Genotypic variants at 2q33 and risk of esophageal squamous cell carcinoma in China: a meta-analysis of genome-wide association studies

Abstract

Genome-wide association studies have identified susceptibility loci for esophageal squamous cell carcinoma (ESCC). We conducted a meta-analysis of all single-nucleotide polymorphisms (SNPs) that showed nominally significant P-values in two previously published genome-wide scans that included a total of 2961 ESCC cases and 3400 controls. The meta-analysis revealed five SNPs at 2q33 with P< 5 × 10(-8), and the strongest signal was rs13016963, with a combined odds ratio (95% confidence interval) of 1.29 (1.19-1.40) and P= 7.63 × 10(-10). An imputation analysis of 4304 SNPs at 2q33 suggested a single association signal, and the strongest imputed SNP associations were similar to those from the genotyped SNPs. We conducted an ancestral recombination graph analysis with 53 SNPs to identify one or more haplotypes that harbor the variants directly responsible for the detected association signal. This showed that the five SNPs exist in a single haplotype along with 45 imputed SNPs in strong linkage disequilibrium, and the strongest candidate was rs10201587, one of the genotyped SNPs. Our meta-analysis found genome-wide significant SNPs at 2q33 that map to the CASP8/ALS2CR12/TRAK2 gene region. Variants in CASP8 have been extensively studied across a spectrum of cancers with mixed results. The locus we identified appears to be distinct from the widely studied rs3834129 and rs1045485 SNPs in CASP8. Future studies of esophageal and other cancers should focus on comprehensive sequencing of this 2q33 locus and functional analysis of rs13016963 and rs10201587 and other strongly correlated variants.

Figure 1.

Association results, recombination and LD plots for the region of 2q33 with risk of ESCC. P-values derived from 1 df trend tests across a region of 2q33.1 bounded by rs12693932 and rs3731707, a distance of 473 kb, were plotted. The five colored-line graphs in the upper panel show likelihood ratio statistics (Y-axis on the right) for recombination hotspots from the SequenceLDhot software. The top horizontal line indicates a P-value of 5.0 × 10⁻⁸, and the bottom horizontal line indicates a likelihood ratio statistic cut-off to predict the presence of a hotspot with a false-positive rate of 1 in 3700 independent tests (32). The five different colored lines represent five independent samplings used to estimate the location of the hotspots. The trend P-values from the Chinese, NCI and combined samples were plotted in green triangles, red circles and blue diamonds, respectively. The bottom panel depicts the LD pattern of the region in r², and solid black arrows indicate two flanking recombination hotspots containing five SNPs that exceed genome-wide significance (P-values <5 × 10⁻⁸). The short red vertical lines on the LD heat map indicate the locations of the five genome-wide significant SNPs. The sets of black arrows point to the two recombination hotspots determined in the randomly selected subsets of controls. Since the two panels are on different scales, each has a set of arrows to indicate the hotspots.

Figure 2.

Comparison of results for genotyped (in black) and imputed (in grey) SNPs at 2q33 for their association with risk of ESCC. We plotted P-values for 426 genotyped SNPs and 3878 imputed SNPs. For the imputation, we used a hybrid reference of 1000 Genomes Asian set and the Asian component of the DCEG Reference Imputation Set. This figure shows that the genotyped SNPs were as strong as any SNPs in the imputation analysis.

Figure 3.

Comparison of GWAS and Margarita P-values for 40 SNPs at 2q33. The GWAS P-values are uncorrected, whereas the Margarita P-values (14) come from permutation tests and do not need correction for multiple comparisons. The figure includes data from 40 SNPs and extends from rs12470378 to rs10931959. Both methods show the strongest signal at rs10201587.

Figure 4.

Haplotypes for 53 SNPs at 2q33 and risk of ESCC. Thirteen monomorphic SNPs among all the haplotypes with frequency >1% were excluded from the figure. Haplotype pairs for each subject were inferred using PHASE v2 (15) and used to generate 100 genealogies. ARG result for haplotypes with a frequency >1% containing five SNPs (rs10201587, rs3769823, rs10931936, rs13016963 and rs9288318) showed complete separation between those predicted to be associated with cancer (red) or not (green). All prediction frequencies were 0 or 1.0. rs10201587 was always the best segregator, but in these haplotypes the other four SNPs were in near perfect LD with this SNP. The ARG model prediction suggests that rs10201587 (allele frequency 28.9%) is the most likely SNP, or in strong LD with the functional variant.