Reclassification of prostate cancer risk using sequentially identified SNPs: Results from the REDUCE trial

Abstract

Background

While the clinical validity of risk-associated single nucleotide polymorphisms (SNPs) for assessment of disease susceptibility has been consistently established, risk reclassification from increasing numbers of implicated risk-associated SNPs raises concern that it is premature for clinical use. Our objective is to assess the degree and impact of risk reclassification with the increasing number of SNPs.

Methods

A total of 3,239 patients from the Reduction by Dutasteride of Prostate Cancer Events (REDUCE) trial were included. Four genetic risk scores (GRSs) were calculated based on sets of sequentially discovered prostate cancer (PCa) risk-associated SNPs (17, 34, 51, and 68 SNPs).

Results

Pair-wise correlation coefficients between sets of GRSs increased as more SNPs were included in the GRS: 0.80, 0.86, and 0.95 for 17 vs. 34 SNPs, 34 vs. 51 SNPs and 51 vs. 68 SNPs, respectively. Using a GRS of 1.5 as a cutoff for higher versus lower risk, reclassification rates of PCa risk decreased: 14.11%, 12.04%, and 8.15% for 17 vs. 34 SNPs, 34 vs. 51 SNPs and 51 vs. 68 SNPs, respectively. Evolving GRSs, nevertheless, provide a tool for further refining risk assessment. When all four sequential GRSs were considered, the detection rates of PCa for men whose GRSs were consistently <1.5, reclassified, and consistently ≥1.5 were 20.8%, 29.67%, and 39.26%, respectively (P-_trend=1.12×10⁻⁸). In comparison, the detection rates of PCa in men with negative or positive family history were 23.75% and 31.78%, respectively.

Conclusions

Risk assessment using currently available SNPs is justified. Multiple GRS values from evolving sets of SNPs provide a valuable tool for better refining risk.

Background

Approximately 100 prostate cancer (PCa) risk-associated single nucleotide polymorphisms (SNPs) have been discovered from genome-wide association studies (GWAS) in recent years and the number of these PCa-risk associated SNPs continues to increase.[1–16] Although each of these SNPs is only modestly associated with PCa risk, their cumulative effects strongly increase PCa risk.[15–18] A potential clinical utility of these PCa risk-associated SNPs is risk stratification for targeted intervention such as PSA screening.[19]

With more PCa risk-associated SNPs identified, PCa risk estimated for an individual patient from these evolving sets of SNPs could be different. For a subset of individuals, their risk category (higher or lower based on a cutoff value) could be reclassified. The concern of this risk reclassification was raised from a recent simulation study by Krier et al. where the risk category was reclassified in 50% of men when using risk-associated SNPs available by 2007 and 2013.[20] This level of risk reclassification, if observed in empirical studies, would suggest that it is premature to use currently available risk-associated SNPs for risk assessment. The objective of this current study is to assess the degree and impact of this risk reclassification in an actual study population. We performed a reclassification analysis on four sets of sequentially identified PCa risk-associated SNPs in a well-defined prospective cohort of patients who were followed for four years for detection of PCa.

Methods

Study population

This study included 3,239 patients enrolled in the Reduction by Dutasteride of Prostate Cancer Events (REDUCE) clinical trial who consented for further genetic studies. This trial examined the effect of Dutasteride on PCa development, and has been previously described in detail.[21] Briefly, eligible men (1) had a serum PSA between 2.5–10.0 ng/mL (50–60 years of age) or 3.0–10.0 ng/mL (61–75 years of age), and (2) had undergone one 6–12 core biopsy within 6 months of enrollment and were not diagnosed with PCa. Participants in this trial were randomized to receive Dutasteride (treatment arm) or to receive a placebo (placebo arm). We utilized the placebo arm (n=1,654) for the primary analyses, due to the effects of Dutasteride on decreasing PCa incidence. The treatment arm (n=1,585 Caucasians) was used as an independent replication cohort. Family history (FH) information was collected during the enrollment. Patients with any first degree relatives who had PCa were considered as FH positive. All patients provided informed consent, and the Institutional Review Boards at all participating institutions granted approval.

Genotyping and Quality Control

DNA samples were genotyped in the Center for Cancer Genomics at Wake Forest University using the Illumina HumanOmniExpress BeadChip, which included 729,755 SNPs. For PCa risk-associated SNPs that were not included in the GWAS array, imputation was performed using IMPUTE 2.2.2 based on the combined data of the 1000 Genomes project and HapMap3 data.[22] A posterior probability of >0.9 was applied to call imputed genotypes. More detailed description of genotyping and quality control procedures were described elsewhere.[18]

Assessment of Genetic Risk

SNP Selection

In this study, we applied stringent criteria to ensure that SNPs used in the analysis are common, independent and validated PCa risk-associated SNPs. The criteria were: 1) discovered from GWAS studies with at least 1,000 cases and 1,000 controls; 2) met the gold standard GWAS significance level of P<5×10⁻⁸; 3) SNPs with minor allele frequency>0.05 in reported studies; and 4) independent, linkage disequilibrium measurement (r²<0.2) between any pair of SNPs. As a result, 68 PCa risk-associated SNPs were selected for analysis in the study. SNPs were ordered based on the time that they were identified, and were evenly divided into 17, 34, 51 and 68, representing four sequential sets of PCa risk-associated SNPs (Supplementary Table 1).

Methods for measuring cumulative effect of SNPs

The primary method for measuring the cumulative effect of SNPs in the study was genetic risk score (GRS). It was calculated for each subject based on the genotypes of the various SNPs and weighted by their ORs and risk allele frequencies.[18,27] GRS was calculated as $\text{GRS}={\displaystyle \prod _{i=1}^n}\frac{{{OR}_i}^{g_i}}{W_i}$, where g_i is the genotype of SNP i for an individual (0, 1, or 2 for individual with homozygous of non-risk allele, heterozygous or homozygous of risk allele, respectively). OR_i is the OR of SNP i estimated from external studies[1–16], W_i is the average population risk of SNP i, calculated as W_i = f_i²OR_i² + 2f_i(1−f_i)OR_i + (1−f_i)², where f_i is the risk allele frequency of SNP I based on the 1000 Genome Project of the CEU population. Therefore, a GRS value of 1.0 represents a population average risk. Two other methods for measuring the cumulative effect of SNPs were also used, including a simple risk allele count ( $\text{sRAC}={\displaystyle \sum _{i=1}^n}{g}_i$) and a polygenic risk score (PRS), an OR-weighted risk allele count ( $\text{wRAC}={\displaystyle \sum _{i=1}^n}{g}_{i}ln{OR}_i$).

Statistical analysis

In this study, a GRS cutoff value of 1.5 was used for risk stratification; men with GRS <1.5 or ≥1.5 were classified as lower or higher risk, respectively. This cutoff value was chosen because it confers a risk of PCa similar to that of having positive FH (odds ratio of FH for PCa diagnosis was 1.5 in the placebo arm of the REDUCE study). The Pearson correlation coefficient was used to evaluate the linear correlation between each pair of GRS values. T-tests were used to test the differences in means of normally distributed variables between two groups. For variables that were not normally distributed, two tests were performed; (1) a nonparametric method using the Wilcoxon rank sum test, and (2) t-tests for different means between two groups after log-transformation. Differences in binary variables were tested using chi-square tests. Area under the receiver operating characteristic curve (AUC) was used to evaluate the performance in discriminating two groups of subjects. The difference between two AUCs was tested using Delong’s test.[23] Cochran-Armitage trend test was used to test the difference of detection rates across risk categories.

Results

The baseline demographic and clinical characteristics of men included in this study, stratified by placebo and treatment arms, are shown (Table 1). Univariate analyses demonstrated that men with PCa differed significantly (P<0.05) from men without PCa for all baseline clinical and demographic variables, with the exception of digital rectal examination. Men diagnosed with PCa had a significantly higher proportion of positive FH of PCa in the placebo arm (17% vs. 12%, P=0.01), but not in the treatment arm (16% vs. 14%, P=0.36).

Table 1.

Baseline clinical and demographic variables of the subjects in the REDUCE study

Variables Baseline clinical variables	Placebo group			Drug group
	PCa (N=410)	Non-PCa (N=1244)	P	PCa(N=304)	Non-PCa(N=1281)	P
Age
Mean(SD), year	63.52(5.99)	62.22(6.01)	0.0001	63.88(5.87)	62.78(5.97)	0.0037
Median(range), year	63(50–76)	62(49–76)		64(51–75)	63(49–75)
No.(%) with positive DRE	20(5)	47(4)	0.33	12(4)	55(4)	0.78
Prostate volume
Mean(SD), mL*	44.2(21.4)	46.76(16.13)	0.03	40.27(1.47)	44.93(1.44)	3.63×10−6
Median(range), mL	41.61(9.01–257)	45.46(3.66–127)		40.69(12.46–108.4)	45.62(11.97–257.17)
Total PSA levels
Mean(SD), ng/mL*	5.78(1.37)	5.52(1.4)	0.01	5.77(1.4)	5.55(1.39)	0.058
Median(range), ng/mL	5.7(2.5–10.2)	5.7(1.8–14.2)		5.8(2–15.6)	5.6(1.9–11.1)
PSA density(PSA/PV)
Mean(SD), ng/mL2*	0.14(1.67)	0.14(0.07)	4.67×10−6	0.14(1.62)	0.12(1.55)	1.10×10−6
Median(range), ng/mL2	0.14(0.02–0.63)	0.12(0.03–0.58)		0.14(0.04–0.56)	0.12(0.02–0.61)
Free-to-total PSA ratio
Mean(SD)	0.16(0.06)	0.17(0.06)	0.02	1.17(1.06)	1.19(1.06)	7.73×10−6
Median(range)	0.16(0.03–0.47)	0.16(0.03–0.47)		1.16(1.04–1.56)	1.18(1.02–1.86)
No. of cores sampled at baseline biopsy
Mean(SD)	8.21(2.27)	8.58(2.39)	0.004	8.08(2.19)	8.65(2.40)	0.0002
Median(range)	8(3–19)	8(2–22)		8(4–16)	8(2–20)
No.(%) with positive family history	68(17)	146(12)	0.01	48(16)	176(14)	0.36

^*P-value were analyzed after taking natural logarithm of the variables.

DRE: digital rectal exam; PCa: prostate cancer; PSA: prostate-specific antigen; PV: prostate volume; SD: standard deviation

Although not every individual SNP was significantly associated with PCa risk (data not shown), combinations of these SNPs, measured by GRS, were strongly associated with PCa susceptibility. In the placebo group, the associations between GRS values calculated from each of the four SNP sets and the detection of PCa were highly significant (all P values <10⁻⁷) (Table 2). The performance (measured by AUC) of these increasing numbers of SNPs in discriminating biopsy outcomes (PCa from non-PCa) increased from 0.58 to 0.61, 0.60, and 0.60 for the 17, 34, 51 and 68 SNPs, respectively. A plateau effect of increasing AUC with the increased number of SNPs was observed with the 51 SNP group. The AUC values of all SNP sets were significantly higher than that of FH (AUC=0.52), all P values <0.05. Similar findings were observed for the treatment group (Supplementary Table 2).

Table 2.

Predictive performance of FH and GRS for prostate cancer in the placebo arm.

Risk factors	OR	P-value	AUC	P-value for comparison
				FH	17 SNPs	34 SNPs	51 SNPs	68 SNPs
FH	1.50	0.012	0.52	-	0.002	3.30×10−5	7.32×10−5	4.06×10−5
17 SNPs	1.70	8.77×10−7	0.58	-	-	0.051	0.21	0.17
34 SNPs	1.76	5.37×10−10	0.61	-	-	-	0.52	0.77
51 SNPs	1.57	3.16×10−8	0.60	-	-	-	-	0.68
68 SNPs	1.54	2.18×10−8	0.60	-	-	-	-	-

AUC: area under the curve; FH: family history; GRS: genetic risk score; OR: odds ratio; SNP: single nucleotide polymorphism;

On a per individual subject level, variability of GRS values calculated from 17, 34, 51, and 68 SNPs was documented, but these GRS values were highly correlated. Scatter-plots of GRS values between two sequential SNP sets are shown in Figure 1 for the placebo arm. The shape of scatter-plots became tighter with an increasing number of risk-associated SNPs: 17 vs. 34 SNPs (a), 34 vs. 51 SNPs (b), and 51 vs. 68 SNPs (c). Correspondingly, the pair-wise correlation coefficient (r) of GRS values increased from 0.80 to 0.86 to 0.95 for 17 vs. 34 SNPs, 34 vs. 51 SNPs and 51 vs. 68 SNPs, respectively. Using GRS values <1.5 and ≥1.5 to define lower and higher risk, respectively, reclassification of PCa risk was observed. However, the rates of reclassification decreased with each sequential SNP set; 14.11%, 12.04%, and 8.15% for 17 vs. 34 SNPs, 34 vs. 51 SNPs and 51 vs. 68 SNPs, respectively (Figure 1 and Table 3). Most risk reclassification occurred in subjects whose GRS values were near the 1.5 cutoff value; 81.03%, 80.81%, and 92.54% of reclassified subjects were between 1.0 and 2.0 for 17 vs. 34 SNPs, 34 vs. 51 SNPs and 51 vs. 68 SNPs, respectively. Similar findings were observed for the treatment arm (Supplementary Figure 1 and Supplementary Table 3).

Figure 1.

Scatter plots of GRS values from two consecutive SNP sets in the placebo arm: (a) 17 vs. 34 SNPs, (b) 34 vs. 51 SNPs, and (c) 51 SNPs vs. 68 SNPs. Blue and red circle indicates patients with or without a diagnosis of PCa, respectively. Two dotted lines indicate cutoff value (1.5) to classify higher or lower risk.

Table 3.

Risk categories based on evolving SNP sets and PCa detection rate in the placebo arm using GRS cutoff value of 1.5.

Set of SNPs	No. (%) of subjects			% of PCa				% of high-grade PCa
	Remain lower	Reclassified	Remain higher	Remain lower	Reclassified	Remain higher	P-trend	Remain lower	Reclassified	Remain higher	P-trend
	17 vs. 34 SNPs	1168 (71.05%)	232(14.11%)	244 (14.84%)	20.98	34.91	34.43	7.80×10−8	5.91	7.76	8.61	0.086
34 vs. 51 SNPs	1172 (71.29%)	198 (12.04%)	274 (16.67%)	21.42	27.27	38.32	5.58×10−9	5.63	8.59	9.12	0.019
51 vs. 68 SNPs	1213 (73.78%)	134 (8.15%)	297 (18.07%)	21.93	27.61	36.03	3.84×10−7	5.61	10.45	8.75	0.020
17, 34, 51, and 68 SNPsa	1053 (64.05%)	428 (26.03%)	163 (9.91%)	20.8	29.67	39.26	1.12×10−8	5.60	7.71	9.82	0.021

PCa: prostate cancer; SNP: single nucleotide polymorphism.

^aReclassification of patient risk based on any GRS. For example, 1053 (64.05%) patients had lower risk for all of the four GRSs, 428 (26.03%) had reclassification for any of the GRSs, and 163 (9.91%) had higher risk for all of the four GRSs.

Multiple GRS values from evolving sets of SNPs, nevertheless, offered a tool for further refining risk prediction. In the placebo arm, when all four sequential GRS values were considered, risk reclassification occurred in 26% of men. The observed detection rate of PCa in the 4-year study period were 20.80%, 29.67%, and 39.26% for men whose GRS values were consistently <1.5, changed between <1.5 and ≥1.5 (reclassified), and were consistently ≥1.5, respectively, P_trend=1.12×10⁻⁸ (Table 3). In comparison, the detection rates of PCa in men with negative or positive FH were 23.75% and 31.78%, respectively. Comparable findings were observed for the treatment arm (Supplementary Figure 1 and Supplementary Table 3).

The performance of multiple GRS values from sequentially discovered SNP sets in refining risk assessment was also supported by the observed detection rate of high-grade PCa (Gleason score ≥7) in the 4-year study period. For example, in the placebo arm, the detection rate of high-grade PCa were 5.60%, 7.71%, and 9.82% for men whose GRS values were consistently <1.5, reclassified (changed between 1.5 and ≥1.5), and were consistently ≥1.5, respectively, P_trend=0.02 (Table 3). In comparison, the observed detection rate of high-grade PCa was 6.04% and 9.81% for men with a negative or positive FH, respectively. Again, similar findings were observed for the treatment arm (Supplementary Table 3).

Parallel results were found using simple risk allele count (sRAC) and PRS for measuring the cumulative effect of multiple risk-associated SNPs. Mean PRS and sRAC of each SNP set in PCa patients were significantly higher than that of non-PCa patients for each set of SNPs (all P values <0.05). For discriminating cases from controls at a population level, the performance (AUC) of PRS and GRS were the same and were both better than sRAC for each set of SNPs (Supplementary Table 4). At an individual level, scores of two sequential SNP sets for both sRAC and PRS were highly correlated and the scatter plots became tighter with evolving SNP sets (Supplementary Figures 2 and 3).

Discussion

Utilizing data from the prospective REDUCE study where all subjects were monitored for the development of PCa, we assessed the degree and impact of risk reclassification from increasing numbers of risk-associated SNPs on PCa diagnosis. We found that GRS values calculated from each of the four sequential sets of PCa risk-associated SNPs (17, 34, 51, and 68 SNPs) were significantly associated with PCa risk and had a better performance in discriminating PCa from non-PCa than FH. Although there was variability in GRS values for individual patients when using these four sequential sets of risk-associated SNPs, they were highly correlated, and their correlations increased with evolving SNP sets.

Using GRS values ≥1.5 to define higher PCa risk, reclassification of risk categories among different SNP sets was observed. However, the reclassification rate continued to decrease with evolving SNP sets and was only 8.15% between the two latest SNP sets (51 vs. 68 SNPs). More importantly, multiple GRS values from evolving SNP sets actually provide a valuable tool for refining risk for all subjects. Risk reclassification effectively captures men with GRS values in a grey zone (near 1.5) that are at intermediate risk, and men who have consistently lower (<1.5) or higher (≥1.5) GRS values from multiple SNP sets are further assured of their low or high genetic risk, respectively. Taken together, the present study suggests that 1) on the basis of comparative effectiveness principle, GRS from currently available risk-associated SNPs should be used to stratify PCa risk, and 2) newly discovered SNPs should be used to calculate new GRS values where a combination of new and previous GRS values should be considered together to further refine risk.

Although there were similarities between the results of this study and Krier et al., several differences were noted.[20] The reclassification rate of PCa risk was higher in the study of Krier et al. (50%) than ours (26%).[20] This difference is likely due to a combination of factors, including differences in the SNP sets for risk assessment, estimates of ORs and allele frequencies of SNPs used in calculations, and cutoff values (2.0 vs. 1.5). A GRS cutoff value of 1.5 was chosen in our study because it confers a risk of PCa similar to that of having positive FH (odds ratio of FH for PCa diagnosis was 1.5 in the placebo arm of the REDUCE study). However, similar findings were observed when a GRS cutoff value of 2.0 was implemented (Supplementary Table 5). The most important difference between the two studies is the availability of follow-up data on observed PCa detection. With the prospective design of our study, we are able to evaluate the impact of risk reclassification and demonstrate that multiple SNP sets do not pose a challenge, but rather provide an effective tool for further refining risk.

Most of the analyses described in this study were performed separately in the two trial arms. The decision to perform these analyses separately was based on the consideration that PCa detection rate was significantly lower in patients treated with dutasteride.[8] Despite these differences, the key findings from the treatment group were consistent with that of the placebo group.

Several methods are commonly used to measure the cumulative effect of multiple risk-associated SNPs, including sRAC, PRS, and GRS.[18,24–26] Although the same conclusion of the study can be made from any of these methods, there are important differences between them. sRAC simply counts the number of risk allele without considering OR of each SNP. In PRS, the risk allele count is weighted by the OR of each SNP. For GRS, the risk allele count is first weighted by the OR of each SNP and then standardized by population average risk. At a population level, all three methods perform similarly in predicting PCa risk; the score of sRAC, PRS, and GRS are significantly higher in cases than in controls. PRS and GRS have the same performance in discriminating cases and controls and both have a better performance than sRAC.[27] At an individual level, scores from GRS are easiest to interpret because values higher or lower than 1.0 indicate higher or lower risk than population average, respectively, regardless of the number of SNPs used for the calculation. Scores from PRS and sRAC are difficult for risk assessment because different values from different SNP sets are needed to indicate higher or lower risk. This limitation of PRS and sRAC also makes it difficult to compare multiple risk scores of the same individual from evolving SNP sets. Therefore, GRS is a better choice for risk assessment at an individual level risk assessment.

Important limitations of the study should be noted. First, considering that all subjects in the REDUCE study were men with initially negative biopsies, caution should be made when we attempt to generalize the current results to the general population. Second, the study was restricted to Caucasian men due to the composition of the REDUCE cohort and the fact that most PCa risk-associated SNPs were discovered and confirmed in this racial group. We hypothesize that major findings derived from Caucasians in the study can be replicated in other racial groups, though this needs to be further tested. Third, using 1.5 (or 2.0) as cutoff value of GRS might be subjective. As GRS provides population standardized risk assessment, GRS=1.5 (or 2.0) represents a 1.5-fold (or 2-fold) increase of risk while GRS=0.5 represents a 0.5-fold decrease of risk, comparing to average risk of PCa in general population. [27] These cutoffs were easier to understand and were comparable to FH. On the other hand, our results were consistent at both cutoffs (1.5 and 2.0), suggesting that our conclusions could be applied to a wider range. Fourth, the GRSs did not include rare mutations in high-penetrance genes. DNA sequencing analysis is required to detect these mutations, but was not performed in the REDUCE study. Future studies on assessment of inherited risk should include common PCa risk-associated SNPs, high-penetrance genes associated with PCa risk, as well as FH.

Conclusions

Risk assessment using currently available SNPs is justified when compared with FH. Multiple GRS values from evolving sets of SNPs provide a valuable tool for better refining risk.

Supplementary Material

Supp Fig S1. Supplementary Figure 1.

Scatter plots of GRS values from two consecutive SNP sets in the treatment arm: (a) 17 vs. 34 SNPs, (b) 34 vs. 51 SNPs, and (c) 51 SNPs vs. 68 SNPs. Blue and red circle indicates patients with or without a diagnosis of PCa, respectively. Two dotted lines indicate cutoff value (1.5) to classify higher or lower risk.

Supp FigS2. Supplementary Figure 2.

Scatter plots of sRAC values from two consecutive SNP sets in the placebo arm: (a) 17 vs. 34 SNPs, (b) 34 vs. 51 SNPs, and (c) 51 SNPs vs. 68 SNPs. Blue and red circle indicates patients with or without a diagnosis of PCa, respectively.

Supp FigS3. Supplementary Figure 3.

Scatter plots of PRS values from two consecutive SNP sets in the placebo arm: (a) 17 vs. 34 SNPs, (b) 34 vs. 51 SNPs, and (c) 51 SNPs vs. 68 SNPs. Blue and red circle indicates patients with or without a diagnosis of PCa, respectively.

List of abbreviations

AUC

Area under the receiver operating characteristic curve

FH

family history

GRS

genetic risk scores

GWAS

genome-wide association studies

PCa

prostate cancer

PRS

polygenic risk score

REDUCE

the Reduction by Dutasteride of Prostate Cancer Events trial

SNP

single nucleotide polymorphisms

sRAC

simple risk allele count