Abstract
Background: A pressing clinical issue in prostate cancer is to distinguish which men will have an indolent or aggressive course of disease. Clinical variables such as Gleason grade and stage are useful predictors of lethal cancer; however, the low predictive values of the common Gleason scores, changes in grading over time, and earlier diagnosis of patients due to screening limits their clinical utility. Identifying genetic variants associated with lethal prostate cancer could inform clinical decision making.
Methods: We conducted a genome-wide association study, comparing lethal prostate cancer cases to cases surviving at least 10 years beyond their initial diagnosis. Genotyping was done with the Affymetrix 5.0 chip [∼500,000 single nucleotide polymorphisms (SNP) and 1,483 copy number variants (CNV)] on DNA from participants in the Physicians' Health Study and Health Professionals Follow-up Study (196 lethal cases, 368 long-term survivors). After excluding SNPs and individuals based on quality control criteria, logistic regression assuming an additive model was done using the PLINK software.
Results: No SNP reached genome-wide significance (P ≤ 1 × 10−7); however, three independent SNPs had P < 1 × 10−5. One top-ranked SNP replicated (P = 0.05) in an independent follow-up study. Although no CNV had genome-wide significance, 14 CNVs showed nominal association with prostate cancer mortality (P < 0.05).
Conclusions: No variants were significantly associated at a genome-wide level with prostate cancer mortality. Common genetic determinants of lethal prostate cancer are likely to have odds ratios <2.0.
Impact: Genetic markers identified could provide biological insight to improve therapy for men with potentially fatal cancer. Larger studies are necessary to detect the genetic causes of prostate cancer mortality. Cancer Epidemiol Biomarkers Prev; 19(11); 2869–76. ©2010 AACR.
Introduction
One of the most urgent clinical questions in prostate cancer is how to predict an individual's course of disease at the time of diagnosis. Prostate cancer is the most common incident cancer (other than nonmelanoma skin cancer) and the second leading cause of cancer mortality in men in the United States (1). However, the vast majority of prostate cancer patients will not die from their cancer. Although early detection and treatment play a role in cancer survival, some treated individuals still succumb to prostate cancer whereas many survive without medical intervention. A recent large trial found that men randomized to prostatectomy had only a small (although significant) absolute reduction in prostate cancer death compared with those randomized to watchful waiting (2). Albertsen et al. (3) followed 767 men with conservatively treated localized prostate cancer for over 20 years and observed that the majority of men (70%) did not die of prostate cancer.
What causes one prostate cancer patient to develop metastases or die from their cancer while others survive with the disease for many years? At present, the most utilized predictors of outcome at diagnosis are age, clinical stage, prostate-specific antigen level, and Gleason score. Gleason score, a measure based on the histologic patterns of prostate tumors, is currently one of the best predictors. In a study using re-reviewed Gleason score from prostatectomy specimens, those with Gleason 8 cancers had a hazard of lethal cancer (dying from prostate cancer or developing distant metastases) that was 7.4 [95% confidence interval (95% CI), 2.5-22] times higher than those with Gleason 3+4; cases with Gleason 9 to 10 had an even higher risk of lethal cancer (hazard ratio, 19.1; 95% CI, 7.4-49.7; ref. 4). However, the positive predictive value for mortality of a higher Gleason score, including the most common Gleason 7 as well as Gleason 8 to 10, is only 29% (5), and therefore far from optimal. Gleason score has additional limitations as a predictor because of scoring changes over time (6, 7) and interobserver variability (8, 9).
Epidemiologic and experimental evidence supports the hypothesis that aggressive cancer has an inherited component. A recent study showed concordance of survival and prostate cancer mortality among fathers and sons with prostate cancer, implying that prognosis itself may have a hereditary component (10). Laboratory experiments using a highly metastatic mouse mammary model crossed with several different strains showed that the genetic background of an animal can influence the metastatic efficiency (11). Further quantitative trait mapping work identified regions on chromosome 19 that were significantly associated with metastatic efficiency, suggesting that inherited variation may influence metastasis (12). Thus far, genetic studies in humans have focused on Gleason score as a proxy for aggressive disease. Several regions have been implicated in linkage scans, but three of the regions (5q31-33, 7q31-33, and 19q12-q13.3) were strongly significantly associated with high-grade cancer (P < 0.001) and replicated in at least two independent studies, suggesting that a locus may be present under these peaks (13-16).
In prostate cancer genetic association studies for risk, a combined analysis of two genome-wide association studies (GWAS) identified a variant at chromosome 9q33.2 in a putative tumor-suppressor gene (DABP2IP) that was associated with risk of aggressive prostate cancer, defined by Gleason grade and clinical stage (17). Another study found that the TT genotype of rs4054823 at 17p12 increased the risk of aggressive cancer compared with nonaggressive cancer, again defined by clinical variables (18). A germ-line deletion at 2p24.3 was more strongly associated with the risk of aggressive cancer than nonaggressive cancer (19).
Substantial longitudinal follow-up is required to capture information on prostate cancer mortality; thus, this outcome is studied less frequently than Gleason score. However, we believe that a large-scale genetic study for the most important prostate cancer outcome is crucial to improve our understanding of prostate cancer aggressiveness. We therefore performed a GWAS for prostate cancer mortality in the Physicians' Health Study (PHS) and Health Professionals Follow-up Study (HPFS), with a replication study in the Dana-Farber Harvard Cancer Center Specialized Programs of Research Excellence (SPORE; Gelb Center) case series. In addition to examining the association of genotypes, we also evaluated whether copy number variants (CNV) were associated with prostate cancer mortality.
Materials and Methods
Study Population
Physicians' Health Study.
The PHS began as a randomized, double-blind trial of aspirin and β-carotene in the prevention of cardiovascular disease and cancer among 22,071 healthy U.S. physicians. Written consent was obtained from each participant at the time of initial enrollment, and the investigation was approved by the Human Subjects Committee at Brigham and Women's Hospital. Men were excluded if they had any serious medical conditions at baseline, including all cancers (except nonmelanoma skin cancer). Blood samples were collected from 68% of the physicians in 1982 to 1984, as described previously (20).
Participants are followed through annual questionnaires to collect data on diet, health and lifestyle behaviors, and medical history, and biannually through postcards to ascertain health end points, including prostate cancer. All self-reported prostate cancer cases are verified through medical record and pathology review. Through this systematic medical record review, we also abstract data on clinical information, such as Gleason score. Cause of death is determined by review of death certificates, medical records, and information from the family by a panel of three physicians. There is a high follow-up rate for both cancer incidence (96%) and mortality (98%). Metastases are reported on follow-up questionnaires sent to all men living with prostate cancer and are confirmed through medical record review.
For the current study, we included incident prostate cancer diagnosed between 1982 and 2003, and restricted participants to self-reported Caucasians to reduce potential population stratification. Due to cost restraints, we were unable to genotype all PHS prostate cancer patients from whom blood had been collected. We therefore examined the two extremes of prostate cancer cases: long-term survivors (patients who survived a minimum of 10 years after diagnosis until death or end of follow-up (March 1, 2008) and did not develop metastases to bone or organs or die from prostate cancer; n = 415) and lethal prostate cancer cases (patients who developed metastases to bone or organs after diagnosis or died from prostate cancer; n = 176).
Health professionals follow-up study.
The HPFS, an ongoing prospective cohort study on the causes of cancer and heart disease in men, consists of 51,529 U.S. health professionals who were of ages 40 to 75 years in 1986 (21). At baseline and then biennially, participants respond to a mailed questionnaire that included questions on demographics, lifestyle, and medical history. Between 1993 and 1995, 18,018 of the men provided a blood specimen. When a participant reports a prostate cancer diagnosis, medical and pathology records are obtained. Study investigators review these records to confirm the diagnosis and to abstract stage at diagnosis and Gleason grade. Deaths among cohort members are identified by reports from next-of-kin, the postal service, or searches of the National Death Index. To increase the number of lethal cancers in this study, we included 46 prostate cancer deaths from the HPFS (self-reported Caucasian) among cases diagnosed between 1993 and 2000; these were selected from a larger nested case-control study and had the most available DNA from a total of 53 prostate cancer deaths.
Dana-Farber Harvard Cancer Center SPORE (Gelb Center).
The Gelb Center is a case series of prostate cancer patients diagnosed between 1976 and 2007. A detailed description of this study has been published previously (22). The study captures detailed clinical information from multiple sources, including medical records and patient registration, and a blood sample collected after diagnosis. Follow-up of the participants occurs at clinic visits to the Dana-Farber Cancer Institute and by searching the National Death Index. Because cause of death is not always available or known, if an individual was known to have metastases before their death, they were assumed to have died from prostate cancer. For this study, after restricting to self-reported Caucasians, we included 155 long-term survivors (end of follow-up July 1, 2007) and 500 lethal cases as a replication set.
Genome-wide association study
Affymetrix scan.
The samples from the PHS and HPFS were included in the genome scan. DNA was extracted from peripheral blood samples for all participants. Genotyping was done with the Affymetrix 5.0 single nucleotide polymorphism (SNP) chip, which contains probes for 500,568 SNPs. Briefly, ∼500 ng of DNA from each sample is digested with Nsp and Sty restriction enzymes. The digested segments were ligated to enzyme-specific adaptors that incorporate a universal PCR priming sequence; PCR amplification by using universal primers was done in a reaction optimized to amplify fragments. The products are fragmented, end-labeled with biotinylated nucleotides, and hybridized to a chip and detected (23). The resulting intensities for each allele are used to make a genotype call. The “Birdseed” calling algorithm, an updated version of the Robust Linear Modeling using Mahalanobis distance calling algorithm developed at the Broad Institute of Harvard and Massachusetts Institute of Technology, was used for this study (24). More information on the technology, calling algorithm, and SNP coverage can be found in ref. (25).
Samples and quality control.
A total of 637 unique samples from PHS and HPFS were included in this study. Deaths and long-term survivors were interspersed across seven 96-well plates, and laboratory personnel were blinded to outcome. Each plate had two empty wells (negative controls) as well as two duplicates to be used for quality control.
We assessed the genotype concordance of 458 SNPs from 500-kb regions of chromosomes 1, 5, 10, 15, and 20 for the 14 duplicate pairs (concordance 99.9%). We also compared the genotype calls for 31 SNPs that had previously been genotyped on a subset of these PHS participants; concordance was 99.3% for >14,000 genotypes.
Data analysis.
The PLINK program (26) was used to analyze these genome scan data (27). Forty-six individuals (33 long-term survivors, 13 deaths) with <90% of genotype calls made were removed from the analysis; the average call rate in the remaining individuals was 98.8%. Of the SNPs genotyped, SNPs missing >10% of genotypes (14, 704), with minor allele frequency <1% (68, 603), or with Hardy-Weinberg equilibrium P < 1 × 10−6 (1, 979) were excluded, leaving 419,613 SNPs for analysis.
To address potential remaining population stratification, we used the Eigenstrat program (28). We ran this program for all participants with the default parameters (5 outlier iterations across the top 10 eigenvectors, with outliers exceeding 6 SD along a top principal component excluded), and output the first two eigenvector values. Several individuals were not assigned values along these eigenvectors due to missing data (as described above) or were designated outliers (14 long-term survivors, 12 deaths); these individuals were excluded from further analysis. Using PLINK, for the main analysis, we ran an unconditional logistic regression model adjusting for the first two eigenvectors (excluding one HPFS death missing age at diagnosis), outputting the additive model results for the association of each SNP with lethal prostate cancer (n = 196) versus long-term survival (n = 368). We then ran secondary analyses additionally adjusting for age at diagnosis and restricting to men with a Gleason score of 7.
Follow-up study
The Gelb Center samples were used for a genetic replication study. We selected and designed assays for SNPs with P < 1 × 10−3 that fell in previously identified linkage peaks for Gleason score (chr5q31-33, n = 6; chr7q31-33, n = 1). We then selected markers to capture the independent variation with P < 5 × 10−4 (n = 72). Genotyping was done with Sequenom iPLEX matrix-assisted laser desorption/ionization-time of flight mass spectrometry technology (see ref. 29 for reaction details). The association of the additive model of these SNPs with lethal prostate cancer versus long-term survival was done using unconditional logistic regression. SNPs were excluded from analysis if they had a <90% genotyping success rate. Of the 79 SNPs genotyped in the Gelb samples, 11 failed genotyping quality control. Replication was declared only if P ≤ 0.05 and the direction of the effect was the same as in the GWAS; for the replicated SNP, a joint analysis with the original GWAS data was done as a meta-analysis with a random-effects model. Analysis was done with SAS v9.1 statistical software.
Copy number polymorphism analysis
We analyzed SNP chip-based copy number polymorphism data as generated by the CNV detection software Canary (30) in the form of summarized intensity scores for 1,483 CNVs and 565 subjects. We followed the subject filtering criteria as described above in our genotype analysis; individuals who were missing considerable data or were found to be outliers by Eigenstrat were excluded. Then, we followed a likelihood ratio approach for testing association between each CNV and the binary status of mortality considered as a trait. The approach jointly fits two linear models, as outlined in Barnes et al. (31), and is described as follows. The first model classifies the summarized intensities for each CNV by fitting a finite mixture of Gaussian densities via an Expectation-maximization-based algorithm that uses Bayesian information criterion to select the optimal number of classes. Upon convergence, the classification assigns every individual subject to a copy number genotype. Given an optimal model with multiple copy number classes, we tested for its association with the subject's trait with this joint model. This is done by fitting of a generalized logit linear model to test the null hypothesis H0 that there is no association between a subject's copy number genotype and his binary prostate cancer mortality trait (in this case, lethal/indolent). If the fitting is correct and there is indeed no association, then the computed likelihood ratio statistic is χ2 distributed with 1 degree of freedom, which leads to a corresponding P value of association. The plots and statistics for CNV classification and the associated distribution of trait were generated with the BioConductor package CNVtools.
Results
GWAS results
A description of the PHS and HPFS participants is provided in Table 1. Although participants were restricted to self-reported Caucasians, residual population stratification was addressed with the Eigenstrat program (28). The correlation of eigenvectors 1 and 2 with outcome status was 0.046 and 0.009, respectively, demonstrating that the overall population structure was not strongly related to outcome; the first two eigenvectors for the lethal prostate cancer cases and long-term survivors are shown in Supplementary Fig. S1.
Description of GWAS and replication study participants
A set of 419,613 SNPs passed quality control and were used for subsequent analyses (see Materials and Methods). A quantile-quantile plot of the results compares the χ2 values obtained in this study with the expected distribution under the null hypothesis of no association between genetic variation and mortality (Fig. 1). Although no SNPs reached genome-wide significance (P ≤ 1 × 10−7), three independent SNPs had P < 1 × 10−5; the plot of P values (Fig. 2) shows that there are peaks on chromosome 2q31.2, 11q12.2, and 11q14.1. The results for all SNPs with P < 1 × 10−3 (n = 277) are provided in Supplementary Table S1.
Quantile-quantile plot comparing observed statistics for all results to those expected based on the null distribution.
P values for the association of SNPs with prostate cancer mortality plotted by chromosome and position.
To determine the associations of SNPs on mortality independent of their possible associations with age at diagnosis, we ran the analysis adjusting for age at diagnosis (continuous) in addition to the top 2 eigenvectors. When adjusting for age at diagnosis, there are 3,767 results with P < 0.01; 19% of these results are not among the 3,803 results with P < 0.01 from our main analysis. However, the top SNPs from the non–age-adjusted results (Supplementary Table S1) all have P < 0.005 in the adjusted analysis, suggesting that the overall effect of SNPs on mortality through age at diagnosis may not be substantial. We also examined the association of SNPs with lethal cancer restricting to cases with Gleason 7; again, no SNPs reached genome-wide significance. With this much smaller number of participants, half of the SNPs with P < 0.001 had P < 0.05 in the main analysis.
We examined the results for previously identified prostate cancer risk SNPs in our scan. Sixteen of the 31 confirmed risk SNPs compiled by Varghese and Easton (32) were either directly genotyped in our scan or had a proxy with R2 > 0.8. The most significant finding was for rs16901979 for which the risk allele decreased the probability of lethal cancer [odds ratio (OR), 0.35; P = 0.006]; all results are reported in Supplementary Table S2.
Replication study results
Because the majority of the top-ranked SNPs from the scan will be false positives, we performed a replication study in the Dana-Farber Harvard Cancer Center SPORE Gelb Center (500 lethal cases, 155 indolent). We selected top-ranked SNPs (P < 10−3) that were located in previously identified Gleason linkage peaks (n = 7). We then selected markers to capture the independent variation with P < 5 × 10−4 (n = 72). Of the 79 SNPs selected, 68 were successfully genotyped. Six of these had P ≤ 0.05, but for five the direction of the effect was not consistent with the scan. The one SNP that replicated with the effect in the same direction, rs6973814 (OR, 1.95; 95% CI, 1.01-3.79; P = 0.05), was ranked 66th in the original GWAS (OR, 3.07; P = 0.0003) and is located on chromosome 7q11.2 (nearest gene, AUTS2, 600 kb away). In a joint analysis with the scan results, the combined OR was 2.50 (95% CI, 1.60-3.90; P = 6 × 10−5). All Gelb Center results are in Supplementary Table S3.
CNV results
The model fitting results and number of classes for all 1,483 CNVs are provided in Supplementary Table S4. For the CNVs for which the classification (based on iterative Expectation-maximization modeling) converged and produced more than one CNV genotype class (n = 341), we examined the association between the number of copies an individual carries and lethal prostate cancer. Fourteen CNVs had a nominal P < 0.05; however, none remained significant after correction for multiple testing (Supplementary Fig. S2; Table 2).
Significant associations (P < 0.05) between CNVs and prostate cancer mortality
Discussion
A number of recent GWAS and follow-up replication studies have identified over 20 bona fide genetic prostate cancer risk loci (33-40). Importantly, these studies have provided a new look into the biology of developing the disease. Some of these variants have been tested for association with aggressiveness, typically using the Gleason grade as a proxy for aggressive disease. However, identifying genetic determinants of lethal cancer could improve on the current clinical predictive ability at diagnosis. Understanding who would and who would not benefit from intervention could affect the selection of appropriate medical therapy for the individual, preventing unnecessary treatments and the physical and psychological side effects. In addition, the markers themselves may provide biological insight that could lead to improved therapy for those with potentially fatal cancer.
In this GWAS for lethal cancer, although no SNPs reached genome-wide significance, we identified one top-ranked SNP that replicated in an independent population. The closest gene to the one SNP that was replicated is AUTS2. A recent study based on mRNA expression data reported that this gene was included in the top 100 potential genetic mediators for nonrecurrent primary prostate cancer (41), suggesting a possible biological function.
As noted by McCaroll (42), it is increasingly possible to extend GWAS to examine CNVs and their association with disease phenotypes. In recent years, the SNP arrays have been redesigned to contain probes at the majority of CNVs, which in turn take advantage of the recent high-resolution maps of the CNV locations (43, 44). In this direction, the present GWAS was extended to study CNV in the same SNP array data based on 1,483 mapped CNVs by using a robust statistical modeling algorithm for classification. Although no CNV achieved genome-wide significance, we identified 14 CNVs nominally associated with prostate cancer mortality. Subsequent data mining with alternate modeling strategies or larger studies may reveal further associations.
Prostate cancer mortality is one of the most important phenotypes of this disease. Unfortunately, due to the long follow-up time and the cost necessary to obtain this information, few studies have information on survival and cause of death or the numbers of lethal cases necessary to study this outcome. A major strength of this study is its ability to examine the primary prostate cancer end point, lethal disease, with a substantial number of participants from cohorts that have been followed for decades. The top results were then evaluated in a large case series that also captures survival data.
Figure 3 shows that we are only powered to detect relatively strong effects (e.g., OR >2 with minor allele frequency >20%). Although this is a limitation of our study, it also provides insight into the genetic variants involved in prostate cancer aggressiveness. Based on our data, no common variant will have a large effect on aggressiveness, but rather will most likely have the same magnitude of effect as the alleles that have previously been identified for risk. Although our one SNP that replicated had a larger combined OR of 2.5, in the replication dataset alone the OR was 1.95, suggesting that the initial finding is likely overestimating the magnitude of the effect.
Power for genome scan. Power was calculated using the number of cases included in the final analysis (196 lethal prostate cancer, 368 long-term survivors) with a α level of 1 × 10−7 across a range of allele frequencies and additive model ORs.
Another possible limitation (albeit one that exists in all studies of prostate cancer mortality that are conducted in screened and treated populations) is misclassification of the outcome. Individuals who were labeled as having indolent cancer because they survived at least 10 years without developing metastases or dying of cancer may only be in this category because they received aggressive medical treatment, without which they would have died. However, as the results of the Swedish randomized trial of prostatectomy versus watchful waiting suggest, the number needed to treat to save the life of one man with prostate cancer is 19 (45); thus, the potential effect of misclassification is likely to be minimal. Additionally, it is important to investigate if these genetic variants predict prostate cancer mortality independent of clinical variables such as treatment or Gleason score; however, missing data limits our ability to conduct these analyses. We performed an analysis restricting to the most common Gleason score of 7. Although the results were somewhat similar to the overall analysis, a larger future study examining these associations among men with Gleason 7 would be interesting and could identify SNPs that are associated with lethal cancer independent of their effects on Gleason. A limitation in the CNV analysis is the number of probes included on this Affymetrix chip; a more comprehensive study of CNVs with prostate cancer mortality should be done.
Although several SNPs have been identified that are associated with risk of prostate cancer, these SNPs in general have not been found to confer an increased risk of aggressive compared with indolent disease. If lethal prostate cancer indeed have a genetic component, this suggests that genetic variants determining aggressive disease are different from those that confer overall risk. It would be of clinical utility if future studies specifically focused on attempting to differentiate lethal from indolent cancer by using germline genetic scans and follow-up studies.
Disclosure of Potential Conflicts of Interest
T. Kurth, received within the last 2 years investigator-initiated research funding from the French National Research Agency, the NIH, Merck, and the Migraine Research Foundation; consultant to i3 Drug Safety and World Health Information Science Consultants, LLC; received honoraria from Genzyme, Merck, and Pfizer for educational lectures; and received travel funds from the Restless Legs Syndrome Foundation.
Acknowledgments
Grant Support: Grant from the Doris Duke Clinical Scientist Development Award. The laboratory work at the Broad Institute Center for Genotyping and Analysis was supported by grant U54 RR020278 from the National Center for Research Resources. The PHS was supported by grants CA-34944, CA-40360, and CA-097193 from the National Cancer Institute and grants HL-26490 and HL-34595 from the National Heart, Lung, and Blood Institute, Bethesda, MD. The HPFS was supported by CA-55075 from the National Cancer Institute. Additional funding was provided by Dana-Farber/Harvard Cancer Center Prostate Cancer SPORE (P50 CA090381-08). K.L. Penney was supported by a National Research Service Awards (T32 CA009001-34).
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.
Footnotes
Note: Supplementary data for this article are available at Cancer Epidemiology, Biomarkers & Prevention Online (http://cebp.aacrjournals.org/).
- Received June 8, 2010.
- Revision received August 3, 2010.
- Accepted August 10, 2010.
References
Volume 19, Issue 11
