Assessing the Cumulative Contribution of New and Established Common Genetic Risk Factors to Early-Onset Prostate Cancer

Prostate cancer is the second leading cause of cancer mortality in men in the United States. In 2014, it is estimated that 233,000 men would be diagnosed with prostate cancer and 29,480 men would die from the disease (1). The major recognized risk factors for prostate cancer are increasing age, African ancestry, and positive family history.

Approximately 10% of men diagnosed with prostate cancer in the United States are diagnosed with the disease prior to age 56 years (1). Men with early-onset prostate cancer are more likely to be aggressively treated for their disease and more likely to die from their disease compared with men diagnosed with prostate cancer later in life with similar clinical characteristics (2–4). As with most cancers, early intervention in men that need it can significantly increase the rate of survival. Given the controversy surrounding PSA testing, identifying subsets of men that would most likely benefit from early screening would have a major impact on the successful treatment of the disease. Early-onset disease is also an indicator for heritable disease (2, 4, 5). An important question is whether we can use the cumulative information across associated variants to reasonably predict who is most likely to be diagnosed with early-onset prostate cancer.

To date, genome-wide association studies (GWAS), including primarily older men with prostate cancer, have identified more than 60 distinct common loci with modest effects associated with the disease in men of European descent, including 23 new loci identified using 19,662 prostate cancer cases and 19,715 controls included in the PRACTICAL consortium (6–20). Several studies have demonstrated the importance of the previously established common variants to early-onset and familial prostate cancer (12, 21–25). Herein, we first test whether these 23 new variants are associated with early-onset prostate cancer (20). We then assess the aggregate value of these 23 new variants, aggregate value of these new variants plus 40 established variants, and the added value of including information from these 23 new variants to the overall burden of risk alleles from the 40 established variants in predicting early-onset prostate cancer. We demonstrate that the total risk-allele burden across prostate cancer GWAS variants can be useful for identifying a subset of men with substantially increased risk for early-onset disease.

This study includes 931 unrelated early-onset prostate cancer cases (diagnosed prior to age 56 years) of European descent from the University of Michigan Prostate Cancer Genetics Project (UM-PCGP). Descriptive information about the cases is presented in Table 1. The average age of prostate cancer diagnosis in these 931 cases was 49.7 years. Approximately 62% (576/931) of the cases had a reported first or second degree relative with prostate cancer. All UM-PCGP subjects provided written informed consent to participate in the study. The protocol and consent documents were approved by the Institutional Review Board at the University of Michigan Medical School.

Publically available unrelated male controls with GWAS variant data were selected from Illumina's iControlDB database (n = 1,126; ref. 26). Controls were selected to have European reported ancestry and genotype data generated from a GWAS commercial platform similar to the platform used in UM-PCGP cases. Limited descriptive information, including age, gender, and ancestry, on selected iControlDB subjects can be obtained from the Illumina web site. Illumina iControls have not been screened for prostate cancer.

Nine hundred thirty-eight European American UM-PCGP early-onset prostate cancer cases were genotyped at Wake Forest University using the Illumina HumanHap 660W-Quad v1.1 BeadChip. The iControlsDB subjects were genotyped previously using the Illumina HumanHap550v1 or HumanHap550v3 commercial genotyping platforms.

Initial genotyping quality control methodology was uniformly applied to all GWAS variants and samples [see Lange and colleagues (21) for details]. Subjects missing >5% of variant genotyping calls across all GWAS variants were excluded from consideration. European ancestry for all subjects, including controls, was verified using the software ADMIXTURE (27); subjects with apparent misidentified ancestry or mixed ancestry were also removed from the study. Principal component analysis was also performed using the software Eigenstrat (28) on the combined sample of cases and controls using a linkage-disequilibrium (LD) pruned set of GWAS variants common across genotyping platforms for UM-PCGP cases and Illumina iControls.

We performed genotype imputation on the combined case–control sample to obtain genotype data on the 63 variants reported to be associated with prostate cancer in Eeles and colleagues (20) and Goh and colleagues (29) using the software package MaCH (30, 31). Genotype imputation was performed, separately, including variants from HapMap phase II (CEU reference samples), HapMap phase III (CEU + TSI reference samples), and the 1000 Genomes Project (Chromosome X only using all reference samples). For the autosomal variants, preference was given to phase III imputation results when a variant was successfully imputed using both phase II and phase III HapMap samples. To reduce any possible bias in imputed genotype assignments due to different coverage of variants in the case and control participants, only variants that were successfully genotyped in >98% of both the cases and controls were included in the target panel prior to genotype imputation.

Logistic regression models, implemented in Mach2dat (31), were constructed to test the association between early-onset prostate cancer and each of the 23 newly reported prostate cancer variants using entirely imputed genotype data, scored as dosage values (expected number of copies of the minor alleles). The logistic regression models included covariate adjustment for the first 10 principal components derived from the GWAS data. A Bonferroni-corrected significance threshold for a one-sided test (one-sided P < 0.0022), with requirement the direction of effect was consistent with the previous report, was applied to maintain an overall type I error rate of 0.05.

To assess the cumulative burden of the 23 recently identified variants on early-onset prostate cancer, we estimated the total number of risk alleles each subject carries. The risk allele for each variant was defined as the allele associated with increased risk of prostate cancer in Eeles and colleagues (20). For each subject, we calculated two risk scores, one based on the unweighted sum of risk alleles and the other based on a weighted sum, with the weight given to each variant risk allele equal to the natural logarithm of the corresponding variant's reported OR. For all variants, we used imputed genotype data, even if the variant was directly genotyped, to minimize the impact of any missing data on risk allele counts. We assessed, using t tests, whether the unweighted total number of risk alleles was associated with prostate cancer. We repeated these analyses for 40 previously established prostate cancer variants for populations of European descent summarized in Goh and colleagues (ref. 29; see Supplementary Table S1 for variant identities and their respective imputation quality). The individual association results for these 40 variants in UM-PCGP subjects have been reported previously (21). Finally, we calculated weighted and unweighted totals of risk alleles across all 63 variants.

To assess the relative ability to correctly classify subjects (with respect to case–control status), we constructed ROC curves and calculated the corresponding AUC for weighted and unweighted aggregate risk allele counts for the 23 new prostate cancer variants, 40 established prostate cancer variants, and the set of 63 total prostate cancer variants.

We additionally hypothesized that there was a subset of men with relatively extreme values of total risk-allele burden that could have their disease more accurately predicted than men with total risk-allele counts in the middle of the corresponding total risk-allele count distribution. Specifically, we performed two separate categorizations of all subjects based on the distribution of total risk alleles in controls. In the first categorization, subjects were assigned to a decile grouping based on their total risk allele count using cutoff values defined by the observed total-risk-score values in controls (e.g., the highest decile group would include cases and controls with observed total risk scores greater than 90% of the total risk scores observed in the controls). For each decile grouping, we calculated the ORs comparing the proportions of cases and controls between the corresponding decile grouping and the lowest decile grouping (the reference group). In the other categorization scheme, we split cases and controls into two groupings defined by total-risk-allele threshold values across a range of percentiles cut-points (lower 2.5%, 5%, 10%, 25%, 50%, 75%, 90%, 95%, and 97.5%) defined by the controls. For each percentile cut-point, we compared the distributions of cases and controls between the participant groupings defined by the percentile cut-point. These contingency table-based analyses were performed using both weighted and unweighted risk-allele counts for the 23 new variants, 40 established variants, and combined set of 63 variants.

Ten out of 23 variants recently reported (20) achieved at least nominally significant evidence (one-sided P < 0.05; direction of effect consistent with prior report) for association with early-onset prostate cancer, including rs3771570 (P = 0.032), rs7611694 (P = 0.014), rs1270884 (P = 0.0028), rs8008270 (0.025), rs7241993 (0.0023), rs2405942 (P = 0.011), rs42445739 (P = 3.0 × 10⁻⁵), rs3096702 (P = 0.0018), rs2273669 (P = 8.6 × 10⁻⁴), and rs1933488 (P = 0.0011; Table 2). The latter four variants were significantly associated with prostate cancer after accounting for multiple testing (one-sided P < 0.0022; Table 2). Of the remaining 13 variants that did not minimally achieve nominal significance, only three (rs1894292, rs7141529, and rs11650494) had a direction of effect that was inconsistent with the discovery study.

Early-onset prostate cancer cases had significantly more estimated total risk alleles than unscreened controls across these 23 variants (prostate cancer cases: unweighted mean = 21.61, SE = 0.10, median = 21.70; controls: unweighted mean = 20.69, SE = 0.09, median = 20.55; P-diff = 2.0 × 10⁻¹²). Adding in the 40 established prostate cancer variants, early-onset cases carried 58.02 (SE = 0.16, median = 57.98) and controls carried 54.49 (SE = 0.15, median = 54.64) risk alleles on average (P-diff = 8.9 × 10⁻⁵⁹) across all 63 variants. Overlapping histograms plotting the distributions of the unweighted and weighted sums of risk alleles for cases and controls across the 23 new prostate cancer variants and 63 total prostate cancer variants are presented in Supplementary Figs. S1 and S2, respectively.

The aggregate burden of the risk alleles across the new variants alone provided a poor ability to discriminate between cases and controls (AUC = 0.59 for both weighted and unweighted sums; Fig. 1). The predictive value was only slightly higher when restricting the burden of risk alleles to the 10 new variants that demonstrated nominal evidence (P < 0.05) of association in our study (AUC = 0.61 for both weighted and unweighted sums). The ability to discriminate was noticeably better for the older established variants (AUC = 0.69 for weighted sums, AUC = 0.68 for unweighted sums). Adding the 23 new variants to the 40 established variants only modestly improved the ability to discriminate (AUC = 0.71 for weighted sums, AUC = 0.69 for unweighted sums) compared with the older variants by themselves.

For all three sets of variants (new, established, and combined) there was a steady increase, across decile categories, in the odds for men having prostate cancer compared with the odds for men in the lowest decile grouping (Fig. 2). For brevity, we focus here on results for the weighted total risk-allele scores [results were similar for unweighted scores (see Supplementary Fig. S3)]. A large jump in the OR was observed between the highest decile group and the next highest decile group for the set of 40 established variants [OR = 10.50 (10th decile group) vs. 4.54 (9th decile group)] and combined set of 63 variants (OR = 9.63 vs. 4.98, respectively; Fig. 2). The OR across the decile groupings for the set of new variants were considerably less striking than for the other sets of variants and there was no large jump in the last decile grouping (OR = 2.58 for the 10th decile group vs. OR = 2.49 for the 9th decile group). Categorizing subjects into two groups, based on percentile cut-points of the total-risk-allele sums in controls, revealed that the strongest OR were observed for both the upper and lower extreme 2.5% tail cut-points of the total-risk-score distribution (see Supplementary Fig. S4), consistent with the observed deficits of cases in the extreme lower tail and deficits of controls in the extreme upper tail of the total risk allele distributions.

More than 60 independent common prostate cancer variants have been discovered through GWAS in men of European ancestry. The initial discoveries, often made with relatively small case–control samples, were made possible by the relatively strong effects (OR > 1.25) of the associated variants. The more recent discoveries, including the 23 newly reported variants in Eeles and colleagues (20), required considerably larger sample sizes due to the associated variants having much smaller effects (OR ∼ 1.10). We have previously demonstrated that many of the older, stronger-effect, prostate cancer variants are individually associated with early-onset prostate cancer (21). Herein, we sought to assess whether there was evidence of association between early-onset prostate cancer and these 23 new variants. We also evaluated the added utility of including the total burden of prostate cancer risk alleles for these 23 new variants in combination with 40 previously established prostate cancer variants on early-onset disease prediction. We note that we found no evidence supporting an association between the cumulative burden of prostate cancer risk alleles and measures of disease severity including Gleason grade, tumor stage, or PSA (data not shown).

We found at least nominal evidence (one-sided P < 0.05; effect direction the same as the original study) supporting the reported associations for 10 of the 23 newly reported variants, including four that reached the conservative Bonferroni significance threshold. Ten of the 13 remaining variants had directions of effect consistent with the discovery report. Thus, despite relatively low power to detect such replication (for example, using a one-sided P = 0.05, we had power = 0.38 to detect an associated variant with minor allele frequency = 0.25 and an OR = 1.10), we were able to provide supportive evidence that many of these variants are associated with early-onset prostate cancer. A recent study that compared 312 hereditary prostate cancer cases and 620 sporadic prostate cancer cases with 587 common controls across these 23 variants found nominal evidence for association between prostate cancer and eight of the variants for hereditary prostate cancer (17/23 variants had consistent directions of effect with discovery study) and five of the variants for sporadic prostate cancer (18/23 variants had consistent directions of effect; ref. 25). No single variant achieved statistical significance after accounting for multiple testing in this study. Future larger replication studies are necessary to further validate each of these variants as a prostate cancer risk variant.

The aggregate burden of risk alleles for the 23 new variants is strongly associated with early-onset prostate cancer, but their cumulative predictive value is relatively poor. Not surprisingly, given their smaller number and smaller effect sizes, their overall predictive value is considerably smaller than was observed for the 40 established variants. Including the burden of these 23 variants to the burden of the 40 more established variants resulted in modestly stronger discrimination, with the greatest additional gains observed in men with extreme values of total risk-allele burden. These results suggest finding and including additional lower-effect common variants could be beneficial in disease prediction, but their added value will likely be small.

Three previous studies have evaluated the predictive value of the cumulative burden of established common risk alleles for prostate cancer diagnosis (25, 32, 33). The recent report by Cremers and colleagues (25) described the cumulative risk for 74 prostate cancer variants, including the same variant or a strong LD proxy for 39/40 of our established variants and all 23 new variants, separately in 312 Dutch hereditary prostate cancer cases (mean age diagnosis 62 years) and 620 sporadic prostate cancer cases (mean age diagnosis 65 years) compared with 587 common controls. Using an unweighted total risk allele score, Cremers and colleagues reported that the discriminative value based on these 74 variants was stronger for the hereditary prostate cancer cases [AUC = 0.73] than for the sporadic prostate cancer cases [AUC = 0.64]. The two earlier studies limited their analyses to established variants that demonstrated evidence for association in their own cohorts, whereas our study and the study by Cremers and colleagues included all previously reported associated variants regardless of evidence in our own studies. In Lindstrom and colleagues (32), 23/25 variants included in their risk calculations were included or had a strong LD proxy among our 40 established variants. Lindstrom and colleagues showed that the predictive value of the burden of common established risk alleles was stronger for men diagnosed with prostate cancer earlier in life [e.g., AUC = 0.66 using men diagnosed age 60 years and younger compared with AUC = 0.60 in men diagnosed after the age of 75 years]. Agalliu and colleagues (33) identified 17/31 established variants that demonstrated at least nominal evidence for significance in a cohort of 979 prostate cancer cases and 1,251 controls of Ashkenazic descent that were subsequently included in the construction of an unweighted total risk score (12/17 variants were included among our 40 established variants). The overall discriminative value of these 17 variants [AUC = 0.64] was similar to the overall value observed by Lindstrom and colleagues [average AUC = 0.63 across all ages for the 25 variants included in their study]. Consistent with Lindstrom, Agalliu and colleagues also observed a stronger association between total risk allele burden and prostate cancer in younger cases. When comparing all men in the upper 25% of the total risk allele distribution to men in the lower 25%, Agalliu and colleagues found higher OR in the younger men (diagnosed at age 60 years or younger; n = 238) with prostate cancer (OR = 5.20; 95% CI: 2.94–9.19) than in the men diagnosed with prostate cancer after age 60 years (OR = 3.30; 95% CI: 2.32–4.68).

One very interesting feature of the distribution of total risk alleles is the lack of evidence for a bi-modal distribution among cases and a noticeable deficit of cases in the lower tail of the total risk allele distribution (see Supplementary Figs. S1 and S2). The shapes of the distributions of total risk alleles in cases looked very similar to those of controls, with the distribution for the cases shifted to the right. This observation would suggest that the burden of common risk-alleles plays an important role in the probability of developing disease irrespective of other risk factors (e.g., rare variants, environmental factors, and epigenetic factors). A widely held hypothesis is that yet to be discovered uncommon high-penetrant risk alleles explain a significant proportion of the increased genetic susceptibility in prostate cancer families and men with early-onset disease. This hypothesis is supported by our recent discovery of such a mutation, G84E, in HOXB13, which has a considerably higher frequency in men with early-onset and/or familial disease (34). It has been reported that a high burden of established common variants increases disease risk even among HOXB13 G84E carriers (35). Consistent with this report, among 23 UM-PCGP prostate cancer HOXB13 G84E carriers in our study, the mean cumulative number of risk alleles across all 63 variants was 59.10 (SD = 4.88) compared with 57.99 (SD = 4.92) in non-G84E carriers.

Disease misclassification in cases and/or controls can create biased estimates of effect. We note all cases in our study were confirmed by pathology report. Our controls were largely young males (average age 20 years) who have not, to our knowledge, been screened for prostate cancer. Although approximately 15% of these men will develop prostate cancer some time in their lives, based on the age distribution of our cases and National Cancer Institute Surveillance, Epidemiology, and End Results Program prostate cancer prevalence rates (36), we would expect a disease misclassification rate of approximately 0.7% (n ∼8/1,126 of our controls) using our unscreened controls compared with a perfectly diagnosed age-matched control sample for our early-onset prostate cancer case sample. To assess the impact of this misclassification, we recalculated the mean number of total risk alleles in our iControls using this misclassification rate and the observed risk allele counts in our cases to get an unbiased maximum likelihood-based estimate (MLE) of mean total risk alleles in our control sample. Using the MLE would decrease the parameter estimates for mean number of total risk alleles from 20.69 (using the uncorrected-sample mean) to 20.68 (MLE-based mean) for the 20 new variants, 33.80 to 33.79 for the 43 established variants, and 54.49 to 54.47 for the combined set of variants. Thus, any bias using these unscreened young controls, relative to age-matched controls, is expected to be small and result in slightly conservative conclusions. Further, we note that our expected rate of disease misclassification in our controls is likely lower than that for most prostate cancer case–control studies of older men that rely on PSA and digital rectal exam (DRE) screening. There is considerable overlap in distributions of PSA for men with and without prostate cancer (37) and DRE screening misses stage T1 prostate cancer and prostate cancer that does not occur peripherally in the posterior and lateral aspects of the prostate gland.

Our study includes several other features worthy of discussion. First, the iControls do not have available prostate cancer family history and thus we cannot assess the additive value of prostate cancer risk variants in conjunction with family history. Second, cases and controls were genotyped at separate times on separate, but similar, genotyping platforms. As we reported previously looking at >450,000 genotyped variants, we saw no evidence for systematic inflation of test statistics when comparing these cases to these controls (21). Still, it is possible that a small number of individual variants could be influenced by small genotype batch effects—though the direction of those batch effects would equally likely make our results conservative or anticonservative, as the determination of the “risk allele” for each variant was based on independent data from previous reports. Third, we used imputed genotype data rather than directly genotyped data for analyses. Not all risk variants were directly genotyped and, when constructing burden scores for genotyped variants, missing data would cause unnecessary variation. We included only variants with high genotyping rates in both cases and controls in the target panel prior to genotype imputation and note that imputation quality was estimated to be excellent [R² > 0.9; see Table 1 and Lange and colleagues (21)] for the vast majority of variants. Fourth, a subset of patients (n = 127) were directly ascertained for inclusion in linkage studies based on having known living relatives with disease and many other cases were symptomatic and identified in a hospital-based setting. Thus, this collection of 931 cases is likely not representative of early-onset prostate cancer cases identified through standard epidemiologic screening studies. Fifth, using previously reported ORs from studies based primarily on older men with disease, we demonstrate a small improvement in disease prediction using weighted total risk-allele counts compared with unweighted total risk-allele counts. Prediction of early-onset disease could be improved further by applying variant weights based specifically on studies of early-onset disease. We note that using weights based on our own individual variant effect estimates in an aggregate variant burden setting would be anticonservative. Appropriate variant weighting for early-onset prostate cancer aggregate risk-allele testing will need to be continuously refined as additional prostate cancer populations are studied and new prostate cancer variants are identified.

In summary, we provide the first significant evidence to support the association of several recently identified prostate cancer variants to early-onset prostate cancer. We establish that a high-burden of common risk alleles is strongly associated with early-onset prostate cancer and that men with an aggregate burden of risk alleles in the tails of the total risk allele distribution have either high (men in the upper tail) or low (lower tail) odds of having early-onset prostate cancer. Given the strong OR observed in the upper tail, men with an unusually high number of risk alleles should be considered candidates for earlier prostate cancer screening. The ability to discriminate between case–control status was largely driven by older established variants; including the 23 new variants only modestly improved disease prediction. Despite OR that were considerably elevated there still remained considerable overlap between the case and control total risk allele distributions. Given this overlap and the apparent diminishing discriminating value of including newly discovered lower penetrant common variants, expanding the search for uncommon high-penetrant risk variants could be especially critical to further improving our ability to accurately predict men who will get early-onset disease.

No potential conflicts of interest were disclosed.

Conception and design: E.M. Lange, K.A. Cooney

Development of methodology: E.M. Lange

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K.A. Zuhlke, A.M. Johnson, S.L. Zheng, K.A. Cooney

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): E.M. Lange, J.V. Ribado, G.R. Keele, J. Li, Y. Wang, Q. Duan, Y. Li, J. Xu, K.A. Cooney

Writing, review, and/or revision of the manuscript: E.M. Lange, K.A. Zuhlke, Q. Duan, Z. Gao, J. Xu, S.L. Zheng, K.A. Cooney

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): G. Li, S.L. Zheng

Study supervision: E.M. Lange, K.A. Cooney

Other (genotyping): Z. Gao

The authors would like to thank all of the men with prostate cancer who participated in this research project. The authors especially appreciate the support of Dr. Joel Nelson and his patients. The authors also express gratitude to Ms. Linda Okoth for assisting with UM-PCGP sample preparations and clinical data collection.

This work was primarily supported by NIH R01- CA136621 (E.M. Lange, K.A. Zuhlke, A.M. Johnson, J. Li, Y. Wang, J. Xu, S.L. Zheng, and K.A. Cooney). Additional financial support provided by NIH P50-CA69568 (K.A. Zuhlke, A.M. Johnson, and K.A. Cooney), NIH R01-HG006292 (Q. Duan and Y. Li), and NIH R01-HG006703 (E.M. Lange, Y. Wang, Q. Duan, and Y. Li). J.V. Ribado was supported by the Post-Baccalaureate Research Education Program, NIH 5R25-GM089569.

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.