Germline Variation in Apoptosis Pathway Genes and Risk of Non–Hodgkin's Lymphoma

Non–Hodgkin's lymphoma (NHL) is the fifth most common cancer overall in the United States, and the lifetime odds of developing NHL is 1 in 45 for men and 1 in 53 for women (1). The remarkable rise in incidence of NHL over the last 50 years suggests a major role for environmental factors in the etiology of this cancer. However, established risk factors to date account for only a relatively small fraction of the cases (2).

The t(14;18)(q32;q21) translocation is a hallmark translocation in follicular lymphoma (FL; ref. 3), one of the most common lymphoma subtypes (2). With this translocation event, BCL2 becomes fused to the immunoglobulin heavy chain (IgH) locus, leading to constitutive Bcl-2 expression and apoptosis inhibition under the control of the IgH enhancer (4, 5). Under normal conditions, lymphocytes must strictly regulate growth and apoptosis to provide adequate immunologic defenses against infections while not overwhelming the organism with inappropriate cell numbers. With the t(14;18)(q32;q21) and other less commonly observed translocations of genes in the apoptosis pathway observed among lymphoma cases (6, 7), it is clear that dysregulation of the balance between cell proliferation and programmed cell death is a central feature in lymphomagenesis (8). Furthermore, evidence that the t(14;18)(q32;q21) translocation is also present in ∼30% of diffuse large B-cell lymphomas (DLBCL) and in 1% to 2% of chronic lymphocytic leukemias/small lymphocytic lymphomas (CLL/SLL; ref. 6), and the deletion/downregulation of Bcl-2 inhibiting microRNA species (mir-15 and mir-16) in CLL (9), suggests a broad role for Bcl-2 and apoptosis in lymphoma.

Bcl-2 is a member of a large family of proapoptotic and antiapoptotic proteins, which coordinate both extrinsic and intrinsic cell signals to activate caspases, the effector enzymes necessary for apoptosis execution (4, 5, 10). The high prevalence of the t(14;18) translocation among healthy individuals, estimated as high at 66% at age 50, would indicate that perhaps overexpression of the Bcl-2 protein as a result of this transformation may not sufficient for malignant transformation (11-13). There is accumulating evidence that other Bcl-2 family proteins, caspase family proteases, and genes that encode and regulate their transcription are important in lymphomagenesis. Somatic mutations in many caspase genes, including CASP3, CASP7, CASP8, and CASP10, have been documented in a wide variety of human cancers including NHL (14, 15). Furthermore, there is evidence of differential expression of both caspase genes and Bcl-2 family member genes among the NHL subtypes (16-18).

The above-mentioned studies have primarily focused on genetic events that affect expression or function of apoptotic proteins within the tumor. However, accumulating epidemiologic evidence suggests that germline genetic variation also plays a role in NHL etiology (19-21). Common variation related to B-cell growth and survival (22), inflammation and immune function (23-28), and DNA repair (29) has been linked to NHL risk. Furthermore, recent genome-wide association studies have identified novel single-nucleotide polymorphisms (SNP) that are associated with risk of developing CLL and FL (30, 31). To our knowledge, only one other group has comprehensively evaluated the role of germline genetic variation in the apoptosis pathway with regard to NHL etiology, including 8 BCL2 family members and 12 caspase family members, in a pooled analysis of three independent case-control studies with a total of 1,946 cases and 1,808 controls (32, 33). Statistically significant gene-level associations of BCL2L11, CASP1, CASP8, and CASP9 with NHL risk were identified.

Here, we independently evaluated the hypothesis that germline genetic variation in genes from the apoptosis pathway is associated with risk of developing NHL, and compare these results with those reported in the pooled study in an attempt to validate the relevance of this pathway in NHL etiology. The 36 candidate genes evaluated (Table 1; Supplementary Table S1) are proapoptotic or antiapoptotic, and are represented in both the intrinsic and extrinsic apoptotic pathways.

This study was reviewed and approved by the Human Subjects Institutional Review Board at the Mayo Clinic, and all participants provided written informed consent. Full details of this ongoing, clinic-based, case-control study conducted at the Mayo Clinic in Rochester, Minnesota have been previously reported (27). This analysis is based on phase 1 of the study, which includes participants enrolled from September 1, 2002 to September 30, 2005. Briefly, eligible patients were within 9 months of their first NHL diagnosis, ages 20 years or older, and residents of Minnesota, Iowa, or Wisconsin at the time of diagnosis. All cases were reviewed and histologically confirmed by a hematopathologist and classified according to the WHO criteria (34). Of the 956 eligible cases, 626 (65%) participated in the study. Clinic-based controls were randomly selected from Mayo Clinic Rochester patients ages 20 years or older, who were residents of Minnesota, Iowa, or Wisconsin, and were being seen for a prescheduled medical examination in the general medicine divisions of the Department of Medicine. Patients were not eligible if they had a history of lymphoma, leukemia, or HIV infection. Controls were frequency matched to cases by 5-year age group, gender, and geographic region (county groupings based on distance from Rochester, Minnesota and urban/rural status). Of the 818 eligible controls, 572 (70%) participated in the study. All participating subjects were asked to complete a self-administered risk-factor questionnaire and to provide a peripheral blood sample for genetic studies. DNA was extracted from blood samples using a standard procedure (Gentra, Inc.).

Genotyping reported here was part of a larger genotyping project to assess the role of immune and other candidate genes in the etiology and prognosis of NHL (27). Most of the genes and SNPs reported here were from the ParAllele (now Affymetrix) Immune and Inflammation SNP panel that included 1,253 genes and 9,412 SNPs (35). The Immune and Inflammation panel was supplemented by a second round of genotyping using a custom Illumina Goldengate (36) OPA that included 384 SNPs from 100 candidate genes. Full genotyping details and quality control measures for both of these genotyping platforms have been previously described (27, 28). Briefly, tagging SNPs were selected using CEPH (European-American) and Yoruba (African) samples from release 16 (Immune and Inflammation panel) and 21 (Illumina panel) of the HapMap Consortium (37). Tagging SNPs covered 5 kb upstream and downstream of each gene with minor allele frequency (MAF) of ≥0.05 and pairwise r² threshold of 0.8. Across both platforms, the overall sample success rate was >98%, the assay call rate was >93% (99.1% for ParAllele and 93.5% for Illumina), and the concordance rate of sample duplicates was >98%; the concordance rate among the 71 SNPs that were duplicated across the two platforms was 99.7%. A total of 916 people (441 cases and 475 controls) were genotyped in both assays and passed all quality control measures (28). This combined master data set was restricted to subjects who reported their race as Caucasian. After the duplicate SNPs with the lower platform-specific SNP call rates were dropped and SNPs that had a MAF of <1% (n = 935) were excluded, 8,034 SNPs remained in the data set. For this analysis, we evaluated 226 SNPs (Supplementary Table S1) from 22 BCL2 and 14 caspase family genes (Table 1).

Allele frequencies from cases and controls were estimated using observed genotype frequencies. The frequencies in the controls were compared with genotype frequencies expected under Hardy-Weinberg equilibrium (HWE) using a Pearson goodness-of-fit test or Fisher's exact test (MAF < 0.05). In this analysis, 14 of the 226 evaluated SNPs had a HWE P < 0.05 (Supplementary Table S2); because no genotype calling errors were identified, these SNPs were not excluded from analysis. We previously found no evidence of population stratification in our data (27).

Two methods were used when analyzing the association between each gene and case-control status. The first approach used a principal components analysis to create orthogonal (e.g., uncorrelated) linear combinations of the SNP minor allele count variables that provide an alternate, and equivalent, representation of the SNP genotype count variables. These component linear combinations were then ranked according to the amount of the total SNP variance explained. The resulting smallest subset that accounted for at least 90% of the variability among the SNPs was included in a multivariable logistic regression model. A gene-specific global test using the resultant principal components was then carried out using a multiple degree-of-freedom likelihood ratio test. This method decreases the dimensionality of the data when SNPs are correlated by reducing the number of independent degrees of freedom that comprise the statistical test. The second method in which the gene-level association was tested used the global score test of Schaid et al. (38) as implemented in the S-Plus program Haplo.stats. Because the haplotype results were similar to the principal components analysis, we only report these results. Gene-level tests with P < 0.05 were declared of interest.

Individual SNPs were examined using unconditional logistic regression to estimate odds ratios (OR) and corresponding 95% confidence intervals (CI) separately for heterozygotes and minor allele homozygotes, using homozygotes for the major allele as the reference. ORs and corresponding 95% CIs were also estimated per copy of variant allele for each SNP, and P_trend was calculated assuming an ordinal (log-additive) genotypic relationship. SNPs with P_trend < 0.05 in the setting of a global gene test of P < 0.05 were declared of interest. We also evaluated the association between SNPs in genes of interest with NHL risk by major NHL subtype (DLBCL, FL, and CLL/SLL). We used polytomous logistic regression to simultaneously calculate ORs and 95% CIs for each subtype relative to controls, and to formally test for heterogeneity of the estimated association between each SNP of interest and lymphoma subtype (39).

To assess the robustness of our results in the setting of multiple hypothesis testing, we used the tail strength methods of Taylor and Tibshirani (40) at both the gene and SNP level. This method tests the global null hypothesis that the distribution of P values from a large set of univariate tests is uniformly distributed. As such, positive tail strength values significantly >0 indicate that the observed number of small P values is greater than would be expected by chance alone. In addition, we have also estimated q values at the SNP level to estimate the strength of the association with respect to the positive false discovery rate (pFDR; ref. 41).

To allow for SNP-level comparison with previously published associations between BCL2 (33) and caspase (32) family genes from a pooled analysis of three case-control studies, we used the MACH 1.0.14 to impute genotypes for SNPs not directly observed in our study population (42). The 60 unrelated HapMap CEU samples (from release 23a/phase II Mar08, NCBI Build 36, dbSNP build 12) were used to obtain the phased chromosomes, and the expected genotype dosage was computed based on the posterior probability. SNPs with imputation r² larger than 0.30 were deemed of sufficient quality and were examined for their association with NHL case-control status using the SNP dosage estimated from MACH.

Analyses were implemented using SAS (version 8; SAS Institute) and S-Plus (version 7.05; Insightful Corp.) software systems. All analyses were adjusted for age and gender.

There were 441 cases and 475 controls available for analysis. For cases, the mean age was 60.1 years and 58% were male, whereas for controls the mean age was 61.7 years and 55% were male. Additional patient characteristics have been previously published (28). The most common NHL subtypes were SLL/CLL (n = 123), FL (n = 113), and DLBCL (n = 69).

We first evaluated gene-level associations between the 36 candidate apoptosis pathway genes and NHL, all subtypes combined. Using principal components analysis, we observed four genes to be significantly associated with NHL risk at P < 0.05 (Table 1): BAG5 (P = 0.026), BCL2L11 (also known as BIM; P = 0.0019), BCLAF1 (P = 0.0097), and CASP9 (P = 0.0022). In addition, BCL2, BCL2L13, BCL2L14, BID, APAF1, CASP7, CASP10, and DFFB each had one or more SNPs at P < 0.05 but in the setting of a gene-level test of P ≥ 0.05, and thus were not considered further. All SNP-level associations from nonsignificant genes are available in Supplementary Table S2.

Next, we formally evaluated SNP-level associations within the four genes with P < 0.05 from the gene-level analysis (Table 2). The single genotyped BAG5 tagSNP (rs7693) was significantly associated with NHL risk at P < 0.05 (OR, 1.24 per T allele copy; 95% CI, 1.02-1.50). For BCL2L11, two of the five tagSNPs were significant at P < 0.05, and variant alleles were associated with decreased NHL risk for both (rs6746608: OR, 0.82 per A allele copy; 95% CI, 0.68-1.00; rs12613243: OR, 0.58 per C allele copy; 95% CI, 0.38-0.87). There was little evidence of correlation between these two BCL2L11 SNPs (r² = 0.046), and both genotyped SNPs remained statistically significant when modeled in a single logistic regression model, confirming that these two SNPs represent separate NHL risk signals (P = 0.011 and P = 0.0015, respectively). Both genotyped BCLAF1 tagSNPs were significantly associated with NHL risk (rs797558: OR, 1.38 per G allele copy; 95% CI, 1.07-1.80; rs703193: OR, 1.42 per T allele copy; 95% CI, 1.10-1.84). These two SNPs were in strong linkage disequilibrium with each other in our population (r² = 0.96). Neither SNP reached statistical significance when both were modeled in the same logistic regression model, indicating that they represent the same signal related to NHL risk. Finally, three of the seven genotyped CASP9 tagSNPs were significant (rs6685648: OR, 1.41 per C allele copy; 95% CI, 1.14-1.73; rs2020902: OR, 0.74 per C allele copy; 95% CI, 0.57-0.95; rs2042370: OR, 0.82 per C allele copy; 95% CI, 0.68-1.00). The pairwise correlation of these SNPs is fairly low (r² between 0.079 and 0.34); however, they seem to represent the same signal when jointly modeled. That is, when we added either rs2020902 or rs2042370 to the logistic regression model with the most significant SNP, rs6685648, the added SNP was nonsignificant based on the likelihood ratio test (P > 0.1). Of note, all of the SNPs that were significantly associated with NHL risk were intronic, with the exception of the BAG5 SNP rs7693, which is in the noncoding region interval of an mRNA transcript.

To evaluate the effect of multiple testing, we estimated both the tail strength of the P values generated in the gene- and SNP-level analyses and the q values with respect to the pFDR for each SNP. The tail strength for the 36 gene-level P values was 0.38 (95% CI, 0.05-0.70), whereas the tail strength estimate for the 226 SNP-level P values was 0.20 (95% CI, 0.07-0.33). As the tail strength estimates and 95% CIs exclude the null in both the gene- and SNP-level analyses, we can conclude that the distribution of P values is more extreme than we would have expected by chance alone, and thus, observed significant associations at both the gene and SNP level remain noteworthy. However, individual SNP q values within the genes of interest (Table 2) ranged from 0.22 to 0.38 in the SNPs we have considered significant for the purpose of this analysis.

In exploratory analyses, similar associations (as assessed by direction and magnitude of ordinal ORs and P_{heterogeneity} obtained from polytomous logistic regression) were observed for significant SNPs from BCLAF1 and BAG5 for the subtypes of CLL/SLL, FL, and DLBCL. In contrast, there is some evidence that the associations between individual tagSNPs within BCL2L11 and CASP9 differ among the three NHL subtypes (Table 3). For BCL2L11, the decreased risk of NHL with copies of the variant A allele at rs6746608 seems similar across NHL subtype (P_{heterogeneity} = 0.57), but the decreased NHL risk with copies of the variant C allele at rs12613243 seems limited to CLL/SLL and FL (P_{heterogeneity} = 0.047). The pattern for CASP9 SNPs seemed most differential by subtype. The decreased NHL risk with copies of the A allele at both rs4646077 and rs2020902 was limited to DLBCL (P_{heterogeneity} = 0.049 and 0.17, respectively); the decreased NHL risk with copies of the A and C variant alleles at rs4646018 and rs2042370, respectively, was limited to CLL/SLL (rs4646018, P_{heterogeneity} = 0.0098; rs2042370, P_{heterogeneity} = 0.0046); the remaining CASP9 SNPs had P_{heterogeneity} > 0.25. However, caution is warranted in interpreting the subtype results due to small sample sizes, especially SNPs with lower MAFs.

There were 17 genes that were evaluated in both this study and the pooled analysis of three case-control studies from the United States and Australia (32, 33): 13 genes that were not associated with NHL risk in either study (BAX, BCL2, BCL2A1, BCL2L1, BCL2L2, BCL2L10, CASP2, CASP3, CASP4, CASP5, CASP6, CASP7, and CASP10), 2 genes (CASP1 and CASP8) that were associated with NHL risk in the pooled study but were not associated with NHL risk in our study, and 2 genes (BCL2L11 and CASP9) that were significantly associated at the gene level with NHL risk in both studies. For these two genes (BCL2L11 and CASP9), we compared individual SNP-level significance (based on ordinal ORs) between these two studies. The results for the overlapping observed and imputed SNPs are presented in Table 4. Figure 1 illustrates the relative position and linkage disequilibrium (based on HapMap CEPH population genotypes) of each BCL2L11 (A) and CASP9 (B) genotyped SNPs.

For BCL2L11, there was no overlap in the SNPs genotyped in the two studies. For the Mayo case-control study, we imputed genotypes for the four SNPs observed to be significantly associated with NHL risk in the pooled study (rs7567444, rs3789068, rs686952, and rs6760053). The estimated NHL risk with variant allele copies was similar in direction and magnitude across all four SNPs, and reached significance for rs3789038 (P_trend = 0.0018) and rs6760053 (P_trend = 0.0031). In exploratory analyses of these four SNPs by NHL subtype, we observed that the association between the imputed SNP genotypes in the Mayo population were strongest in the CLL/SLL subtype but were also observed in FL at a magnitude consistent with the pooled study, which observed an association specific to the FL subtype (Table 5).

The CASP9 SNP rs2020902 was genotyped in both studies, although the results were inconsistent. The observed association between copies of the variant G allele and NHL risk was OR = 0.74 (95% CI, 0.57-0.95; P_trend = 0.019) in the Mayo case-control study and OR = 1.02 (95% CI, 0.89-1.16; P_trend = 0.82) in the pooled study. Alternatively, two CASP9 SNPS were observed to be significantly associated with NHL risk in the pooled study (rs4661636 and rs4646047), and the magnitude and direction of the estimate associated with Mayo imputed genotypes was consistent for both. An inverse association with the number of variant T alleles at both rs4661636 and rs4646047 was observed in both studies, although associations did not reach significance in the Mayo case-control study.

The pooled study did not genotype BCLAF or BAG5, so we were not able to compare our findings for these two genes. Further, we did not have the original genotyping data from the pooled analysis, so we were not able to impute genotypes in their population for comparison with SNPs significant in the Mayo case-control study.

In this case-control study, we show gene- and SNP-level association of BCL2L11 (BIM), BCLAF1, BAG5, and CASP9 with NHL risk that remained noteworthy after accounting for multiple testing. Whereas the pFDR q values estimated for BCL2L11 and CASP9 SNPs do indicate that there is a moderate chance that any of these associations individually may be false positives, the tail strength estimates indicated that the distribution of P values for the group of SNPs from these genes is more extreme than we would have expected by chance alone. Moreover, the significant gene-level associations for BCL2L11 and CASP9 were consistent with previously published data from a pooled analysis of three studies [National Cancer Institute–Surveillance, Epidemiology, and End Results (NCI-SEER), Connecticut, and New South Wales] that included 1,946 cases and 1,808 controls (32, 33). Although there was minimal overlap in the individual tagSNPs genotyped across our study and the pooled study, we were able to impute genotypes for the SNPs in our study that were not directly observed, and we found associations that were largely consistent in magnitude and direction across all four significant BCL2L11 SNPs and two of three CASP9 SNPs from the pooled study, making this the first independent replication of these results. The associations from our study with BCLAF1 and BAG5 have not been tested in an independent data set and thus require replication.

Strengths of our study include a carefully designed case-control study, central pathology review, and high quality genotyping. Although this study was not population based, both case and control participation was restricted to those residing in the region surrounding Mayo Clinic (Minnesota, Iowa, and Wisconsin), thus minimizing the effect of referral bias and increasing the internal validity of using frequency-matched general medicine controls from the same region. Common HapMap SNPs were used to tag genes of interest, and through other genotyping projects, we have ruled out the presence of significant population stratification (27). The major limitations are the use of an exclusively Caucasian population, which limits generalizability, and the relatively small sample size, which in particular precludes robust estimation of NHL subtype associations. There are several apoptosis genes evaluated in this study for which the genotyped tagSNPs provided <70% gene coverage, and thus, an association between these genes and NHL risk cannot yet be ruled out. In addition, there are >100 human genes more broadly identified with apoptosis pathway involvement in which germline variants may play a role in NHL risk, including genes with caspase recruitment domains, death domains, and death effector domains (43). The genes in the current analysis represent only the core of a larger set of genes involved in apoptosis-related pathways.

The two genes that replicated have strong biological plausibility in NHL pathogenesis. BCL2L11 balances the antiapoptotic influence of BCL2 and coordinates proapoptotic signaling through the intrinsic apoptosis pathway (4). In addition, BCL2L11 is required for negative selection of autoreactive lymphocytes (44). Functional silencing of BCL2L11 through methylation has been observed in Burkitt lymphoma cell lines and primary tumor biopsies, and reduced BCL2L11 mRNA and protein expression has also been documented in other tumors including renal cell carcinoma, melanoma, and colon cancer (45). Of note, we did not identify any association between BCL2 and NHL risk at the gene level, and only 1 of the 53 genotyped SNPs was significant at the SNP level. Although these 53 SNPs comprised only 74% gene coverage, it does suggest that BCL2 germline variation may not play a role in NHL risk. This would be consistent with the hypothesis that most of the BCL2 variation in lymphoma tumors is a result of hypermutation following the t(14;18) translocation event (4). This hypothesis should be further explored by comparing somatic and germline mutation among patients stratified by the t(14;18) translocation.

CASP9, the other gene to be replicated, is a proapoptotic protease integral to the intrinsic apoptotic pathway, and is responsible for effector caspase activation and apoptosis execution following activation by Apaf-1 bound to cytochrome c released from mitochondria (46). Hyperphosphorylation of caspase-9 may lead to aberrant apoptosis inhibition, and the relevance of this process has been shown in several other cancer types (46). Also of note, on mitochondrial release of cytochrome c via both the intrinsic and extrinsic apoptosis pathways, the cytoplasmic protein Apaf-1 binds caspase-9 to form the apoptosome, in turn activating the caspase cascade (4). APAF1 did not reach gene-level statistical significance in our population, although 6 of the 13 genotyped tagSNPs in this gene were significant at P < 0.05. Given the gene-level significance of CASP9, there may be some clinical relevance of the individual tagSNP significance in APAF1, and further follow-up on this gene is warranted.

To our knowledge, this is the first report of an association between germline variation in BAG5 and BCLAF1 with regard to lymphoma risk. BCLAF1 and BAG5 are both Bcl-2 family members that suppress BAX (proapoptotic) gene expression, in turn suppressing the APAF1 gene and inhibiting apoptosis (4, 47). These associations should be confirmed in follow-up with an independent study population.

Although underpowered to assess NHL subtypes, our data provide some evidence that there may be subtype-specific associations in the apoptosis pathway, particularly BCL2L11 and CASP9. Evidence of differential association of germline variants in the BCL2 and CASP families was observed in the pooled analysis of three case-control studies (32, 33), although the BCL2L11 association was largely limited to FL (33), a pattern we observed for both FL and CLL/SLL in our study.

In conclusion, our results support an association of four genes from the apoptosis pathway, with NHL risk, and these associations may vary by NHL subtype. In light of the importance of the apoptosis pathway to human lymphomagenesis, further characterization of the key players within this pathway is warranted.

No potential conflicts of interest were disclosed.

We thank Sondra Buehler for her editorial assistance.

Grant Support: NIH, National Cancer Institute grant R01 CA92153. J.L. Kelly was supported by NIH, National Heart Lung and Blood Institute grant HL007152.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.