Meta-analysis of three genome-wide association studies identifies susceptibility loci for colorectal cancer at 1q41, 3q26.2, 12q13.13 and 20q13.33

Genome-wide association (GWA) studies of colorectal cancer (CRC) have vindicated the hypothesis that part of the heritable risk is caused by common, low-risk variants ¹. Our previous analyses, based on two GWA studies from the UK (UK1/CORGI) and Scotland (Scotland1/COGS) have identified 10 common variants that are associated with CRC risk ². These variants map to 8q24.21 (rs6983267), 8q23.3 (rs16892766, EIF3H), 10p14 (rs10795668), 11q23 (rs3802842), 14q22.2 (rs4444235, BMP4), 15q13 (rs4779584), 16q22.1 (rs9929218, CDH1), 18q21.1 (rs4939827, SMAD7), 19q13.1 (rs10411210, RHPN2) and 20p12.3 (rs961253).

The discovered effect sizes of individual associations and the need for stringent thresholds for establishing statistical significance inevitably constrain the power of individual GWA studies to detect common variants. In order to augment our ability to detect additional CRC loci, we have undertaken a further GWA analysis of a set of cases from the VICTOR and QUASAR2 clinical trials of adjuvant therapy in potentially curable colorectal carcinoma These trials recruited patients from throughout the United Kingdom. The controls comprised a UK population-based 1958 Birth Cohort (58BC) for which genotype data are publicly available. Together, this case-control set (henceforth referred to as VQ58) comprised 1,432 cases and 2,697 controls.

The VQ58 cases were genotyped in-house using the Illumina Hap300/370 SNP arrays. After filtering of both the VQ data and the publicly-available control data to remove SNPs and individuals that fell below pre-determined quality control standards (see Methods), we examined associations between genotype and CRC status. A Q-Q plot demonstrated no evidence of systematic inflation of the allelic test statistic (λ_gc=1.018). No individual SNP showed a significant association with CRC under dominant, additive or recessive models at genome-wide significance (set at P ≤ 1.0x10^-7, based on a Bonferroni correction). This was not unexpected, given the power of the VQ58 set to detect associations of the magnitudes found in our previous analyses of the UK and Scottish GWA studies ². We therefore directly proceeded to a combined analysis of the three GWA studies, comprising UK1/CORGI and Scotland1/COGS in addition to VQ58 (Supplementary Table 1). Quality control measures were standardised throughout the sample sets. We used principal components analysis (PCA) to examine whether there was evidence of distinct genetic sub-groups within the three GWA studies. After removal of 88 outliers and six duplicate samples, the Scottish and UK (UK1/CORGI and VQ58) samples essentially clustered together, minor variation in the first component reflecting the known North-West to South-East cline in the UK (Supplementary Figure 1).

The UK1 and Scotland1 samples had been genotyped using Illumina Hap550 arrays. We therefore imputed genotype probabilities in the VICTOR and QUASAR2 samples at SNPs not present on the Hap300/370 arrays. 94,867 of 214,649 imputed SNPs passed our threshold of ≤ 5% missing genotypes and an Information Score ≥ 0.5. We then conducted a meta-analysis of the three data sets (Supplementary Table 1) using the Mantel-Haenszel method under fixed- and random-effects models. Only one SNP (rs4939827, chromosome 18q21.1), previously shown to be associated with CRC risk ^3–5, achieved formal genome-wide significance for association.

At this stage, we considered whether to include in the meta-analysis data we had generated from two additional, large UK case-control sets: UK2/NSCCG (2,854 cases and 2,822 controls) and Scotland2/SOCCS (2,024 cases and 2,092 controls ) (Supplementary Table 1). These additional samples had been genotyped at 55,000 SNPs with the strongest evidence of association from meta-analysis of UK1/CORGI+Scotland1/COGS GWA studies ². If we were to include these extra data, essentially we had to weigh two factors, (i) the extra power afforded by UK2/NSCCG and Scotland2/SOCCS versus (ii) the probability that a true CRC SNP had not been taken forward into the top 55,000 from the UK1+Scotland1 meta-analysis, but did make it into a (smaller) set of “top” SNPs in a VQ58+UK1+Scotland1 meta-analysis. Power calculations showed that, except for rare alleles with small effects for which the power of detection was in any event low, the extra power provided by the UK2 and Scotland2 samples more than compensated for the loss of a few truly disease-associated SNPs that would not have reached the significance threshold for genotyping in UK2 and Scotland2 (Supplementary Figure 2).

We therefore undertook a meta-analysis of VQ58, UK1/CORGI, Scotland1/COGS, UK2/NSCCG and Scotland2/SOCCS (Figure 1). Seven SNPs achieved formally significant associations (P<10^-7). All these SNPs had previously been shown to be associated with CRC risk. After exclusion of SNPs in strong pairwise LD (r²>0.7), we selected seven SNPs (rs11805285, rs6687758, rs6691170, rs10936599, rs7136702, rs11169552, rs4925386) with nominal associations at P<5.0x10^-5 All of these SNPs had been genotyped, rather than imputed, in VQ. The 7 SNPs underwent validation testing in 9,883 CRC cases and 10,655 controls from six independent, northern European case-control series (COIN/NBS, Helsinki, UK3/NSCCG, UK4/CORGI2BCD, Scotland3/SOCCS and Cambridge; Supplementary Table 1). This threshold for follow-up did not exclude the possibility that other SNPs represented genuine association signals, but was simply a pragmatic strategy for prioritising replication. After replication, significant associations were confirmed for six SNPs mapping to four loci: rs6687758 (P=2.27x10^-9) and rs6691170 (P=9.55x10^-10) at 1q41, rs10936599 (P=3.39x10^-8) at 3q36.2, rs7136702 (P=4.02=x10^-8) and rs11169552 (P=1.89x10^-10) at 12q13.13 and rs4925386 (P=1.89x10^-10) at 20q13.3 (Figure 2, Table 1, Supplementary Table 2). There was no significant between-study heterogeneity for these SNP associations (P_het >0.05 for all SNPs, Table 1) and no SNP showed any evidence of association with age or sex in any data set (P>0.05).

rs6691170 (chr1:220,112,069) and rs6687758 (chr1:220,231,571) lie 125kb from each other on chromosome 1q41 (Table 1). The region containing these two SNPs (Figure 3) is flanked by recombination hotspots close to rs3003888 (chr1:220,049,548) and rs6687797 (chr1:220,296,043). Between these sites, LD relationships are complex and blocks are not easily defined, although a minor recombination hotspot exists at chr1:220,137,516, between rs6691170 and rs66867758. rs6691170 and rs66867758 respectively lie 250kb and 125kb upstream of DUSP10, a dual-specificity phosphatase that inactivates p38 and SAPK/JNK. The region otherwise contains few genes, but several spliced ESTs. In the UK data sets, rs6691170 and rs6687758 were in modest pairwise LD (r²=0.22; D’=0.71) raising the possibility that these SNPs may represent independent signals of association. We assessed this using multiple logistic regression analysis stratified by sample series in which genotypes at one SNP were assessed conditional on those at the other SNP. We found that rs6691170 was associated with an odds ratio [OR]=1.07 (P=6.15x10^-5) and rs6687758 with OR of 1.06 (P=1.92x10^-4). Individuals with the high-risk haplotype (TG) at rs6691170 and rs6687758 had a 1.15-fold increased risk of CRC compared with the low-risk haplotype (GA) (P=5.39x10^-8).

rs10936599 (chr3:170,974,795) is flanked by recombination hotspots at chr3:170,837,364 and chr3:171,082,143 (Figure 3). rs10936599 lies on chr3q26.2, within the myoneurin (MYNN/OZSF) gene which encodes a zinc finger protein of unknown function that is expressed principally in muscle. rs10936599 is also close to the actin-related protein M1 and the TERC telomerase loci.

rs7136702 (chr12: 49,166,483) and rs11169552 (chr12:49,441,930) lie about 275kb apart, within what is essentially a large, poorly-defined haplotype block (Figure 3), composed of a set of smaller blocks, but with considerable long-range LD between markers (chr12:48,658,293-49,505,968). rs7136702 is just telomeric to the myeloproliferative oncogene binding-protein gene LARP4 and 30kb proximal to disco-interacting protein 2B (DIP2B), which may have a role in determining epithelial cell fate. rs11169552 is just telomeric to DIP2B, and proximal to activating transcription factor 1 (ATF1) . ATF1 is the 3’ partner in recurrent translocations with the EWSR1 gene (chr22q12) that contribute to the development of soft tissue clear cell sarcomas ⁶. rs7136702 and rs11169552 map close to a known chromosomal fragile site, but we have found that colorectal tumours rarely show somatic chromosomal breakpoints at this site ⁷. rs7136702 and rs11169552 are not strongly correlated (r²=0.11, D’=0.76 in the UK samples). We therefore tested independence of these signals using conditioned logistic regression analysis as for the chromosome 1 signals. In this combined analysis, the rs11169552 signal nearly retained global significance (OR=0.91, P=4.33x10^-7), whereas the strength of association at rs7136702 was reduced (OR=1.06, P=4.34x10^-4). Individuals with the high-risk haplotype (TC) at rs7136702 and rs11169552 had a 1.14-fold increased risk of CRC compared with the low-risk haplotype (CT) (P=6.90x10^-8).

rs4925386 (chr20:60,354,439) is within a very small haplotype block (chr20:60,330,882-60,355,038), although it shows moderate LD with distal markers outside the block (Figure 3). rs4925386 lies within the large laminin A5 (LAMA5) gene which is required for the production of noggin, a secreted BMP antagonist. It is notable that other BMP pathway SNPs are likely to be involved in CRC predisposition ². rs4935386 is in moderate/strong LD (r²>0.5) with four non-synonymous SNPs, which lead to substitutions Ala1908Thr, Arg2226His, Asp2062Asn and Val1900Met, although all of these are predicted to be benign changes.

For both 1q41 and 12q13.12, the two signals, if independent, might have resulted from two causal variants or from a single causal variant strongly associated with disease and correlated with both SNPs in the region. For each region, we addressed the latter possibility by imputing SNPs from the HapMap2 CEU samples between the flanking recombination hotspots. We conducted logistic regression analysis of GWA studies and, UK2/NSCCG and Scotland2/SOCCS, conditioning on the genotypes at each of the two identified SNPs. Although a small number of imputed SNPs from 12q13.12 had a stronger predicted association than the genotyped SNPs (Supplementary Figure 3), no single imputed SNP was able to account for the dual signals in either the 1q41 or 12q13.12 region.

In order to explore whether any of these novel CRC associations resulted from cis-acting regulatory elements, we examined whether any of the 6 SNPs tagged reported expression Quantitative Trait Loci (eQTLs) for nearby genes. Although four SNPs had no association with known eQTLs, rs7136702 was in moderate-strong LD (r²=0.47-0.61, D’=0.80-0.84) with four SNPs (rs11169520, rs11169524, rs3742062 and rs2280503) that had been associated with DIP2B expression in lymphoblastoid cell lines ⁸. Furthermore, rs492536 was in moderate/strong LD (r²=0.61, D’=0.78) with rs13043313, an eQTL for LAMA5 expression in the liver ⁷.

Using a case-only design, we searched for pairwise gene-gene interactions between the 6 new CRC susceptibility SNPs and between these 6 SNPs and the 10 previously identified risk SNPs (rs6983267, rs16892766, rs10795668, rs3802842, rs4444235, rs4779584, rs9929218, rs4939827, rs10411210 and rs961253) ². Although there was suggestive evidence of epistasis between rs6687758 and rs7136702 (P=7.70x10^-4), this did not meet the threshold for formal significance after adjustment for 120 comparisons (P=0.09). There was no evidence to suggest any functional relationships between genes close to these SNPs. No other evidence of gene-gene interactions was found (details not shown).

We have identified four new CRC predisposition loci, none of which maps to previously reported cancer-predisposition genes of high or low penetrance. At two of these loci, there exists the possibility that two SNPs independently predict risk. Our study illustrates other general issues that currently affect large-scale studies to identify common predisposition alleles. Allelic ORs of CRC were less than 1.10 for each of the CRC SNPs we identified. Power to detect the effects of such loci was therefore modest, the likelihood of discovery being highly sensitive to small chance differences in genotype frequencies, especially in the three GWA study data sets. Therefore, many more CRC loci of similar effect size may exist. While the new CRC risk alleles we have identified collectively account for ~1.5% of the familial CRC risk, in concert with other alleles they have potential to impact significantly on disease risk and thus have application to risk stratification at a population level. Finally, the loci we have identified are likely to provide fresh insights into the aetiological basis of CRC.