Haplotype Analysis of Common Vitamin D Receptor Variants and Colon and Rectal Cancers

The vitamin D receptor (VDR) mediates signals from circulating vitamin D to the cell nucleus, affecting transcription of a number of genes, including those encoding the insulin-like growth factor–binding proteins (1-3), which modulate insulin-like growth factor effects on cell growth. Thus, the antiproliferative effect of vitamin D on cell cultures, and its hypothesized protective effect against cancer, is thought to occur through binding to the VDR (4).

Several common polymorphisms of the VDR gene have been studied as possible genetic contributors to cancer susceptibility. A group of polymorphisms are found near the 3′ end of the of the VDR, including a single nucleotide polymorphism (SNP) in an intron between exons 8 and 9 detected by the Bsm1 restriction enzyme and a polyadenosine [poly(A)] length polymorphism in the 3′ untranslated region (5). These polymorphisms are in extensive linkage disequilibrium (6). Although these specific DNA sequence changes do not affect VDR function (7), epidemiologic studies, reviewed by Thakkinstian (8), show that these VDR variants influence bone density, providing evidence that the 3′ region polymorphisms are markers for functionally different alleles. A SNP detected by the Fok1 restriction enzyme is recognized as functional, altering a translation initiation site (9) and affecting bone mineral density (10). The Fok1 polymorphism has been reported to exhibit little linkage disequilibrium with other VDR polymorphisms (11).

The potential associations between risk of colorectal cancers or adenomas and the VDR 3′ region polymorphisms and/or the Fok1 polymorphism have been investigated in several studies (12-22). The implications of the evidence from these studies are not completely clear, as some studies detected associations with the 3′ region polymorphisms and others report associations with the Fok1 polymorphism. Some of the published studies have considered only one of the known common variants, whereas others evaluate genotype for two or more. With the exception of one small study (23), studies of these VDR variants in colorectal cancer risk have not applied haplotype-based analysis. The functional effect of genetic variation on gene expression, protein structure, or enzyme activity is expected to be better described by a summary of all genetic variants present on an allele (i.e., a haplotype) than by the presence or absence of individual polymorphisms (24, 25).

In the present study, we constructed VDR haplotypes from genotype data determined for three VDR polymorphisms: the intronic SNP detected by the Bsm1 restriction enzyme, the 3′ untranslated region poly(A) repeat length polymorphism, and the Fok1 SNP. Using data from population-based, case-control studies of colon and rectal cancer, we estimated haplotype frequencies for four race or ethnic groups, and evaluated associations between haplotype and cancer risk. Further, we compared VDR haplotype variables to VDR genotype variables for their ability to explain case-control differences.

We report on participants in collaborative population-based case-control studies of cancers of the colon and rectum. The colon cancer study was conducted in eight counties in Utah, the Kaiser Permanente Medical Care Program in northern California, and seven counties in Minnesota. Individuals with diagnoses of colon cancer (ICD02 18.0 and 18.2-18.9) between October 1991 and September 1994 were identified as cases for the study through the Utah Cancer Registry, the Northern California Cancer Registry, the Sacramento Cancer Registry, and the Minnesota Cancer Surveillance System. The rectal cancer study was conducted in the state of Utah and Kaiser Permanente Medical Care Program of northern California only; individuals with diagnoses of rectal cancer (ICD02 19, 20) between May 1997 and May 2001 were identified through the registries in those states. In northern California, only cases who were members of the Kaiser Permanente Medical Care Program were eligible for the studies. In Utah, control subjects ages ≤64 years were randomly selected from computerized drivers' license lists and those ages ≥65 years from Center for Medicare Studies (formerly Health Care Finance Administration) lists. Membership lists of the Kaiser Permanente Medical Care Program were used as a sampling frame for controls in northern California. In Minnesota, controls were selected from drivers' license lists. Control groups were frequency-matched to the age and sex distributions of incident cancer cases. Additional eligibility criteria for both cases and controls were as follows: being 30 to 79 years of age, being able to complete an interview in English, and having no previous diagnosis of colon or rectal cancer. Details of the procedures for recruitment of participants have been described (26-28). Recruitment and interviewing were conducted in accordance with human subject research protocols approved at each responsible institution. Participants took part in a structured interview. Each participant was asked to identify his or her race and ethnic background. Interview response rates among colon cancer study cases and controls were 67.7% and 52.4%, respectively, and in the rectal cancer study were 65.2% and 65.3% (27, 29).

Participants were asked to provide a blood sample for DNA extraction. An A>G SNP in the untranslated region (rs154410), intron 8 to 9, was detected by PCR and RFLP using the Bsm1 enzyme, according to a published method (30) with modifications as described (22). Presence of the Bsm1 restriction site is represented by “b” and its absence as “B.” The length of the poly(A) repeat was determined by PCR amplification as described (22, 30); the size of PCR products was determined after separation by gel electrophoresis and categorized as short “S,” 14 to 17 repeats, or long, “L,” 18 to 22 repeats. The T>C SNP (rs10735810) that affects the first of two transcription start codons (ATG) was detected by PCR and Fok1 RFLP (10); the presence of the Fok1 restriction site (first ATG intact) is represented by “f” and its absence (first ATG start codon lost, protein product will be shorter by three amino acids) by “F.”

This analysis is restricted to subjects who provided a blood sample and for whom all three VDR polymorphisms were successfully genotyped: 1,452 colon cancer cases (73% of 1,993 interviewed cases), 1,811 colon cancer study controls (75% of 2,410 interviewed), 679 rectal cancer cases (75% of 952 interviewed), and 905 rectal cancer study controls (77% of 1,205 interviewed). Subjects were categorized as belonging to one of several mutually exclusive racial or ethnic groups: White (Caucasian or White; not Hispanic), Hispanic (Hispanic or Latino; not African-American or Black), African-American (African-American or Black), or Asian (Asian or Pacific Islander), based on self-report of race and ethnicity. A few individuals reported race as Native American, but they were excluded from this analysis because there were too few subjects to conduct haplotype analysis for this group.

We used genotype data from the control subjects, considering each racial or ethnic group separately, to infer chromosomal phase of linked loci and to estimate haplotype frequencies. We used the expectation-maximization algorithm (31), implemented in SAS/Genetics software (SAS/Genetics, Cary, NC, 2002), to develop maximum likelihood estimates of population haplotype frequencies. The algorithm converges on the haplotype frequencies that have the highest probability of generating the observed genotypes. An omnibus test comparing haplotype distribution of two groups (32) was used to test for differences by race. Lewontin's D′ statistic for linkage disequilibrium was calculated for each pair of polymorphisms. The values of Lewontin's D′ can range from −1 to 1; we present the absolute value of D′, with zero indicating complete independence between two markers and one representing complete linkage disequilibrium. The expectation-maximization algorithm output also estimates each subject's probabilities of carrying particular pairs of haplotypes, based on his or her genotypes. These probabilities can be used in regression models to evaluate disease association (33). Haplotype probabilities for cases were assigned according to the estimated haplotype probabilities of a control from the same race or ethnic group with the same genotypes. Eight haplotypes were possible, based on the three polymorphisms considered. The haplotype probabilities were used to create haplotype dose variables (34). The values of the haplotype dose for an individual can range from zero, indicating that the haplotype is not possible based on the subject's genotypes, to two, indicating homozygosity for the haplotype; the sum of the eight haplotype probabilities equals to two.

Case-control comparisons were made for each of three race or ethnic groups, White, Hispanic, and African-American; small numbers of Asian subjects precluded meaningful case-control analysis for this racial group. An omnibus test for differences in the distribution of haplotypes between cases and controls was run, implementing a model-free statistical comparison (32). For our primary analysis, the haplotype dose variables were entered, as continuous variables, into logistic regression models to estimate odds ratios (OR) for the association between the haplotype and colon or rectal cancer. We first ran logistic regression models that included all haplotype variables. One model included all White, Hispanic, and African-American subjects and was adjusted for age, sex, study center, and race. Additional models were run separately for Whites, Hispanics, and African-Americans, adjusting for age, sex, and study center. We then used stepwise selection procedures to identify haplotype variables most likely to predict case-control status and to estimate ORs for haplotypes retained in the more parsimonious model. We also ran logistic models using haplotype variables coded 0, 1, 2, representing each subject's most likely pair of haplotypes, instead of haplotype dose. Finally, we compared VDR haplotype variables to VDR genotype variables, which have been reported on previously in this population (15, 22), on ability to explain colon cancer case-control differences. Using the logistic model, we evaluated improvement in −2 log-likelihood values; in other words, the improvement in model fit when a haplotype or genotype variable (or set of variables) was added to a restricted logistic model.

The allele frequencies for the VDR Bsm1, poly(A), and Fok1 polymorphisms among controls, by race and ethnicity, are shown in Table 1. The genotype distributions were in Hardy-Weinberg equilibrium within each group, with the exception that a smaller than expected number of heterozygotes for the Bsm1 and poly(A) variants was observed among the small sample (n = 26) of Asian controls. The absolute values of the D′ statistics indicated that the linkage disequilibrium between the Bsm1 b and poly(A) L genotypes was almost complete among White and Hispanic subjects. Linkage disequilibrium was also seen between Bsm1 and poly(A) among African-American and Asian subjects, but was not as strong. There was little evidence of linkage disequilibrium between the Bsm1 and Fok1 polymorphisms, and similarly little between the poly(A) polymorphism and Fok1. When data were analyzed separately by study center, genotype frequencies and linkage disequilibrium coefficients within each race or ethnic group were essentially identical for each center.

The haplotype frequencies predicted by the expectation-maximization algorithm are shown in Table 2. The haplotype bLF was the most frequent haplotype among all race and ethnic groups studied. The four haplotypes containing Bsm1 b with poly(A) S, or Bsm1 B with poly(A) L, were uncommon in all groups. The four uncommon haplotypes were particularly rare, <1% frequency for each haplotype, in the White and Hispanic populations, in keeping with the strong linkage disequilibrium between Bsm1 and poly(A) polymorphisms in these groups. The haplotype distributions differed significantly by race.

In case-control comparisons for colon cancer (Table 3), the omnibus test detected differences between cases and controls in the distribution of VDR haplotypes for all races combined (P = 0.0004) and for Whites (P = 0.004). The logistic regression model results from the full model, including haplotype dose variables for all eight haplotypes, indicated that the haplotypes bLF and BLF were associated with increased risk of colon cancer, and that the haplotype bLf was associated with reduced risk of colon cancer. The confidence intervals (CI) on the ORs for each of these three haplotypes excluded one in the model for all three races and in the model limited to Whites. ORs for bLF and colon cancer among African-Americans and Hispanics, and for BLF and colon cancer in African-Americans, were >1, suggestive of an increased risk of a similar or greater magnitude to the association observed for Whites, but CIs were wide in these smaller groups of subjects.

In the stepwise modeling analysis, with P values of 0.05 specified for entry into the model, the bLF variable, with an OR of 1.15 (95% CI, 1.03-1.28), and the BLF variable, with an OR of 2.40 (95% CI, 1.43-4.02), were selected in the model that included colon cancer cases and controls of all races. The same two haplotypes were selected by the stepwise model in an analysis restricted to Whites. For Hispanics and African-Americans, no variables were entered by the stepwise model. The OR for the bLF haplotype dose variable represents a 15% increased risk of colon cancer for individuals with one allele for the haplotype bLF compared with those with no copies of this allele, and individuals who are homozygous for the bLF haplotype are at 30% increased risk. The haplotype bLF was common, with an estimated frequency >35% in all of the race and ethnic groups that we studied, so ∼12% of the population is expected to be homozygous for this haplotype. The OR for the BLF haplotype dose variable indicates a >2-fold increased risk for individuals carrying one BLF allele. Because of the low frequency of the BLF haplotype, homozygosity for BLF would be extremely rare.

The omnibus test did not detect differences in haplotype distributions between rectal cancer cases and controls, and there were no haplotype variables that were significantly associated with rectal cancer in the full logistic model (Table 3). The stepwise selection procedure did not identify any haplotype variables as associated with rectal cancer at P = 0.05 nor at P = 0.10.

When we ran the logistic models using variables coded 0, 1, 2 for each haplotype, representing each subject's most likely pair of haplotypes, results were very similar to those shown in Table 3, although some OR estimates were very slightly closer to the null. For example, in the full model for colon cancer for all races, the ORs for bLF and BLF were 1.12 (95% CI, 1.01-1.24) and 2.14 (95% CI, 1.36-3.40), respectively, from the most likely pairs analysis, compared with 1.13 (95% CI, 1.01-1.26) and 2.26 (95% CI, 1.36-3.78) from the haplotype dose model.

Logistic models containing VDR genotype or VDR haplotype variables were compared (Table 4). The difference in the −2 log likelihoods associated with two haplotype-dose variables, bLF and BLF, was 16.58 units, whereas the corresponding value for adding dummy variables representing Fok1 Ff and ff genotypes was 7.32; both represent statistically significant improvement in model fit, and thus significant association with case-control status, but the ability of the haplotype variables to explain variability in the model is better than that of genotype variables. The model with the two haplotype variables selected by the stepwise model was associated with more improvement in −2 log-likelihood values than any genotype variable or variables. However, for the model containing one haplotype dose variable for the common haplotype bLF, the model improvement, 4.91 units, was not greater than that associated with the Fok1 genotype variable, 7.29 units.

We evaluated haplotypes based on three common variants of the VDR gene in a data set from large, population-based case-control studies of cancers of the colon and rectum. We observed tight linkage disequilibrium between the Bsm1 and poly(A) polymorphisms, whereas there was almost no linkage disequilibrium between the Fok1 polymorphism and the other loci. The absence of linkage disequilibrium between the Fok1 and other SNPs is consistent with what has previously been reported (11). There was variation in the extent of linkage disequilibrium among race or ethnic groups in this U.S. study population, as has been reported for Bsm1 and poly(A) in another population (6).

The overall distribution of VDR haplotypes differed between colon cancer cases and controls, but not between rectal cancer cases and controls. The number of Hispanic and African-American subjects in this study population was too small to detect significant associations between haplotype and disease within these groups analyzed separately. However, there were qualitative similarities of the ORs between Whites, Hispanics, and African-Americans, indicating that, despite the differences in VDR haplotype distributions among these groups, it was appropriate to combine the groups in a race-adjusted model for case-control comparisons. For the combined analysis, after stepwise selection of haplotype variables, ORs for colon cancer risk associated with haplotype dose were significantly greater than one for bLF, the most common haplotype, and for BLF, a relatively rare haplotype.

We considered haplotype variables as an alternative to variables representing genotypes for modeling colon cancer risk related to VDR variants. We compared logistic regression models that included the same study subjects and covariates, differing only in the variables used to represent VDR haplotype or genotype. The combination of the bLF and BLF haplotypes conferred the most improvement in model fit. However, a variable representing the Fok1 genotype was also associated with a very significant improvement in model fit. Among the three polymorphisms examined, the Bsm1 and poly(A) were in almost perfect linkage disequilibrium, but only very weak linkage disequilibrium existed between these two and the Fok1 SNP. With this linkage disequilibrium pattern, genotype variables served almost as well as haplotypes to model disease risk; contrasts between explanatory power of haplotype versus genotype variables might be stronger for genes or polymorphisms with more complex linkage disequilibrium patterns.

The number of possible haplotypes in a data set can be large, so that when case-control comparisons of haplotypes are conducted, there is a possibility that a significant association for a single haplotype may represent a chance finding among multiple comparisons. It is important to consider results of a single test for differences in the distribution of haplotypes between cases and controls when making comparisons using haplotypes. In our analysis of VDR haplotypes, the detection of a highly significant difference in distribution of haplotypes between colon cancer cases and controls based on the omnibus test, a model-free comparison, improves our confidence in the conclusion that differences are not explained by chance. The omnibus test has certain limitations: It does not describe the magnitude of the association and it does not take into account covariates such as study center or race. In the present study, haplotype distributions were consistent across centers and case-control differences were similar by race. Logistic regression models were used to estimate ORs associated with haplotype dose, adjusted for covariates.

It has been suggested that only haplotypes with a population frequency of ≥5% should be evaluated in case-control comparisons. However, associations between rare alleles and disease may be overlooked if this rule of thumb is applied. Infrequent haplotypes cannot be meaningfully analyzed in small data sets, but the present study evaluated a large number of colon and rectal cancer cases and controls so that we were able to retain rare haplotypes for analysis. We applied a stepwise modeling procedure to select haplotypes that best predicted disease risk; the BLF haplotype, which had a frequency of <1% among Whites and Hispanics, proved to be statistically significantly associated with colon cancer. A possible drawback of including all haplotypes in a model is that these highly correlated variables may produce output that is difficult to interpret. In this example, however, the ORs from the full and reduced models were very similar for the two haplotypes selected.

The traditional approach to investigating candidate genes in molecular epidemiology has been to select one or a few individual polymorphisms hypothesized to confer disease susceptibility through an effect on protein function or on the expression of the gene and to examine genotypes for these polymorphisms in relation to disease risk. More recently, investigators have been encouraged to take into account variability throughout the candidate gene by identifying haplotype-tagging SNPs and analyzing these markers in study populations (25, 35, 36). The second approach will be more resource-intensive, requiring that genotypes be determined for a larger number of markers of unknown functional significance, and only a limited number of examples have appeared in the cancer epidemiology literature (e.g., ref. 34). Analysis based on tag SNPs will become more feasible as our understanding of the human genome expands, the efficiency of genotyping methods improves, and statistical methods for haplotype analysis are further developed. Still, the traditional approach will remain valuable when there is knowledge of biological function associated with a polymorphism. A third, hybrid approach is possible: to focus on polymorphisms believed to be functional, but to assay multiple polymorphisms in the gene and to construct haplotypes based on the combination of these markers. The present study is an example of this hybrid approach. The VDR Fok1 polymorphism is recognized to be functional, with the C allele altering a translation start codon so that the protein would differ in length by three amino acids. The Bsm1 and poly(A) polymorphisms are among a group of variants that are associated with functional differences in bone density, although the polymorphisms are thought to be markers rather than functional variants. In our analysis, the comparison of likelihoods from logistic models showed that, using data from the same polymorphisms, the VDR haplotype variables were only somewhat better predictors of disease risk than VDR genotype variables.

We thank Michael Hoffman and Sandie Edwards for technical assistance with this study.