Genetic Susceptibility to Cancer: Title and subTitle BreakThe Role of Polymorphisms in Candidate Genes FREE

During the last few decades, extensive effort has been invested in identifying sources of genetic susceptibility to cancer. Both the International Human Genome Sequencing Project and the International HapMap Project have generated a very large amount of data on the location, quantity, type, and frequency of genetic variants in the human genome.^{1 - 4} Facilitated by continuing technological advances that allow faster and cheaper genotyping results, a large and increasing number of observational studies investigating the association between variants in candidate genes and cancer risk have emerged.⁵

This increasing number of studies prompted us to assess the overall contribution of these studies to our current understanding of the genetic susceptibility to cancer. Quiz Ref IDOne of the main criticisms of genetic epidemiology has been a lack of replication. There are several examples of studies exploring a previously published statistically significant finding for a genetic variant and failing to reproduce those findings, suggesting a large number of false-positive reports.^{6 - 7} The size of these genetic association studies is also an important methodological concern that has prompted the utilization of meta-analyses and pooled analyses to combine both statistically significant and nonsignificant results from individual studies and weighting these results by their precision (a function of sample size).^{8 - 10}

To evaluate the overall progress of candidate gene association studies in identifying genetic variants associated with cancer risk, we systematically examined the results of all published meta-analyses and pooled analyses on genetic polymorphisms and risk of cancer and report observed point estimates, 95% confidence intervals (CIs) and P values. Just as 3 parameters are needed to fully evaluate medical diagnostic tests (specificity, sensitivity, and predictive value of a positive test), 3 analogous parameters are needed to evaluate fully statistical tests of an association (eg, between a genetic variant and cancer).¹¹ The P value, the probability of obtaining a more extreme estimate than the one observed when the null hypothesis of no association (odds ratio [OR], 1.0) is true, is analogous to 1 minus specificity (the likelihood of a test classifying a person as having the condition when he/she truly does not have the condition). Study power, the likelihood of detecting an association when one exists, is analogous to sensitivity (the likelihood of a test classifying someone as having the condition when he/she truly has it.) However, it is well established in medical diagnostics that specificity and sensitivity can be high, but the predictive value of a positive test can still be low. This is because, if the condition is rare, positive diagnostic test results will mostly be false positives. This is less appreciated but also important in evaluating statistical tests of hypothesized associations: when the prior probability is small that an exposure-disease hypothesis is true, then a statistically significant finding has a high chance of being a false positive. Quiz Ref IDThe false-positive report probability (FPRP) is defined as “the probability of no association given a statistically significant finding”¹² and is analogous to 1 minus the predictive value of a positive test. Thus, it is the FPRP rather than the P value that answers the question of how probable the hypothesis, as tested, actually is.

In this article, we evaluate the results of candidate gene-cancer association studies by presenting the P value, power, and FPRP for all statistically significant associations as reported in meta-analyses or pooled analyses. The FPRP is calculated from the statistical power of the test, the observed P value, and a given prior probability for the association.¹² Because the prior probabilities are not easily determined, we calculated the FPRP for 2 levels of prior probabilities that are appropriate for a range of hypotheses; from low probabilities, appropriate for polymorphisms with known functional consequences in important candidate genes, to very low probabilities, appropriate for randomly selected variants as used in genome-wide association studies.

This review presents information on knowledge generated thus far by candidate gene association studies conducted to identify cancer susceptibility genes and can also be used to direct future studies toward areas that remain unclear. Furthermore, results from this analysis provide information on the allelic frequency and expected effect size (strictly speaking, strength of association), which can be helpful for planning (genome-wide) association studies.

We identified published meta-analyses and pooled analyses that had evaluated the association between genetic polymorphisms and cancer risk in observational studies (ie, case-control and nested case-control studies) and indexed in PubMed through March 15, 2008. Meta-analyses and pooled analyses are defined as tools that integrate results from individual studies that, alone, may not have sufficient power to detect a statistically significant association.^{8 - 10} In brief, the data (ie, crude and adjusted ORs) used for a meta-analysis are extracted from published results, whereas original data sets acquired from a number of independent studies are used for a pooled analysis. We performed a literature search of the PubMed database using the following search terms for our literature searches: the keyword combinations of cancer plus meta plus gene, cancer plus pooled plus gene, cancer plus consortium plus gene, and the keyword combinations of gene plus cancer and genetic plus cancer restricted to publication type “meta-analysis.”

We considered 794 articles identified through our search methods, screened in detail 224 articles, for a final 161 articles included (Figure 1). Studies included in our review had to meet all of the following criteria: (1) included at least 500 cases combined from all summarized studies, (2) evaluated cancer risk as the outcome (analyses of survival, neoplastic markers, or precursors, such as polyps, were excluded), (3) excluded HLA antigen genetic markers, and (4) published in English. Quiz Ref IDFurthermore, because this review focuses on common variants, meta-analyses and pooled analysis of low-frequency, high-penetrance genes, such as APC and BRCA1 or BRCA2, were excluded. In addition, although statistically significant associations were reported for HRAS1 polymorphisms and risks of breast and lung cancer, these associations have been questioned because of flawed genotyping methods. Thus, these are not reported with other statistically significant associations. To avoid duplication of results from more than 1 meta-analysis or pooled analysis addressing the same association, we selected the most recent publication, which typically had the largest number of cases (sometimes smaller, due to stricter inclusion criteria). Data extracted from each meta-analysis or pooled analysis included cancer site, gene name, genetic variant, point estimate (ie, relative risk [RR] or OR) and 95% CI, allelic frequency (if provided), number of studies, number of cases, test of study heterogeneity (eg, Q test), and test of publication bias (including the Begg test, Egger test and funnel plots). Random-effect estimates from meta-analyses were presented, unless only fixed-effect estimates were available.

We calculated summary estimates to describe published reports identified through our search. Differences in the number of studies and cases were evaluated by t test. Associations were considered statistically significant if the reported P value was <.05 or if the 95% CI excluded 1.0. P values were determined by first calculating a z score based on the reported OR and 95% CI, z score = ln(OR)/[(ln(upper CI) – ln(lower CI))/(2 × 1.96)], and then comparing it to a normal distribution.

For each statistically significant association reported, we estimated the FPRP using methods described by Wacholder et al.¹² The FPRP value is determined by the P value, the given prior probability for the association, and the statistical power of the test. Assigning a prior probability should be determined before obtaining results from a study and should be independent of any data used in the analysis. Prior probabilities are subjective and are influenced by both previous epidemiologic findings and experimental evidence about known functions of a genetic variant. Therefore, we chose to calculate FPRP values for 2 levels of prior probabilities: at a low prior that would be similar to what would be expected for a candidate gene (0.001) and at a very low prior that would be similar to what would be expected for a random single-nucleotide polymorphism (SNP) (0.000001), thus, allowing readers to evaluate the association using their own judgment about the supporting evidence for a given loci. Wacholder et al¹² suggest estimating statistical power based on the ability to detect an OR of 1.5 (or its reciprocal, 0.67 = 1/1.5 for ORs <1.0), with an α level equal to the observed P value.¹² But given the recent attention to much smaller ORs this estimate may be too conservative; thus, we have chosen to present results for both an OR of 1.5 and 1.2 (or its reciprocal 0.83 = 1/1.2). To evaluate whether an association is noteworthy, we used a FPRP cutoff value of 0.2, as suggested by the authors¹² for summary analyses. Hence, FPRP values less than 0.2 indicate an association that remained robust for a given prior probability and will be referred to as noteworthy in the present article. Statistical power and FPRP were computed by the Excel spreadsheet provided by Wacholder et al.¹²

The genes (National Center for Biotechnology Information Entrez identification numbers) for which the variants have a noteworthy association with cancer are as follows: MDM2 (4193), XPD (name changed to ERCC2) (2068), RNASEL (6041), GSTT1 (2952) XRCC1 (7515), TGFB1 (7040), CASP8 (841), NAT2 (10), MTHFR (4524), CHEK2 (11200), and GSTM1 (2944).

We identified 161 published meta-analyses and pooled analyses, encompassing 18 cancer sites and 99 different genes. These 161 meta-analyses and pooled analyses addressed 344 gene-variant cancer associations with an average of 7.3 studies and 3551 cases per investigated association (range, 508-19 729 cases). As expected, most analyses were conducted for common cancers, such as breast (n = 119), prostate (n = 42), and lung (n = 34) cancer; there are very few evaluations of genetic associations in rare cancers, such as cervical and esophageal (Table 1). Across all cancer sites, variants in genes involved in DNA repair (eg, XRCC1 and XPD; n = 81) and genes encoding metabolizing enzymes (eg, cytochrome P450 (CYP) variants, n = 58; or glutathione S-transferases (GSTs), n = 31) were most often evaluated. Meta-analyses and pooled analyses that found a statistically significant association evaluated a higher number of studies but included a lower number of cases than those that found a nonsignificant association (P = .02 and P = .05, respectively; Table 1). A complete table that lists all data extracted from each of the 344 associations identified in our search is available upon request.

Among the 344 gene-variant cancer associations evaluated, the summary OR for 98 (28%) associations (excluding those involving HRAS1) were statistically significant (P values between .05 and 8.6 ×10⁻¹⁵; Figure 2 and Figure 3 and Table 2, Table 3, Table 4, and Table 5). Thirty of these 98 associations were inverse for the variant, with a mean OR of 0.73 (median, 0.75; range, 0.32-0.92). The other 68 analyses reported ORs higher than 1.0, with a mean of 1.47 (median, 1.34; range, 1.07-3.13). Statistically significant associations were found among 16 cancer sites, predominantly among studies investigating breast cancer, glioma, and lung cancer.

To evaluate the robustness of these findings, we calculated FPRP values at 2 levels of prior probabilities. Among the 98 associations, 85 gene-variant cancer associations had FPRP values higher than 0.2 across the prespecified prior probabilities (0.001 and 0.000001); these results are not considered noteworthy. For example, although the summary OR from the pooled analysis for XRCC1, Arg399Gln indicated a statistically significant positive association with risk of breast cancer (OR, 1.6; 95% CI, 1.1-2.3), FPRP values were higher than 0.2, at any of the 2 prior probabilities; hence, the finding is not considered noteworthy. At a prior probability level of 0.001 and statistical power to detect an OR of 1.5, 13 gene-variant cancer associations remained noteworthy (FPRP ≤0.2) for MDM2 SNP309 and lung cancer (OR, 1.27; P = .0002)⁵⁵ ; XPD Lys751Gln and lung cancer (OR, 1.30; P = .0002)⁵⁸ ; RNASEL Asp541Glu and prostate cancer (OR, 1.27; P = .0001)⁶⁴ ; GSTT1 null and colorectal cancer (OR, 1.37; P = 8.1 × 10⁻⁵)²⁸ ; XRCC1 Arg399Gln and lung cancer (OR, 1.34; P = 5.2 × 10⁻⁵)⁵⁹ ; TGFB1 Leu10Pro and breast cancer (OR, 1.16; P = 6.9 × 10⁻⁵)¹⁴ ; CASP8 Asp302His and breast cancer (OR, 0.89; P = 5.7 × 10⁻⁶)¹⁴ ; NAT2 slow acetylator and bladder cancer (OR, 1.46; P = 2.5 × 10⁻⁷)³⁶ ; MTHFR C677T and gastric cancer (OR, 1.52; P = 4.9 × 10⁻⁸)⁴⁵ ; CHEK2 *1100delC and breast cancer (OR, 2.4; P = 2.5 × 10⁻⁹)¹⁵ ; GSTT1 null and acute leukemia (OR, 1.19; P = 3.5 × 10⁻⁸)⁶⁸ ; GSTM1 null and bladder cancer (OR, 1.5; P = 1.9 × 10⁻¹⁴)³⁴ ; and GSTM1 null and acute leukemia (OR, 1.20; P = 8.6 × 10⁻¹⁵).⁶⁸ At a very low prior probability of 0.000001, 4 of these 13 gene-variant cancer associations remained noteworthy: MTHFR C677T, NAT2 slow acetylator, and GSTM1 null. This number further reduced to 2 (GSTM1 null with bladder cancer and GSTM1 null with leukemia) when we calculated statistical power based on a lower OR of 1.2. Consistent with the FPRP, associations noteworthy at a very low prior probability were highly statistically significant (P values between 10⁻⁷ and 10⁻¹⁵).

Overall, close to one-third of all gene-variant cancer associations from published meta-analyses and pooled analyses were reported to be statistically significant. Thirteen of these associations were noteworthy at a prior probability of 0.001 and statistical power to detect an OR of 1.5, of which 4 remained noteworthy at even a lower prior probability similar to one appropriate for a randomly selected SNP in a genome-wide association study (1/1 000 000 = 0.000001) with P values between 10⁻⁷ and 10⁻¹⁵. These associations are thus less likely to be false positives and have a high likelihood of being true associations with cancer risk. Specifically, we observed that, among the noteworthy associations, genes encoding for phase II metabolizing enzymes made up the majority of noteworthy associations.

Continuing advances in genotyping technologies have led to the feasibility of testing a large number of genetic variants; with this has come the potential for the publication of a large number of false-positive results due to the widely used strategy of declaring significance based on a P value of <.05. A key feature of the Bayesian approach using the FPRP is that it is based, not only on the observed P value but also on both the power and prior probability of the hypothesis, allowing the user to incorporate prior knowledge, including functional information, of the specifically tested variants. Although the FPRP calculation allows an evaluation at different scenarios of prior probability, statistical power, and noteworthiness criterion, the choice for these parameters should be determined a priori using empirical evidence from past studies. Accordingly, it may be reasonable to claim that SNPs of relevant candidate genes with known or predicted function (based on experimental studies or in silico tests) are more likely to be associated with cancer risk and hence justify higher prior probabilities. However, choice of a single prior probability will be subject to debate; hence, herein, we provide readers with the opportunity to use their own judgment about the body of evidence for a given candidate gene or variant. In this article, we chose a more agnostic approach to evaluating associations by applying 2 levels of prior probability (0.001 and 0.000001) and statistical power (OR of 1.5, recommended by Wacholder et al and similar to the average reported OR in our review; as well as OR of 1.2, close to the median reported OR in our review) to all statistically significant associations. As suggested by Thomas and Clayton,⁷² the prior probability for studies evaluating candidate genes will usually exceed 1000:1 (or 0.001). Thus, at a prior probability of 0.001, 13 associations were noteworthy and may plausibly be true associations. The likelihood of being a true association, however, is even greater for the 4 associations that remain noteworthy at a very low prior probability (0.000001).

GSTM1 and GSTT1 belong to a family of phase II enzymes, the glutathione S-transferases, that are involved in the metabolism and biotransformation of toxic xenobiotics and endobiotics.⁷³ Deletion of GSTT1 was associated with an increased risk of colorectal cancer²⁸ and acute leukemia⁶⁸ and the GSTM1 deletion was statistically significantly associated with risk of bladder cancer³⁴ and acute leukemia⁶⁸ ; and the latter 2 were found to be among the most noteworthy findings across all meta-analyses and pooled analyses. Individual studies conducted subsequent to the meta-analyses continue to support findings for GSTT1^{74 - 79} and GSTM1,^{80 - 85} except for one study that reported a statistically significant inverse association between GSTT1 null and colorectal cancer⁸⁶ and a few small studies on GSTT1 and leukemia providing inconsistent results.^{83 ,85 ,87 - 88} The prevalence of GSTT1 null ranges from 20% in whites to 60% among Asians,⁸⁹ and approximately 50% of humans (ranging from 22% in Africa to 62% in Europe) are GSTM1 null.⁹⁰ Quiz Ref IDGSTT1 and GSTM1 are involved in the elimination of carcinogens in the body, such as products of oxidative stress and polycyclic aromatic hydrocarbons from tobacco smoke.⁹¹ Deletion of the GSTT1 and GSTM1 gene results in the variant called GSTT1/GSTM1 null and a complete loss of enzymatic activity.⁹² An individual with the null variants is thus expected to have an impaired ability to detoxify carcinogens and have an increased risk of cancer, potentially affecting multiple cancer sites. This and the fact that GSTT1 and GSTM1 result in noteworthy associations with risk of various cancers lends support to the theory that these 2 variants, in particular GSTM1 are functional and truly impact cancer risk.

Another finding that was among the most noteworthy was the association between NAT2 slow acetylator phenotype and bladder cancer.³⁶ This meta-analysis was published recently; thus, no additional studies were identified subsequent to the meta-analysis. NAT2 is 1 of 2 N-acetyl transferase isoforms expressed in humans, which are involved in the detoxification of heterocyclic or aromatic amines and their metabolites.⁹³ NAT2 is highly polymorphic and several nonsynonymous polymorphisms result in poor expression, an unstable protein, or decreased catalytic activity, all of which result in the slow acetylator phenotype.⁹⁴ The prevalence of NAT2 slow acetylators in European whites is about 56% and approximately 11% among Asians.³⁴ The change in the rate of acetylation is expected to alter the effect of carcinogens on cancer risk, but the effect of this change may differ by cancer site. The NAT2 slow-acetylator phenotype is associated with an increased risk of bladder cancer (due to decreased detoxification of carcinogens from tobacco smoke), but has been associated with decreased risk of colorectal cancer (due to reduced activation of carcinogens).^{31 ,93 - 94} Taken together, the strong evidence supporting a functional effect of the NAT2 slow acetylator and the highly statistically significant association with bladder cancer supports the hypothesis that this variant is likely to modify cancer risk.

The recently published association between MTHFR C677T and gastric cancer was also among the most noteworthy associations.⁴⁵ MTHFR, 5,10-methyletetetrahydrofolate reductase, plays a key role in the 1-carbon metabolism pathway. Specifically, MTHFR converts 5,10-methylenetetrahydrofolate to 5-methyltetrahydrofolate which then allows for the metabolism of homocysteine and the provision of methyl groups. Enzyme activity among individuals homozygous for MTHFR C677T is much reduced, approximately 30% of expected enzyme activity, compared with those who are homozygous for the common variant.^{95 - 96} Consequently, the reduced ability of MTHFR has been associated with alteration in methylation patterns and potentially aberrant DNA synthesis, repair, and chromosomal instability.⁹⁷ Due to its role in a key pathway, the MTHFR C677T variant may have a true impact on cancer risk.

Among associations noteworthy at prior probabilities of 0.001 were 3 genes associated with DNA repair (CHEK2, XPD, and XRCC1). Pathways involving these genes are responsible for repairing DNA damage and errors that may occur during DNA replication. There have been no studies published subsequent to the meta-analysis on CHEK2 *1100delC and breast cancer.¹⁵ Studies conducted subsequent to the meta-analysis on XPD Lys751Gln and lung cancer^{98 - 99} have drawn the same conclusions as our review. The statistically significant finding for XRCC1 was present among Asians only, and 1 of the 3 subsequent studies conducted among Asians^{100 - 102} found a statistically significant association between XRCC1 Arg399Gln and lung cancer. Quiz Ref IDOverall, it is biologically plausible that genes associated with DNA repair have an impact on the risk of cancer and our review lends support toward the likelihood of these associations.

RNASEL Asp541Glu, MDM2 SNP309, TGFB1 Leu10Pro, and CASP8 Asp302His are additional variants identified through our review as being noteworthy; they belong to key pathways plausibly influencing cancer susceptibility. RNASEL plays an important role in the inflammatory response pathway and was first identified as a candidate gene for prostate cancer risk due to its location within the hereditary prostate cancer 1 (HPC1) region.^{103 - 104} Because the meta-analysis has been published recently, only 3 subsequently published studies were identified but with conflicting results for prostate cancer.^{105 - 107} MDM2 encodes for the human homologue of mouse double minute 2, a nuclear phospholipoprotein that binds and inhibits p53, a tumor suppressor.¹⁰⁸ A further study published after the meta-analysis lends support when analysis was restricted to never smokers.¹⁰⁹ TGFB1, which encodes transforming growth factor beta 1, has been implicated as both a tumor suppressor and a tumor promoter.^{110 - 111} An additional study published subsequently did not find an association.¹¹² CASP8 encodes for Caspase 8 which plays a central role in the initiation and activation of a cascade of caspases leading to apoptosis.¹¹³ The decreased risk with CASP8 Asp302His for breast cancer observed in the pooled analysis is further supported by findings from a recent association study.¹¹⁴

Very recently, results from the first genome-wide association studies of cancer have become available, in which hundreds of thousands of variants were genotyped across the entire genome. These studies detected several highly statistically significant variants in the human chromosome 8q24 region that were associated with prostate, colorectal, and breast cancer susceptibility; however, there are no known characterized genes within this region.^{115 - 122} Variants located within SMAD7,¹²¹ a gene involved with cell signaling, and DAB2IP,¹²³ a putative tumor suppressor gene, have also been associated with colorectal and prostate cancer, respectively. Three follow-up genome-wide scans in prostate cancer have confirmed the previously identified loci and identified several additional loci that may be associated with prostate cancer risk.^{124 - 126} The loci that were identified in at least 2 of the studies were as follows: 8q24, HNF1B (17q12), MSMB (10q11), NUDT10/11 (Xp11.22), and 17q24. Six highly statistically significant variants associated with breast cancer susceptibility have also been identified through genome-wide studies, of which 3 are located within genes associated with control of cell growth or cell signaling (TNRC9, MAP3K1, and LSP1).^{122 ,127 - 128} Two variants were located in the 8q24 and 2q35 regions, and the sixth within FGFR2, a tumor suppressor gene overexpressed in breast cancer. The substantial evidence supporting these variants, including sizeable power and replication in large samples, indicates that these associations are likely to be true and yet none of the statistically significant variants had been previously identified because most did not reside in “interesting” candidate regions. Genome-wide association studies of cancer have also demonstrated that the effect size of statistically significant genetic variants is overall quite modest (point estimates between 1.1 and 1.5 for an additive mode of inheritance), which is consistent with the weak associations found in most meta-analyses and pooled analyses.

We attempted to review all published meta-analyses and pooled analyses covering the topic of genetic variants and cancer risk through several iterations of search criterion; however, it is possible that we have missed some studies. Many of the noteworthy variants identified were deletions (which may not be well captured by genome-wide association studies) and nonsynonymous SNPs, but this may be due to the fact that these types of mutations tend to be the most commonly studied. Our focus was strictly on results from candidate-gene association studies and did not take into account results from linkage studies to identify high-penetrance genes. Another potential limitation of this review is that associations were confined to those summarized in a meta-analysis or pooled analysis. We are aware of individual studies with potentially much larger sample sizes and hence more power to find a statistically significant association than some meta-analyses and pooled analyses; some of these studies have been conducted subsequent to the meta-analyses or pooled analyses and some prior. To address this issue in part, we reviewed studies conducted subsequent to the latest meta-analyses or pooled analysis for associations considered noteworthy at a low prior probability to determine whether evidence continued to support the previously observed associations. Another limitation of our review is that our results are susceptible to reduced quality and breadth of the meta-analyses or pooled analysis as a result of publication bias. However, most analyses included herein tested for publication bias and heterogeneity, as noted in the accompanying tables. Because the power to assess gene-gene and gene-environment interactions is even lower than that to assess main effects and most meta-analyses and pooled analyses focused on main effects, we only reported on main effects of genetic variants. Therefore, we may have missed important subgroup effects, for it is possible that certain genetic variants may only be relevant when “the system is under stress,” eg, smoking, concurrent illness, or malnutrition. Most analyses evaluated single-candidate polymorphisms; however, because genotyping has become increasingly affordable in recent years, this now allows investigators to test for genetic variants across entire candidate genes and pathways and most recently across the entire genome. Although results from single SNPs are easy to compare, this approach is certainly less comprehensive and does not rule out that other SNPs in the same gene may be related to cancer risk. As the number of articles on genetic variants published in the past decade has increased considerably and continues to grow, we accept that this review will not long remain current but does provide a snapshot of progress in the field.

We observed 98 statistically significant gene-variant cancer associations, of which 13 were considered noteworthy at a prior probability of 0.001. At a very low prior probability (0.000001), 4 remained noteworthy of which all were highly statistically significant (P values between 10⁻⁷ and 10⁻¹⁵). A majority of the most noteworthy associations identified are not SNPs but deletions, 4 involve GST variants. Results from meta-analyses and pooled analyses were helpful in synthesizing published results and may guide future genetic studies toward areas that require further clarification and away from those that do not.