Systematic Meta-Analyses and Field Synopsis of Genetic Association Studies in Colorectal Cancer

Abstract

Colorectal cancer is a major global public health problem, with approximately 950 000 patients newly diagnosed each year (www.who.int; International Agency for Research on Cancer GLOBOCAN). The risk of developing colorectal cancer increases steeply with age and its incidence is rising in many industrialized countries as life expectancy and the number of elderly people increase. Incidence is also rising in many developing countries as diet and lifestyle change and increasingly resemble those in industrialized countries. Overall, 5-year survival of colorectal cancer remains at best only approximately 50%. Because of the rising disease burden, health care costs, including costs for diagnosis and treatment, are substantial.

Approximately one-third of variance in colorectal cancer is attributed to inherited genetic factors (1) and disease risk is increased by two- to fourfold for first-degree relatives of patients. Excess risk of familial cancer can be accounted for by a combination of rare high-penetrance mutations and large numbers of common genetic variants, each of which confers a small risk. These variants combine to confer a range of susceptibilities in the population (2).

The associations between genetic variants and human diseases described so far were highly dependent on the study designs used to identify them. Highly penetrant mutations with large effects have been identified in a number of genes responsible for heritable colorectal cancer susceptibility syndromes (eg, DNA mismatch repair genes, APC, SMAD4, LKB1/STK11, MUTYH, and a linkage region on 15q) and these mutations account for about 3% of incident colorectal cancers in industrialized countries. Identification of further high-penetrance loci has proven intractable to genetic linkage approaches, probably because of locus heterogeneity, and confounding, because of the segregation of multiple alleles in collected families. Association studies conducted in general population samples using common genetic markers have typically identified variants with very small effects. To date, genome-wide association studies (GWASs) have reported 14 common genetic variants that influence the risk of colorectal cancer and account for approximately 6% of the excess familial risk (2–10). Future resequencing studies are expected to identify rarer variants (eg, with 0.05%–5% prevalence) with intermediate, or perhaps even large, effects (2). GWASs of structural variations will likely identify deletions, amplifications, and other copy-number variations that may also influence risk of colorectal cancer.

For this report, we have undertaken a comprehensive review of genetic factors that appear to be associated with colorectal cancer by using previously published guidelines for the assessment of cumulative evidence on genetic association studies (11,12) and following a format similar to previous overview meta-analyses (13–15). We have cataloged all genetic association studies published in this field and conducted meta-analyses of variants with genotypes available in four or more independent case-control studies; any variants that had been typed in two large GWASs were also included in the meta-analyses. The results of the search strategy and meta-analyses are publicly available on a regularly updated Internet database (CRCgene). This represents the first attempt to systematically capture all published genetic association data for colorectal cancer and conduct a meta-analysis.

The core aim is to provide an up-to-date systematic review of the state of the art across the field of colorectal cancer genetics for the research community. We conducted a critical review of all published “candidate gene” study data, incorporating relevant candidate gene data from the GWASs available to us, and then performed meta-analyses of these data. This approach enabled us to summarize available evidence from larger sample sizes, thus gaining greater precision in odds ratio estimates. We have presented these data within a defined statistical and causal inference framework to aid correct interpretation of data (16).

The broader medium-term aim of this work is to identify genetic variants for which there is robust evidence of influence on risk of colorectal cancer. This will help to inform future research efforts and to identify variants that can serve as a basis for providing risk estimates for population groups. It will also provide new insights into the fundamental biological mechanisms involved in colorectal carcinogenesis.

Methods

Literature Search and Data Collection

The first step was to undertake a comprehensive systematic literature review of all current published data on genetics and colorectal cancer. To identify gene association studies aimed at risk of colorectal cancer, we used the Medline database via the Ovid gateway and the search terms comprising medical subject headings (MeSH) and keywords relating to colorectal neoplasms, the MeSH heading “genetic predisposition to disease,” and the keywords “gene$” and “associate$” were applied to terms in the entire article. We cross-checked these 10 125 findings against those listed in the HuGENet phenopedia (17). Of the 10 145 articles thus identified, first we screened the abstracts for eligibility, and then, if necessary, the full texts. We used the following inclusion and exclusion criteria. The article must have evaluated the association between a polymorphic genetic variant [one with a minor allele frequency (MAF) ≥0.01 in the general population based on the data on the reference panel of the 1000 GenomesTable 1; (18)] and sporadic colorectal cancer. Studies that examined associations with only premalignant conditions such as adenomas, polyps, or dysplastic tissue were not included. In addition, studies of hereditary colorectal cancer syndromes, such as familial adenomatous polyposis, hereditary nonpolyposis colorectal cancer, juvenile polyposis syndrome, and Gardner’s syndrome were excluded because our focus was on sporadic colorectal cancer. All studies needed to relate to human participants; any study that was concerned solely with investigating the progression or histological phenotype of colorectal cancer was excluded. Case–control studies and appropriate cohort and GWASs were included. The study had to be published in English in a peer review journal before June 30, 2010. For variants that were identified through the GWAS, the search was repeated and extended until March 31, 2011. Any research that had only been reported in abstracts (eg, presented in scientific conferences yet to be fully published) was excluded. Nine family-based studies were also excluded. We generated a list with all variants to be summarized using meta-analysis and compared it with a list of variants that were included in two GWASs, from Scotland and Canada, for which we had access to individual-level data. The Scottish study comprised patients aged 16–79 years who were diagnosed with colorectal cancer throughout Scotland during the period 1999–2006 (6,19). Potentially eligible population-based control subjects were selected through the Community Health Index, a national register of all individuals registered with a general practitioner in Scotland, and individually matched with patients on age (± 2 years), sex, and area of residence. The Canadian study was based on patients aged 20–74 years who were diagnosed during the period July 1, 1997, to June 30, 2000, who were identified and recruited from the population-based Ontario Cancer Registry (3,20). Random digit dialing was used to select population-based control subjects who were frequency-matched with patients on sex and 5-year age group. In both of these studies, DNA from blood was genotyped using multiple platforms; here, we included data obtained in both studies from the Illumina HumanHap300 and HumanHap240S arrays on the Infinium platform. If a variant was found to be included in either of these GWASs, then genotype counts were included in the meta-analysis of this variant. For the other studies, we contacted the authors of studies with missing data, with an approximate successful response rate of 0.2.

Data Entry, Management, and Abstraction

Once the search was completed, the references of the articles in the search were entered into a web-based database, “RefWorks” (http://www.refworks.com/), thus freezing the search at that point in time. Their abstracts and/or full texts were screened to assess their eligibility for inclusion in the field synopsis. Review articles and meta-analyses on genetic associations of colorectal cancer were also kept alongside so that the references they used could be screened for eligibility if they had been missed in the Medline search.

Data from all studies that met the final inclusion and exclusion criteria were abstracted into two standardized tables. The first of these listed study characteristics and the second table listed allele and genotype frequencies. We abstracted key variables with regard to the study identifiers and context, study design and limitations, intervention specifics, and outcome effects.

Statistical Analysis

Statistical analysis was conducted using Intercooled STATA, version 11.0 (21). Meta-analysis was performed for all variants with case-control data available from four or more independent samples. We obtained summary crude odds ratios (ORs) and 95% confidence intervals (95% CI) for two additive models (variant [var]/wild-type [wt] vs wt/wt; and var/var vs wt/wt), one recessive model (var/var vs var/wt and wt/wt), and one dominant model (var/var and var/wt vs wt/wt). We applied either the fixed-effect model (Mantel–Haenszel method) or, in case of heterogeneity, the random-effect model (DerSimonian–Laird method). Between-study heterogeneity was quantified by calculating the Q statistic, with a P value less than .05 being the threshold. We also calculated the I² heterogeneity metric and its 95% confidence interval (CI). Although sometimes we summarized studies that were very heterogeneous, we recognized that because of the variation in study methods and outcome definitions, the meta-estimates should be interpreted cautiously. To assess any small-study effects, we performed funnel plot analyses and tested for statistical significance using the Harbord modification of the Egger test, implemented in STATA (http://ideas.repec.org/a/tsj/stataj/v9y2009i2p197-210.html). A negative result from small study–effects testing does not entirely exclude publication bias. In addition, the test for small-study effects may be underpowered if there are approximately 10 or fewer studies and it may be inappropriate in the presence of large heterogeneity (22). We also estimated the power that each meta-analysis had to detect a statistically significant effect by using the Power and Sample Size Program (23) and specifying α = .05 as the level of statistical significance and the effect sizes and allele frequencies estimated from the meta-analyses (an integral component of the BFDP analysis).

The relative risk of colorectal cancer in a sibling that was attributable to a given single-nucleotide polymorphism (SNP) was calculated using the following formula (24,25):

(1)

where p is the population frequency of the referent allele, q = 1– p, and r₁ and r₂ are the relative risks (estimated as ORs from the meta- analyses) for heterozygotes and variant homozygotes, relative to wild-type homozygotes. Assuming a multiplicative interaction, we calculated the proportion of the familial risk attributable to an SNP as log(λ*)/log(λ₀), where λ₀ is the overall familial relative risk estimated from epidemiological studies, assumed to be 2.2. Although family studies are normally used to estimate the relative risk for siblings, this formula can be used in population-based studies (24,25).

Finally, we repeated the analysis by excluding any studies that were conducted in a nonwhite population (with white populations defined as Europeans, North Americans, and Australians) for the SNPs that were found to be statistically significantly associated with risk of colorectal cancer in any of the genetic models at a threshold level of α = .05.

Credibility of Genetic Association

To assess the credibility of genetic associations, we considered the BFDP (26) and the Venice criteria (11,12). The BFDP assesses the noteworthiness of an observed association. For variants that were found to be statistically significantly associated with risk of colorectal cancer in any of the genetic models (at P < .05), the BFDP was estimated using the Excel Calculation Spreadsheet (http://faculty.washington.edu/jonno/cv.html). The BFDP threshold for noteworthiness was set up to be equal to 0.20, based on the assumption that a false discovery would be four times more costly than a false nondiscovery. We chose to calculate BFDP values for two levels of prior probabilities: at a medium or low prior level (0.05 to 10^–3) that would be close to what would be expected for a candidate gene; and at a very low prior level (10^–4 to 10^–6) that would be close to what would be expected for a random SNP.

With regard to the Venice criteria, we operationalized the criterion of volume of evidence on the basis of statistical power to detect an association of the desired magnitude: A, 80% or more; B, 50%–79%; or C, less than 50%. For replication, we used the I² criterion proposed by Ioannidis et al. (11). For protection against bias, we considered that complete reporting was problematic. The phenotype definition was addressed by our inclusion criterion ––namely, that case subjects would have newly incident colorectal cancer. In general, genotyping error rates are low (27), and the criterion of replication across studies in part addressed potential concern about variation in genotyping quality between studies; some genotyping issues are discussed in relation to specific findings below. Whereas population stratification may impact gene discovery (28,29), the effect on the magnitude of association in general appears to be small (30,31).

We classified the genetic associations in three categories according to the findings after the BFDP analysis and the application of the Venice criteria. Associations were classified as positive if they fulfilled the following criteria: 1) they were statistically significant (at P < .05) in at least two of the genetic models; 2) they had a BFDP less than 0.20 (at least when P < .05); 3) they had a statistical power greater than 80%; and 4) they had an I² less than 50%. Associations were classified as less-credible positives if they were statistically significant (P < .05) in at least one of the genetic models, their BFDP was greater than 0.20, and their statistical power was between 50% and 79% (I² ranged from 0% to 89% , but this criterion was not taken into account for this category). All other associations were classified as negative.

Model-free Approach

For those SNPs that were identified as true positives after applying the BFDP and Venice criteria, we applied the model-free meta-analysis approach described by Thompson et al. (32). This model uses a maximum likelihood estimator, assumes a fixed-effect meta-analysis and is similar to a bivariate meta-analysis of the effects for the heterozygotes and variant homozygotes. It gives an estimate of the lambda (λ), which represents the underlying genetic model (and it represents the heterozygote effect as a proportion of the homozygote effect), with its confidence interval limits. If the confidence interval of λ is too wide, there is not enough power to estimate λ. The value of λ is not restricted. Values equal to 0, 0.5, and 1.0 correspond to the recessive, additive, and dominant genetic models, respectively. Values greater than 1 or smaller than 0 suggest heterosis, which is when the risk of the heterozygotes is higher or lower, respectively, than the risk of either of the homozygote genotypes. However, heterosis is relatively uncommon and λ values greater than 1 or smaller than 0 could represent chance fluctuation, reflecting low statistical power to estimate the magnitude of effect for heterozygotes.

Results

Literature Search and Data Collection

After screening approximately 10 145 titles and abstracts (Supplementary Table 1, available online) dated until June 30, 2010 (or March 31, 2011 for GWAS-identified hits), we identified and extracted data from 635 publications reporting on 445 polymorphisms in 110 different genes. More than 9231 (88%) of these studies were published after 1995, and about half of those were published during the past 3 years.

Statistical Analysis

Meta-analyses were conducted for 92 polymorphisms in 64 different loci (Table 1) with genotypes available in four or more candidate or GWA studies. On average, these meta-analyses were based on data from 5281 case patients (median interquartile range = 13 472 – 3384) and 6594 control subjects (median interquartile range = 16 102 – 4534) originating from eight (median interquartile range = 13 – 6) case–control samples. Unpublished data from a Scottish and/or a Canadian GWAS were included in the analyses of 37 SNPs. Summary crude odds ratios (ORs) and 95% confidence intervals for two additive models (var/wt vs wt/wt and var/var vs wt/wt) for variants that were identified from candidate studies are presented in Table 2 and for variants that were identified from GWAS, in Table 3. Summary crude odds ratios and 95% confidence intervals for a recessive (var/var vs var/wt and wt/wt) and a dominant model (var/var and var/wt vs wt/wt) are presented in Supplementary Table 2 (available online), for variants that were identified from candidate studies, and Supplementary Table 3 (available online), for variants that were identified from GWAS. For 37 associations in which the 95% confidence intervals excluded unity, we repeated the analysis excluding studies in nonwhite populations (Supplementary Tables 4 and 5, available online). We checked the linkage disequilibrium (LD) for all the polymorphisms that were included in our study and we found that the following variants were in LD (r² threshold ≥ 0.60): 1) rs4939827 (in SMAD7) and rs12953717 (in SMAD7, r² = 0.60); 2) rs3813867 (in CYP2E1) and rs2031920 (in CYP2E1, r² = 0.87); 3) rs10505477 (in 8q24) and rs6983267 (in 8q24, r² = 0.88); 4) rs731236 (in VDR) and rs1544410 (in VDR, r² = 1.00).