using a standard normal distribution. Cochran’s Q statistic to test for heterogeneity and the I² statistic³⁵ to quantify the proportion of the total variation due to heterogeneity were calculated.

We used Haploview software (v3.2) to infer the LD structure of the genome in the regions containing loci associated with disease risk. Patterns of risk for associated SNPs were investigated by logistic regression, coding the SNP genotypes according to additive, dominant and recessive models. Models were then compared by calculating the Akaike information criterion and Akaike weights for each mode of inheritance. Associations by site (colon/rectum), MSI status, family history status (at least one first-degree relative with CRC), gender and age at diagnosis (stratifying into two groups by the median age at diagnosis) were examined by logistic regression in case-only analyses, using all cases from replication phases for whom the clinicopathological variable being tested was available. Results for gender and age at diagnosis were based on all case series apart from London phase 1, VCQ58 and FCCPS; for site, on data from London phase 2, Edinburgh phases 1 and 2, London replication, SEARCH, Canada, DACHS and Kiel; for family history status and MSI status, on London phase 2 and London replication. The combined effect of each pair of loci identified as associated with CRC risk was investigated by logistic regression modeling, and evidence for interactive effects between SNPs assessed by likelihood ratio test assuming an allelic model. The OR and trend test for increasing numbers of deleterious alleles was estimated based on the London and Edinburgh phase 2 data by counting two for a homozygote and one for a heterozygote.

The sibling relative risk attributable to a given SNP was calculated using the formula^36,37

$${\lambda }^{*}=\frac{p{(p{r}_2+q{r}_1)}^2+q{(p{r}_1+q)}^2}{{(p^2{r}_2+2pq{r}_1+q^2)}^2}$$

where p is the population frequency of the minor allele, q = 1 − p and r₁ and r₂ are the relative risks (estimated as OR) for heterozygotes and rare homozygotes relative to common homozygotes. Assuming a multiplicative interaction, the proportion of the familial risk attributable to a SNP was calculated as log(λ^*)/log(λ₀), where λ₀ is the overall familial relative risk estimated from epidemiological studies, assumed to be 2.2 (ref. ³⁸). A naïve estimation of the contribution of all of the loci identified to the excess familial risk of CRC under an additive model was calculated using the formula³⁹

$$\frac{{\text{OR}}_{\text{per allele}}-1}{2({\lambda }_{\mathrm{o}}-1)}$$

We based our estimate of study power on a joint analysis of phases 1 and 2 assuming a multiplicative model for each SNP⁴⁰. Samples in phase 1 of both GWAs were potentially enriched for genetic susceptibility by virtue of either having familial disease or being diagnosed young. To estimate the power of the meta-analysis, we therefore also computed power estimates accounting for genetic enrichment. For familial cases, the sample size required to detect a common disease susceptibility allele is typically reduced by more than twofold⁴¹. Similar statistical considerations apply to early-onset cases. Hence, for the power calculation, we conservatively inflated phase 1 case samples size by a factor of 2. The power to identify epistatic interactions between SNPs was estimated assuming a multiplicative model of gene–gene interaction using the Lubin and Gail approach^42,43.

Relationship between SNP genotypes and expression levels

To examine for a relationship between SNP genotype and expression levels of MYC, LOC120376, CDH1, CDH3, BMP4, BMP2 and RHPN2 in lymphocytes, we made use of publicly available expression data generated from analysis of 90 European-derived Epstein-Barr virus–transformed lymphoblastoid cell lines using Sentrix Human-6 Expression BeadChips (Illumina)^23,24. Online recovery of data was done using WGAViewer Version 1.25 Software. We compared differences in the distribution of levels of mRNA expression between SNP genotypes using a Wilcoxon-type test for trend⁴⁴.

URLs

Haploview, http://www.broad.mit.edu/personal/jcbarret/haploview/; QUASAR, http://www.octo-oxford.org.uk/alltrials/trials/q2.html; European Collection of Cell Cultures, http://www.hpacultures.org.uk; WGAViewer, http://www.genome.duke.edu/centers/pg2/downloads/wgaviewer.php; Genetic Lung Cancer Predisposition Study, http://pfsearch.ukcrn.org.uk/StudyDetail.aspx?TopicID=1&StudyID=781, http://www.dh.gov.uk/assetRoot/04/01/45/13/04014513.pdf; NSCCG, http://pfsearch.ukcrn.org.uk/StudyDetail.aspx?TopicID=1&StudyID=1269; 1958 Birth Cohort, http://www.cls.ioe.ac.uk/studies.asp?section=000100020003.

Appendix

Note: Supplementary information is available on the Nature Genetics website.

AUTHOR CONTRIBUTIONS

The study was designed and financial support obtained by R.S.H., I.P.M.T., M.G.D. and H.C. The manuscript was drafted by R.S.H. and E.W. with contributions from M.G.D. and I.P.M.T. Statistical analyses were conducted by E.W., with contributions from M.C.D.B., J.-B.C., S.L.S. and A.T. E.W. and A.M.P. performed bioinformatic analyses.

Institute of Cancer Research: Subject recruitment and sample acquisition to NSCCG were undertaken by S.P. Coordination of sample preparation and genotyping was performed by P.B. Sample preparation and genotyping performed by A.M.P., K.S., J.V. and S.L. Histology review was by I.C. Testing of MSI was performed by S.L. and I.C.

Wellcome Trust Centre for Human Genetics: Subject recruitment and sample acquisition were undertaken by E.B., M.G., L.M. and members of the CoRGI Consortium and D.K. Sample preparation was performed by K.H., S.L.S. and E.J. Genotyping was performed and co-ordinated by L.C.-C., K.H., A.S.T., S.L.S., A.W., E.J. and E.D.

Colon Cancer Genetics Group, Edinburgh: Study design and funding: M.G.D., H.C., A.T., S.M.F. and M.E.P. Subject recruitment and sample acquisition was led by M.G.D., H.C., M.E.P., R.C., S.M.F. and members of the SOCCS/COGS study teams. Sample preparation was co-ordinated by S.M.F. and R.A.B. Histology review: M.G.D. and S.M.F. Genotyping was performed and co-ordinated by S.M.F. and M.G.D. A.T. performed statistical analyses. J.G.P. performed bioinformatic analyses. The following authors from the various collaborating groups conceived the local study, undertook assembly of case/control series in their respective regions, collected data and samples, variously undertook genotyping and analysis: I.N., S.T., A.K. and L.A.A. in Finland; T.K. and P.D.P. in Cambridge; S.S., H.V., S.B., C.S., J.T. and J.H. in Kiel; J.C.-C., M.H. and H.B. in Heidelberg; B.W.Z., A.M., T.J.H. and S.G., in Toronto and Montreal. All undertook sample collection and phenotype data collection and collation in the respective centers.

Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/

The full list of authors and affiliations is as follows:

Group 1: Richard S Houlston², Emily Webb², Peter Broderick², Alan M Pittman², Maria Chiara Di Bernardo², Steven Lubbe², Ian Chandler², Jayaram Vijayakrishnan², Kate Sullivan², Steven Penegar² & Colorectal Cancer Association Study Consortium²⁰

Group 2: Luis Carvajal-Carmona³, Kimberley Howarth³, Emma Jaeger³, Sarah L Spain³, Axel Walther³, Ella Barclay³, Lynn Martin³, Maggie Gorman³, Enric Domingo³, Ana S Teixeira³, CoRGI Consortium²⁰, David Kerr⁴, Jean-Baptiste Cazier⁵, Iina Niittymäki⁶, Sari Tuupanen⁶, Auli Karhu⁶, Lauri A Aaltonen⁶ & Ian P M Tomlinson³

Group 3: Susan M Farrington⁷, Albert Tenesa⁷, James G D Prendergast⁷, Rebecca A Barnetson⁷, Roseanne Cetnarskyj⁸, Mary E Porteous⁸, Paul D P Pharoah⁹, Thibaud Koessler⁹, Jochen Hampe¹⁰, Stephan Buch¹⁰, Clemens Schafmayer^11,12, Jurgen Tepel¹², Stefan Schreiber^11,13, Henry Völzke¹⁴, Jenny Chang-Claude¹⁵, Michael Hoffmeister¹⁶, Hermann Brenner¹⁶, Brent W Zanke¹⁷, Alexandre Montpetit¹⁸, Thomas J Hudson^18,19, Steven Gallinger¹⁹, International Colorectal Cancer Genetic Association Consortium²⁰, Harry Campbell⁷ & Malcolm G Dunlop⁷

²Section of Cancer Genetics, Institute of Cancer Research, Sutton, SM2 5NG, UK. ³Molecular and Population Genetics, Wellcome Trust Centre for Human Genetics, Roosevelt Drive, Oxford OX3 7BN, UK. ⁴Department of Clinical Pharmacology, Oxford University, Old Road Campus Research Building, Oxford, OX2 6HA, UK. ⁵Bioinformatics and Biostatistics, Wellcome Trust Centre for Human Genetics, Roosevelt Drive, Oxford OX3 7BN, UK. ⁶Department of Medical Genetics, Genome-Scale Biology Research Program, Biomedicum Helsinki, University of Helsinki, Helsinki, Finland. ⁷Colon Cancer Genetics Group, Institute of Genetics and Molecular Medicine, University of Edinburgh and MRC Human Genetics Unit, Edinburgh EH4 2XU, UK. ⁸Clinical Genetics, Western General Hospital, Edinburgh EH4 2XU, UK. ⁹Cancer Research UK Laboratories, Strangeways Research Laboratory Department of Oncology, University of Cambridge, Cambridge, UK. ¹⁰Department of General Internal Medicine, University Hospital, Schleswig-Holstein, Campus Kiel, Schittenhelmstraβe 12, 24105 Kiel, Germany. ¹¹POPGEN Biobank, University Hospital Schleswig-Holstein, Campus Kiel, Schittenhelmstrasse 12, 24105 Kiel, Germany. ¹²Department of General and Thoracic Surgery, University Hospital Schleswig-Holstein, Campus Kiel, Arnold-Heller-Strasse, 24105 Kiel, Germany. ¹³Institute for Clinical Molecular Biology, University Hospital Schleswig-Holstein, Campus Kiel, Schittenhelmstrasse 12, 24105 Kiel, Germany. ¹⁴Institut für Epidemiologie und Sozialmedizin, University Hospital Greifswald, Walther-Rathenau-Strasse 48 Greifswald 17487, Germany. ¹⁵Division of Cancer Epidemiology, German Cancer Research Center (DKFZ) and ¹⁶Division of Clinical Epidemiology and Aging Research, German Cancer Research Center (DKFZ), Im Neuenheimer Feld 280, 69120 Heidelberg, Germany. ¹⁷The Ottawa Health Research Institute, Division of Epidemiology, 501 Smythe Road, Ottawa, Canada, K1H 8L6. ¹⁸The McGill University and Genome Quebec Innovation Centre, 740 Dr Penfield Avenue, Montreal Quebec, Canada H3G 1A4. ¹⁹Cancer Care Ontario, 620 University Avenue, Toronto Ontario M5G 1L7 and Ontario Institute for Cancer Research, 101 College Street, Toronto, Canada M5G 2L7. ²⁰A full list of members is provided in the Supplementary Note online.