NCBI Bookshelf. A service of the National Library of Medicine, National Institutes of Health.

Little J, Wilson B, Carter R, et al. Multigene Panels in Prostate Cancer Risk Assessment. Rockville (MD): Agency for Healthcare Research and Quality (US); 2012 Jul. (Evidence Reports/Technology Assessments, No. 209.)

Cover of Multigene Panels in Prostate Cancer Risk Assessment

Multigene Panels in Prostate Cancer Risk Assessment.

Show details

Results

The literature search yielded 1,998 unique citations. In total, 1,303 (65 percent) were excluded from further review following the initial level of title and abstract screening. Because of the complexity of the content area, and challenges in defining the ‘clinical relevance’ of the reported evaluations, full text screening was conducted in three phases. The first phase was conducted by EPC staff and focused on the most straightforward assessment of the overall study against eligibility criteria; the second phase was conducted by investigators and focused on establishing the eligibility of the specific SNPs within the panel reported; the third phase was also conducted by the investigators and focused on deciding whether the SNP panel could be considered ‘available’ and whether the evaluation context could be considered, at least to some extent, clinically relevant. Therefore, out of the 695 citations promoted to full text screening, 457 were excluded at the first phase, 127 were excluded at the second phase, and 97 at the third phase. This left 14 articles 188-201 retained for the review, which proceeded to data abstraction and quality assessment. All 14 focused on the assessment of clinical validity (KQ2). Figure 2 depicts the flow of studies through the screening process, and reasons for study exclusion. The remainder of this chapter describes the evidence for the key questions (KQs) and a quality assessment of the studies.

Figure 2 is a flow diagram indicating the numbers of citations included and excluded at each level of screening. In addition the reasons for exclusion at full text screening and sorting are included. The literature search yielded 1,998 unique citations. In total, 1,303 citations were excluded from further review following the initial level of title and abstract screening. Out of the 695 citations promoted to the initial level of full text screening, 454 were excluded and 238 proceeded to full text sorting. Of the 238 articles sorted, 127 were excluded. The remaining 111 citations were subject to full text panel screening where 97 were excluded. This left 14 articles that passed full text screening and proceeded to data extraction and quality assessment.

Figure 2

Flow diagram depicting the flow of studies through the screening process.

One challenge that became evident during the assembly of source material for review was a lack of published data describing the technical protocols and analytical accuracies achieved for specific SNPs, and in particular, their analytical validation. There was also a paucity of information describing the laboratory protocols used to demonstrate the analytical validation of SNP panels used for clinical service testing. The reviewers sought but did not receive additional unpublished details about the analytical and clinical validation of proprietary commercial panels from the providers of these services. Therefore, from the articles eligible for KQ2 (clinical validity), we abstracted any information that was relevant to KQ1 (analytic validity).

Characteristics of the Studies

All but two of the studies were of case-control design with the number of cases ranging from 203 to 2,899 and the number of controls from 560 to 1,781 (Tables 8 through 10). One study was a cross-sectional study of 5,241 men who had undergone prostate biopsy,200 and one was an investigation of survival in 2,875 men diagnosed with prostate cancer.201 The studies were carried out in Canada,195 Sweden,188,192,198,200,201 the United States,189,191,194,196,197,199 (clarified in an email from W. Catalona, M.D. ( gro.ffmn@anolataCW) in February 2012) and in both Sweden and the United States.190,193

Table 8. Characteristics of included studies.

Table 8

Characteristics of included studies.

Table 9. Characteristics of included studies: SNPs.

Table 9

Characteristics of included studies: SNPs.

Table 10. Characteristics of included studies: Analysis and results.

Table 10

Characteristics of included studies: Analysis and results.

There was complete overlap in the participants included from five of the six studies that included Sweden: a risk model was initially developed for a panel of 5 SNPs,188 extended to 11 SNPs192 in data from the same participants, then 14 SNPs,193 and then 28 SNPs;198 the study of prostate cancer survival used a 16-SNP panel.201 For the initial 5-SNP model, validation was undertaken in King County (Washington, United States),189 and a combined estimate of the cumulative effect of the five risk variants was made, which incorporated these data and the Swedish data.190 For the 14-SNP model, data from the United States were used for confirmation;193 the U.S. data in this study was based on the same participants (in the Prostate, Lung, Colorectal and Ovarian (PLCO) Cancer Trial)90 as in one of the U.S. studies used to validate the 5-SNP model.190 There was also overlap between the studies in the United States, first of participants recruited at the Johns Hopkins Hospital, Baltimore 1999-2006,190,194 second in participants recruited in King County, Washington 1993-2002 to 2002-2005,189,197 and third in participants recruited in Chicago 2002-2008191 and 1997-2009.199

Nine of the studies were concerned solely with the development of models for the prediction of risk for prostate cancer,188,192,194,196-201 two solely with model validation,189,190 and three with both development of new models and validation of previously-developed models.191,193,195 All five of the studies that carried out model validation used data independent of those in which the models had been developed. However, in two of the studies, the teams of investigators validating the models included some who were also involved in the model development.190,193

Most of the studies related to participants of European origin. In all but one200 of the studies of Swedish participants,188,190,192,193,198,201 the ethnicity was not explicitly specified. All but one of the studies of United States' participants were limited to men of European origin.189-191,193,194,197,199 The exception to this presented a stratified analysis for non-Hispanic European (54 percent of controls), Hispanic (33 percent), and African-American origin (13 percent).196 The study including Canadian subjects also related to ethnically diverse participants: European origin (81 percent of controls), Asian (8 percent), black (7 percent), and other (4 percent); some analyses were adjusted for ethnicity and some were restricted to participants of European origin.195

In one study, estimates were presented separately for cases from families in which two additional first-degree relatives had been diagnosed with prostate cancer and for cases that were recruited irrespective of family history.194

Eight studies presented information on the proportion of cases and controls with a family history of prostate cancer. In five, this was specified as relating to first-degree relatives – in three different analyses of the same Swedish participants, the proportion of cases with a family history was 19 percent and controls 9.4 percent,188,192,198 in a study in King County, WA, the proportions were 21.6 percent and 11.1 percent, respectively,189 and in the study in which cases were recruited in Chicago, IL, and St. Louis, MO, the proportions were 36.4 percent and 14.9 percent.199 In one study, family history referred to first- and second-degree relatives, and the proportion of cases for which such a history was reported was 11.6 percent and of controls 6.1 percent.193 In the other two studies, the degree of relationships included in “family history” were not defined: in the Canadian study, the proportion for cases was 16.4 percent and for controls was 12.1 percent,195 while in the Stockholm study, the proportions were 29.0 percent and 21.9 percent respectively.200

Ten articles were based on newly incident cases, one that related to the Canadian study (cases detected following referral for prostate-specific antigen (PSA) ≥4.0ng/mL or abnormal digital rectal examination without previous history of prostate cancer),195 six to data on the same participants from Sweden,188,190,192,193,198,201 one to the Stockholm study,200 and two to partially overlapping studies from the United States.189,197

Two publications (one of which also reported on participants from Sweden) reported analyses on prevalent cases from overlapping studies in the United States.190,194 One study in the United States was based on a mixture of newly incident and prevalent cases.196 In another two, it was unclear whether the cases were newly incident or prevalent – it was stated only that the cases were recruited after radical prostatectomy.191,199

The mean age of cases ranged from 56.8 years193 to 70.5 years.197 There was no obvious pattern according to inclusion of newly incident or prevalent cases.

As might be expected given trends in PSA testing, there appeared to be a pattern that the average PSA level at diagnosis of cases was lower for more recent study periods. The proportion of cases with a PSA level of ≤4ng/ml varied between under 8 percent in Canada 1999-2007195 and Sweden 2001-2003,188,192,198 to 13.6 percent in Washington State (United States) 1993-1996 and 2002-2005,202 and 22 percent in Chicago 2002-2008.191

Where reported (n=9), the proportion of cases with a Gleason score of ≤6 at diagnosis ranged from 51 percent (Physicians' Health Study) 1982-2008197 to 81 percent (Chicago and St. Louis 1997-2009).199 Only one study190 explicitly referred to having used the revised scoring as described by Epstein, et al.,62 for the Johns Hopkins Hospital component of the study. The stage at diagnosis was reported for the Swedish cases,188,192,198,200,201 in the study comprising three sets of cases and controls in the United States,197 and the Chicago study;191 over two-thirds of the cases were stage T2 or less at diagnosis. All of the cases in the Chicago-St. Louis study were stage T1c at diagnosis.199

In some of the studies, cases and controls clearly derived from the same study base. Thus, in the Canadian study, controls were selected from the same group of men referred to the prostate cancer centers of the University of Toronto who had either a PSA value ≥4.0ng/ml or an abnormal digital rectal examination (DRE), and who had no biopsy evidence of prostate cancer.195 In five of the studies including Swedish cases, the controls were population-based and selected from the Swedish population registry.188,190,192,193,198 In the Stockholm study, participants had undergone at least one prostate biopsy.200 The cases from the PLCO Trial were compared with controls participating in the trial.193,203 Cases arising in the Physicians' Health Study197 and cases from the San Antonio cohort196 were compared with controls selected from the same cohorts. Cases with prostate cancer in King County, Washington were compared with men without a self-reported history of prostate cancer who were resident in the county and identified by random digit dialing (participation rate 44.5 to 51.6 percent).189,197 Cases from the Johns Hopkins Hospital series, all of whom had undergone radical prostatectomy, were compared with men undergoing surgery for prostate cancer at the Johns Hopkins Hospital and in the greater Baltimore metropolitan area who had normal DRE, PSA <4.0ng/ml, and were aged >55 years.190,194 Cases for the Northwestern Memorial Hospital series, all of whom had undergone radical prostatectomy, were compared with 777 healthy male volunteer controls; from these, 247 may have been selected for the Icelandic genealogical database or from other genome-wide association (GWA) studies at deCODE, while the remaining participants were from a prostate cancer screening program done in April 2007 (it is not stated where this occurred).191,204 In the Chicago-St. Louis study,199 203 stage T1c cases (who had undergone radical prostatectomy, had a PSA <4.0ng/ml and a nonsuspicious DRE) were compared with 611 controls who had a PSA <4.0ng/ml, normal DRE, and no prior history of prostate biopsy that are stated to have been selected from a GWA study that included participants from the University of Chicago and Northwestern,205 per an email from W. Catalona. M.D.( gro.ffmn@anolataCW) on February 2, 2012.

Source of Funding and Conflict of Interest

All of the studies were publicly funded. In addition, two studies received support from deCODE Genetics.191,199 All but five studies189,190,194,198,199 included conflict of interest statements. Of the nine studies in which there was such a statement, two referred to the filing of a patent application188,192 and two indicated specific nonpublic funding received by one of the authors.191,197

Overview of the SNP-Based Genotype Panels

There were 15 panels identified from the included studies (Tables 11 and 12). The number of SNPs included in the panels ranged from two to 35. Almost all of the individual SNPs had been discovered and replicated as being associated with prostate cancer in GWA studies.

Table 11. Focus 5 test.

Table 11

Focus 5 test.

Table 12. Summary of SNPs and other variables included in test panels.

Table 12

Summary of SNPs and other variables included in test panels.

Apart from overlap for the five SNPs included in the Focus 5 test panel, there were considerable differences between the panels assessed (Table 12).

The first test panel included five SNPs as described in the article of Zheng, et al.,188 and is the basis of the Focus 5 predictive test for prostate cancer. A patent application has been filed by Xu, et al.,206 “Methods and compositions for correlating genetic markers with prostate cancer risk.” The test has been marketed by Proactive Genomics.207 Four other articles assessed this test in independent data.189-191,195

The second test, again initially proposed by Zheng, et al.,188 included family history with the five SNPs included in the first test, and two of the articles that assessed the first test panel also assessed this test.189,190 In two of these studies, family history was defined to include first degree relatives.188,189

The other 13 tests were reported in 11 articles191-201 (Table 11). Four of these included family history, two in first-degree relatives,192,198 one in first- and second-degree relatives,193 and one in relatives of unspecified degree.200

deCODE markets the deCODE ProstateCancer test, which tests for 27 genetic variants associated with prostate cancer in men of European descent (including the five SNPs included in the Focus 5 test), a subset of 9 variants for African-American men, and a subset of 12 variants for men of East Asian descent (Table 13); the specific variants in the subsets are not specified in the Web site (www.decodhealth.com/prostate-cancer).208 If the deCODE ProstateCancer is sought separately, it has to be obtained through a licensed health professional. The test can also be ordered as part of the deCODEme Complete Scan, which analyzes genetic risk factors for 47 traits and conditions ($1,100 USD as of 19 June 2011) or the deCODEme Cancer Scan, which analyzes genetic risk factors for seven types of cancer ($500 USD).209 A patent application was filed by Gudmundsson, et al., in May, 2010.210

Table 13. Genetic variants tested for by deCODE ProstateCancer.

Table 13

Genetic variants tested for by deCODE ProstateCancer.

KQ1. What is the analytic validity of available SNP-based panels designed for prostate cancer risk assessment?

1. What is the accuracy of assay results for individual SNPs in current test panels?

No data addressing this question were identified in the literature search. Companies known to offer testing for the risk for prostate cancer based on SNP panels were approached in May 2011, as were companies known to offer genetic testing more generally. As of September 1, 2011, no response had been received. From the articles that were identified as providing information relevant to the assessment of the clinical validity of SNP panels (KQ2), no data were presented on the analytic validity of individual SNPs from which the panels were composed.

2. What is the analytic validity of current test platforms whose purpose is, or includes, predicting risk of prostate cancer?

5-SNP panel. The 5-SNP panel that is the basis of the Focus 5 test, and the test that incorporates family history of prostate cancer, was genotyped using the Mass ARRAY QGE iPLEX system (Sequenom) in the report in which these models were developed.188 The same method was applied in samples from the Johns Hopkins Hospital190 and Canada.195 Some of the analytic validity information relevant to the initial study in Swedish samples188 are reported in other articles which relate to the same platform, including the initial five SNPs as well as additional SNPs.192,193,201 A call rate of 98.3 percent was reported,192,193,201 with a concordance rate for duplicate SNPs of >99 percent, and the genotypes for each SNP conformed to Hardy-Weinberg equilibrium (HWE) in controls.188,192,193 (For the purpose of this report, call rate was defined as the proportion of samples for which genotypes are called for a converted marker). It was not reported whether genotyping was done blind to case-control status.

The 5-SNP panel was genotyped with one modification (substitution of rs6983561 for rs16901979; it was stated that there was perfect correlation between these two SNPs in HapMap CEPH individuals), in a study using the Applied Biosytems (ABI) SNPlex Genotyping System.189 There was perfect agreement for the five SNPs between 140 blind duplicate samples distributed across all genotyping batches. Genotyping was done blind to case-control status. All genotype frequencies observed in controls were consistent with HWE.

One of the sets of samples used to assess the 5-SNP panel was the PLCO trial.190 Four of the SNPs had already been genotyped as part of a GWA.159 The genotyping had been undertaken by means of Sentrix HumanHap300 and Sentrix HumanHap240 platforms (Illumina).158,161 The fifth SNP (rs16901979 in 8q24) was imputed from the adjacent genotyped SNPs at 8q24.190

9-SNP panel. In the study of Helfand, et al.,191 it is stated that genotyping was done by deCODE and reference is given to previous papers describing genotyping methods, quality control, and genotyping accuracy (5 companion papers).159,160,162,165,205 The methods include the Illumina Infinium Human Hap300 SNP chip, for which it is stated that samples with a call rate of <98 percent were excluded from analysis.159,160,162,165 In addition, the Centaurus (Nanogen) platform was used159,160,162,165,205 and the concordance rate of SNPs genotyped by both the Illumina and Centaurus methods was stated to be >99.5 percent.159,160 It is also stated that all genetic variants were in HWE.191

17-SNP panel. In the Chicago-St. Louis study,199 as for the 9-SNP panel, it is stated that genotyping was also done by deCODE and reference is given to the same companion papers describing genotyping methods, quality control, and genotyping accuracy.159,160,162,165,205 It is also stated that all 17 genetic variants were in HWE in controls.199

11-SNP panel. This panel was genotyped using the Mass ARRAY QGE iPLEX system (Sequenom).192 A call rate of 98.3 percent was also reported, with an average concordance rate for duplicate SNPs of 99.8 percent, and the genotypes for each SNP conformed to HWE in controls.192 It was not reported whether genotyping was done blind to case-control status.

14-SNP panel. In the Swedish samples in this study, this panel was genotyped using the Mass ARRAY QGE iPLEX system (Sequenom).193 A call rate of 98.3 percent and a concordance rate between duplicate samples included in each-96-well plate of 99.8 percent was reported. For the samples from the PLCO Trial included in this study, it is stated that 13 SNPs had been genotyped already as part of a companion paper,161 and one (rs16901979 in 8q24) was imputed. In the PLCO samples, genotyping was undertaken by means of Sentrix® HumanHap300 and Sentrix HumanHap240 platforms (Illumina).158,161 It is stated that tests for HWE in control participants in each of the two sets of samples were made, but results are not presented. It was not reported whether genotyping was done blind to case-control status.

16-SNP panel. This panel was genotyped using the Mass ARRAY QGE iPLEX system (Sequenom).201 A call rate of 98.3 percent was reported, with an average concordance rate for duplicate SNPs of 99.8 percent. As the study examined survival in prostate cancer cases, conformity of the genotypes to HWE was only assessed in the cases; each SNP was stated to be in equilibrium.201

28-SNP panel. No specific information was presented in the article where this panel was reported.198

Three SNPs in 8q24. The three SNPs included in this test were part of 12 SNPs at 8q24 that were genotyped using the Mass ARRAY QGE iPLEX system (Sequenom), with a call rate of >98 percent and an average concordance rate between duplicate samples included in each-96-well plate of >99 percent.194 Genotype proportions were consistent with HWE in controls.

4-SNP test: KLK2, HPC1, TNF, ETV1 and 8q24, 17q24, TNF, ETV1. The Sequenom iPLEX technology was applied in the genotyping of the Canadian study used to develop these tests. The call rate was >90 percent for 25 SNPs; six of these were not in HWE and were excluded from further analysis.195 The call rate of SNPs significantly associated with prostate cancer was >95 percent.

Test for three SNPs in steroid hormone pathway genes. The three-SNP test in non-Hispanic whites was developed on the basis of the genotyping of 120 SNPs in the steroid hormone pathway by different methods.196 One hundred and four of the SNPs were genotyped using the GoldenGate assay (Illumina), four by TaqMan, and the remainder by methods described in four publications.109,211-213 It is stated that >80 percent of SNPs were successfully genotyped in >90 percent of the samples. Three SNPs failed (rs632148 within SRD5A2; rs280663 in HSD97B3; rs10877012 in CYP27B1) and one was not polymorphic (rs9332900 in SRD5A2). Three of the remaining SNPs were not in HWE in non-Hispanic whites and were excluded from the analysis of this ethnic group.

Test for two SNPs in steroid hormone pathway genes. The two-SNP test in Hispanic whites was developed on the basis of the genotyping of 120 SNPs in the steroid hormone pathway by different methods.196 One hundred and four of the SNPs were genotyped using the GoldenGate assay (Illumina), four by TaqMan, and the remainder by methods described in four publications.109,211-213 It is stated that >80 percent of SNPs were successfully genotyped in >90 percent of the samples. Three SNPs failed (rs632148 within SRD5A2; rs280663 in HSD97B3; rs10877012 in CYP27B1) and one was not polymorphic (rs9332900 in SRD5A2). Two of the remaining SNPs were not in HWE in Hispanic whites and were excluded from the analysis of this ethnic group.

6-SNP panel. This panel was developed to predict risk for prostate cancer in two sets of samples, and to predict risk for prostate cancer mortality in three, on the basis of genotyping six 8q24 and two 17q variants.197 The Sequenom iPLEX technology was used to genotype samples from the Physicians' Health Study and the Gelb Center; there was >99 percent concordance for six SNPs that were assessed on a subset (n=1,370) of specimens twice.197 The Applied Biosytems (ABI) SNPlex Genotyping System was used to genotype the samples from King County, Washington. None of the eight SNPs violated HWE in either set (Physicians' Health Study or King County, Washington) of controls. The call rate for the eight SNPs genotyped was >94 percent.

35-SNP panel. This panel was developed by genotyping 36 SNPs validated in previous studies using matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectrometry based on allele-specific primer extension using the Sequenom iPLEX technology.200 Genotyping rs2660753 (at 3p12) failed completely. For the remaining 35 SNPs, a 98.6 percent average call rate was reported. Hardy-Weinberg equilibrium was assessed in controls, no departure from HWE was observed (per an email from H. Grönberg, M.D., Ph.D. ( es.ik@grebnorG.kirneH) on February 2, 2012). The genotyping was performed at a core mutation analysis facility in Huddinge and was fully blinded to the case-control status (clarified in emails from H. Grönberg, M.D., Ph.D. ( es.ik@grebnorG.kirneH), and M. Aly, M.D. ( es.ik@yla.sukram) on February 2, 2012).

deCODE ProstateCancer test. The company's Web site states that the deCODE ProstateCancer test is performed by Illumina I-Select Bead Chip method – and based on proprietary Illumina technology using DNA amplification hybridization and fluorescent detection.208 Greater than 99.9 percent accuracy is claimed.

3. What are the sources of variation in accuracy or analytical validity across different test panels?

No evidence to address this question was identified.

KQ2. What is the clinical validity of available SNP-based panels designed for prostate cancer risk assessment?

1. How well do available SNP-based genotyping platforms predict the risk of prostate cancer in terms of

a. stratifying future risk and/or screening for current disease?

5-SNP panel (Focus 5) with and without inclusion of family history. Zheng, et al.,188 developed a model for the cumulative effect of five SNPs, selected as the most significant of 16 SNPs genotyped in five chromosomal regions (three at 8q24, and two at 17q). The number of genotypes associated with prostate cancer was counted for each subject and showed a significant trend of association, with the odds ratio (OR) for four or more genotypes compared with none being 4.47 (95% CI, 2.93 to 6.80, adjusted for age, geographic region, and family history). When family history was included in the risk score for each subject, the OR for five or more factors (genotype or family history) was 9.96 (95% CI, 3.62 to 24.72, adjusted for age and geographic region). Receiver operating curves were calculated. The area under the curve (AUC) for a model including age and geographic region was 57.7 percent (95% CI, 56.0 to 59.3), for a model adding family history to these factors was 60.8 percent (95% CI, 59.1 to 62.4), and for a model further adding in the number of genotypes associated with prostate cancer was 63.3 percent (95% CI, 61.7 to 65.0). These data were also presented in a later paper focusing on the development of a 28-SNP panel.198 In the later analysis, the sensitivities and specificities of a risk score combining the five SNPs and family history in first-degree relatives were presented for cutoffs of onefold, twofold, and threefold the median risk score (Table 4). As would be expected, sensitivity decreased and specificity increased with increasing cutoffs of absolute risk. The positive predictive value of the five SNPs (family history excluded) was 34 percent.

Table 4. Sensitivity and specificity for absolute risk of prostate cancer for risk score based on 5-SNP and family history (FHx) in first-degree relatives in Swedish study.

Table 4

Sensitivity and specificity for absolute risk of prostate cancer for risk score based on 5-SNP and family history (FHx) in first-degree relatives in Swedish study.

The model was tested in independent data from men of European origin in King County, Washington,189 in data from the Johns Hopkins Hospital and the PLCO Cancer Screening Trial,190 in a Canadian study,195 and in a study in which cases underwent radical prostatectomy in a hospital in Chicago.191 The pattern of association with risk score was attenuated compared with the original study of Swedish data,188 with the OR for four or more genotypes compared with the reference category of no risk genotypes being 3.36 (95% CI, 1.90 to 6.08, adjusted for age and family history) in King County, 2.42 (95% CI, 1.4 to 4.1) in the Canadian study, 2.84 (1.30 to 6.21) in Johns Hopkins Hospital, 3.09 (95% CI, 1.62 to 5.90) in the PLCO Trial, and 3.19 (95% CI, 1.85 to 5.50, adjusted for age) in Chicago. In the Canadian study, the AUC for a baseline model that included age, family history of prostate cancer, ethnicity, urinary symptoms, PSA, free: total PSA ratio, and DRE was 72 percent (95% CI, 70 to 74), and with the addition of five SNPs, 73 percent (95% CI, 71 to 75).195 In these studies, the proportion of controls with four or more risk genotypes ranged between 1.6 percent190 and 3.4 percent,191 while the population with five or more risk factors (one of which could be family history of prostate cancer) was 0.3 percent or less.188-190

When family history was included in the risk score, the ORs for five or more risk factors compared with none was 4.92 (95% CI, 1.58 to 18.53, adjusted for age) for King County,189 and 20.68 (95% CI, 2.61 to 163.85) for the PLCO trial.190 In the King County data, the AUC for a model including age, serum PSA level, and history of prostate cancer in a first-degree relative was 63 percent, which increased to 66 percent when the five SNPs were added (difference 3 percent, 95% CI, -12 to +6); this difference was not statistically significant.189

9-SNP panel. Helfand, et al.191 extended the 5-SNP model, adding four variants at 2p15, 10q11, 11q13, and Xp11. The OR associated with having six or more of the nine risk genotypes was 5.75 (95% CI, 2.50 to 13.24), and the proportion of controls in the category of highest risk was 2.5 percent. For the model with five genetic variants, the crude AUC was 58 percent, and with adjustment for age, 65 percent. With inclusion of the four additional variants, the AUCs were 61 percent and 66 percent, respectively.

17-SNP panel. In the Chicago-St. Louis study (Helfand et al.),199 the 9-SNP model was modified by changing one variant at 2p15, and adding one variant at 3q21.3, 11q13, 17q12, 19q13.2, and two at 5p15 and 8q24. The study differed from the others in that it was limited to men with a PSA level <4.0ng/ml and with normal DRE, and cases were limited to clinical stage T1c. Compared with men who had four or fewer variants, the OR for men with 11 or more variants was 10.6 (95% CI, 2.7 to 42.0), and the proportion of controls in this highest risk category was 2.5 percent. When history in first-degree relatives was added to the risk score, compared to men with zero to five variants/family history, the OR for men with 11 or more variants was 11.2 (95% CI, 4.3 to 29.2), and the proportion of controls in this highest risk category was 3.2 percent. The AUC for the model including all the carrier numbers of the 17 SNPs was 0.66; this was not significantly different from an AUC of 0.62 for age alone. The AUC of a model containing the 17 SNPs and family history was 0.71, which was statistically significantly higher than the model based on age alone.

11-SNP panel. Zheng, et al.,192 examined the effect of including 14 additional SNPs in the same Swedish study participants as in the original 5-SNP model.188 On the basis of an SNP by SNP analysis, 12 remained associated with prostate cancer risk after adjustment for age, family history, geographic region, and the other SNPs. However, one of these SNPs was not included in further analysis because it was originally discovered in this study population and “has not been extensively confirmed in other study populations.”192 Thus, further evaluation focused on counts of risk alleles for 11 SNPs and family history. The AUC for a model involving age only was 58 percent (95% CI, 56 to 59), for age and family history was 61 percent (95% CI, 59 to 62), and for age, family history, and all eleven SNPs was 65 percent (95% CI, 63 to 66). Stratified analysis of data on sensitivity and specificity by number of risk factors did not show differences by disease aggressiveness or age at diagnosis. These data were also presented in a later paper focusing on the development of a 28-SNP panel.198 In the later analysis, the sensitivities and specificities of a risk score combining the 11 SNPs and family history in first-degree relatives were presented for cutoffs of onefold, twofold, and threefold the median risk score (Table 5). As would be expected, sensitivity decreased and specificity increased with increasing cutoffs of absolute risk. The positive predictive value of the 11 SNPs (family history excluded) was 37 percent.

Table 5. Sensitivity and specificity for absolute risk of prostate cancer for risk score based on 11-SNP and family history (FHx) in first-degree relatives in Swedish study.

Table 5

Sensitivity and specificity for absolute risk of prostate cancer for risk score based on 11-SNP and family history (FHx) in first-degree relatives in Swedish study.

14-SNP panel. The Swedish data were also investigated in development of a prediction model of absolute risk for prostate cancer using 14 SNPs and family history, and using data for the PLCO trial for confirmation.193 The number of risk alleles could range from zero to 27 (because one of the risk alleles was on the X chromosome), with the mode being 11 for controls. In the Swedish data, the OR for prostate cancer in men who had ≥14 risk alleles and positive family history (which occurred in 1 percent of control men) compared with men with 11 risk alleles and no family history of prostate cancer was 4.92 (95% CI, 3.64 to 6.64). The corresponding OR for the PLCO trial data was 3.88 (95% CI, 2.83 to 5.33). In the Swedish data, the risk did not differ between aggressive and nonaggressive disease. With regard to absolute risk in Sweden, a 55 year old man with ≥14 risk alleles and a positive family history was estimated to have a 52 percent risk of being diagnosed with prostate cancer in the next 20 years, compared to a risk of 8 percent for men with seven or fewer risk alleles and no family history. The corresponding estimates for the men in the United States were 41 percent and 6 percent, respectively.

28-SNP panel. The Swedish data were also used in the development of a 28-SNP panel.198 The AUC for the panel was 0.62, compared with 0.61 for the 11-SNP panel, and 0.60 for the 5-SNP panel; these differences were statistically significant. The sensitivities and specificities of a risk score combining the 28 SNPs and family history in first-degree relatives were presented for cutoffs of onefold, twofold, and threefold the median risk score (Table 6). As would be expected, sensitivity decreased and specificity increased with increasing cutoffs of absolute risk. The positive predictive value (PPV) of the 28-SNPs (family history excluded) was 37 percent. When the SNPs and family history were sorted on the basis of their contribution to genetic variance, from highest to lowest, at each cutoff of onefold, twofold, and threefold population median risk, the PPV increased only slightly with increasing numbers of SNPs.

Table 6. Sensitivity and specificity for absolute risk of prostate cancer for risk score based on 28-SNP and family history (FHx) in first-degree relatives in Swedish study.

Table 6

Sensitivity and specificity for absolute risk of prostate cancer for risk score based on 28-SNP and family history (FHx) in first-degree relatives in Swedish study.

Three SNPs in 8q24. One study in the Johns Hopkins Hospital investigated multiple variants of 8q24 in men with prostate cancer who had at least two additional first-degree relatives with prostate cancer, men who did not fall into this category, and controls.194 To assess the combined effects of variants in three regions of 8q24, one variant from each region was selected. Compared to men with no risk genotype, the OR of prostate cancer for men with 2+ affected first-degree relatives for two or more risk genotypes was 2.94 (95% CI, 1.68 to 5.15), and for prostate cancer without such a family history was 2.23 (95% CI, 1.52 to 3.28).

4-SNP test: KLK2, HPC1, TNF, ETV1. In a Canadian study,195 in addition to examining the 5-SNP model of Zheng, et al.,188 a model comprising four SNPs, one each in KLK2, HPCI, TNF, and ETV1 was evaluated. The OR associated with presence of all four variants compared with none was 2.53 (95% CI, 1.6 to 4.1). The proportion of controls that had variants of all four SNPs was 3.2 percent. The AUC for the baseline model that included age, family history of prostate cancer, ethnicity, urinary symptoms, PSA, free: total PSA ratio, and DRE was 72 percent (95% CI, 70 to 74), and with the addition of the four SNPs was 73 percent (95% CI, 71 to 74).

4-SNP test: 8q24, 17q24, TNF, ETV1. In the same Canadian study,195 a model comprising four SNPs, one each from 8q24, 17q24.3, TNF, and ETV1, was evaluated. The OR associated with presence of all four variants compared with none was 6.07 (95% CI, 2.0 to 18.5). The proportion of controls that had variants of all four SNPs was 0.3 percent. The AUC for the baseline model that did not include SNPs (see above) was 72 percent, and with the four SNPs included was 74 percent (95% CI, 72 to 76). Using two thirds of the data, the investigators developed a nomogram that incorporated these SNPs, age, family history of prostate cancer, ethnicity, urinary voiding symptom, PSA level, free: total PSA ratio, and DRE in predicting all prostate cancer, and predicting prostate cancer with a Gleason score of 7 or more. Predicted and actual probabilities were compared in the remaining one third of the data, and the incremental drop in AUC for each predictor variable when removed from the nomogram model was assessed. The incremental drop was greater (1.4 percent) for the SNP combination than PSA (0.1 percent), family history of prostate cancer (0.3 percent), urinary voiding symptom (0.1 percent), and DRE (1.0 percent), but not age (2.2 percent) or free: total PSA ratio (6.6 percent).

Test for three SNPs in steroid hormone pathway genes. Beuten, et al.,196 examined SNPs in the steroid hormone pathway. They presented information on the cumulative effect of three risk variants, (one in HSD3B2, two in CYP19) in non-Hispanic whites. There was a trend with an increasing number of risk genotypes. The OR for three risk genotypes compared with none was 2.87 (95% CI, 1.64 to 5.02, adjusted for age), with 3.6 percent of controls in the category of highest risk.

Test for two SNPs in steroid hormone pathway genes. In the investigation of SNPs in the steroid hormone pathway described in the preceding subsection, Beuten, et al.,196 presented information on the cumulative effect of two risk variants (one in CYP19, different from those in non-Hispanic whites, one in CYP24A11) in Hispanic whites. Again, there was a trend with an increasing number of risk genotypes. The OR for two risk genotypes compared with none was 4.58 (95% CI, 2.19 to 9.61, adjusted for age), with 5.6 percent of controls in this category of risk.

6-SNP test. Penney, et al.,197 evaluated eight SNPs, six in 8q24 and two in 17q, in data from the Physicians' Health Study (PHS) and from King County, Washington. Four of the 8q24 and the two 17q SNPs were significantly associated with prostate cancer in the two data sets, and the association with a risk score obtained by adding up the alleles was evaluated. The risk of prostate cancer increased by 19 percent for each additional risk allele in the PHS, and 23 percent in King County.

35-SNP panel. Aly, et al.,200 focused their analyses relating to clinical validity of a 35-SNP panel on men with a PSA level ≤10 ng/ml as they considered that there is most debate over recommending a prostate biopsy in this group than in men with a higher PSA level. A genetic score was calculated by summing the number of risk alleles (0,1, or 2) at each of the 35 SNPs multiplied by the logarithm of the OR for that SNP. In univariate analysis, the OR associated with this score was 1.93 (95% CI, 1.85 to 2.01), with an AUC of 0.61 (95% CI, 0.59 to 0.63). In multivariate analysis, adjusting for PSA, the ratio of free-to-total PSA, age, and family history, the OR was 1.52 (95% CI, 1.45 to 1.59). The AUC for PSA, the ratio of free-to-total PSA, and age was 0.63 (95% CI, 0.60 to 0.65); the addition of family history increased this to 0.64 (95% CI, 0.62 to 0.66) and adding both family history and the genetic score increased the AUC to 0.67 (95% CI, 0.65 to 0.70).

Different risk cutoffs were assessed for: 1) the model comprising PSA, the ratio of free-to-total PSA, age, and family history; 2) the addition of the genetic score to this model; and, 3) a hypothetical genetic model based on a score variable constructed from SNPs explaining 100% of the population genetic risk. Comparisons were made of how these would affect the numbers of biopsies performed and cancer detected per 1,000 men with a clinical prostate biopsy (Table 7). The addition of the 35 SNPs (Model 2) to the factors included in Model 1 would reduce the number of biopsies conducted but increase the number of missed cancers. For the hypothetical genetic model (Model 3), the number of biopsies would be further reduced compared with Model 2, and the increase in proportion of missed cancers reduced.

Table 7. Comparison of effects on biopsies conducted and cancer detected per 1,000 men with a clinical prostate biopsy between three models of risk prediction for prostate cancer and two cutoffs.

Table 7

Comparison of effects on biopsies conducted and cancer detected per 1,000 men with a clinical prostate biopsy between three models of risk prediction for prostate cancer and two cutoffs.

deCODE ProstateCancer test. The deCODE Prostate Cancer Web site states that the predictive accuracy of the 27-SNP ProstateCancer test panel, the 9-SNP subset for African-American men, and the 12-SNP subset for men of East Asian descent is essentially independent of, and therefore complements, the risk confirmed by family history of the disease.208 The validity is reported to be based on the evaluation of risks associated with single SNPs; it is stated that the validity of multiplying together the risk conferred by different markers is based on the lack of significant interaction or overlap of impact between markers in two studies.165,168

b. Distinguishing between clinically important and latent/asymptomatic prostate cancer

5-SNP panel. In a case-only analysis of combined data from the Swedish, Johns Hopkins Hospital, and PLCO Trial participants, there was no statistically significant association between the five genetic variants, Gleason score, aggressiveness of prostate cancer,214 or age at diagnosis.190

14-SNP panel. In the Swedish data investigated in the development of a prediction model of absolute risk for prostate cancer using 14 SNPs and family history, the OR for aggressive prostate cancer in men who had ≥14 risk alleles and positive family history compared with men with 11 risk alleles and no family history of prostate cancer was 4.77 (95% CI, 3.41 to 6.69).193 The corresponding OR for nonaggressive prostate cancer was 5.05 (95% CI, 3.66 to 6.96). In addition, the risk associated with each increase in the number of risk alleles did not differ between aggressive and nonaggressive disease.

11-SNP panel. In the analysis of Zheng, et al.,192 which developed a model comprising counts of risk alleles for 11 SNPs and family history, stratified analysis of data on sensitivity and specificity by number of risk factors did not show differences by disease aggressiveness or age at diagnosis.

35-SNP panel. In the study of Aly, et al.,200 aggressive disease was defined as T3-4 N1 M1 or Gleason 4+3 and higher, and nonaggressive disease as T0-2 N 0/X M 0/X or Gleason 3 +=4 and lower. The increase in AUC for aggressive disease between a SNP-based model (35-SNPs) and a non-SNP-based model based on PSA, the ratio of free-to-total PSA, age, and family history was not statistically significant.

c. How well do available SNP-based genotyping panels predict prognosis in individuals with a clinical diagnosis of prostate cancer?

5-SNP panel (Focus 5) with and without inclusion of family history. In the study in King County,189 described above, the predictive ability of the SNP panel for prostate cancer specific mortality over an average length of followup of 7.6 years was evaluated. There were 45 deaths among 1,207 men with followup data; there was no association with the SNPs individually or in combination, and they did not increase the AUC for a model that included age at diagnosis, serum PSA at diagnosis, Gleason score, and tumor stage (difference in AUC between model including SNPs compared to one without 0.5 percent, 95% CI, -1 to +2).

6-SNP test. In a survival analysis of the six SNPs found to be associated with prostate cancer in the data from the PHS and King County using the Cox proportional hazards model, there was no significant association between these variants and prostate cancer mortality.197 In addition, comparison was made between prostate cancer deaths and men alive more than 10 years after diagnosis in a combined analysis that included both of these samples, together with a series of cases from the Dana-Farber Harvard Cancer Center diagnosed over the period from 1976 to 2007. The total number of risk alleles was not associated with mortality.

16-SNP panel. In a population-based study of survival after prostate cancer diagnosis in 2,875 men in Sweden over an average of 4.9 years (range 3.7 to 6.8 years), there was no association between prostate cancer mortality in a comparison with the average number of risk alleles, in a test for trend with an increasing number of risk alleles, or in relation to specific individual variants within the panel.201

None of the studies reported above presented data on risk reclassification or performance in simulation analyses.

3. What other factors (e.g., race/ethnicity, gene-gene interaction, gene-environment interaction) affect the predictive value of available panels and/or the interpretation of their results?

Beuten, et al.,196 developed separate tests for SNPs in steroid hormone pathway genes for non-Hispanic whites and Hispanic whites (see above).

deCODE markets the ProstateCancer test, which tests for 27 genetic variants (Table 13) associated with prostate cancer in men of European descent (including the five SNPs included in the Focus 5 test), a subset of nine variants for African-American men, and a subset of 12 variants for men of East Asian descent; the specific variants in the subsets are not specified in the Web site (www.decodhealth.com/prostate-cancer).208

KQ3. What is the clinical utility of available SNP-based panels designed for prostate cancer risk assessment?

Process of care

1. Does the use of panels alter processes of care and behavior?

  1. screening or management decisions, and the appropriateness of these decisions, by patients and/or providers
  2. alteration in health-related behaviors of patients (e.g., adherence to recommended screening interventions and/or other lifestyle changes)?

No data addressing this question were identified.

Health outcomes

2. Does the use of panels lead to changes in health outcomes?

  1. all-cause mortality
  2. cancer-specific mortality
  3. morbidity

And do any changes vary by race or ethnicity?

No data addressing this question were identified.

Harms

3. Does the use of panels lead to harms?

  1. psychological harms
  2. other negative individual impacts (e.g., discrimination) and do any such harms vary by race or ethnicity?

No data addressing this question were identified.

Costs

4. What is known about the costs, cost-effectiveness, and/or cost utility of using SNP-based panels for prostate cancer risk assessment, compared to current practice?

No data addressing this question were identified.

Quality Assessment of Individual Studies

All included studies were related to clinical validity, which usually lends itself to a medical test framework for quality assessment. However, we decided to use the Newcastle-Ottawa Scale (NOS)185 (Table 14a) because all but one of the studies had a case-control design (the exception being a cohort study of prostate cancer survival201), and because it is not clear how well the QUADAS186 tool would apply to genetic tests. We supplemented this with selected items from the QUADAS186 tool to assess the risk prediction aspect of the included studies. These were: (1) whether the spectrum of participants was representative of the patients who would receive the test in practice; (2) whether the selection criteria were clearly described; and, (3) whether un-interpretable, indeterminate, or intermediate test results were reported (Table 14b). Other QUADAS186 criteria considered when assessing the risk of bias of the studies included whether or not: 1) the whole sample or a random selection of the sample received verification using the reference standard; 2) participants received the same reference standard regardless of the index test result; 3) the reference standard was independent of the index test; 4) the execution of the index test was described in sufficient detail to permit its replication; and, 5) the same clinical data were available when the test results were interpreted as would be available when the test is used in practice.

The reference standard for cases was histopathological diagnosis in all of the studies, but checking for latent or undiagnosed cancer was not conducted in control groups with two exceptions.195,200 Autopsy studies in men over 50 years of age who had died from other causes have demonstrated a frequency of histologically proven prostate cancer of 30 to 40 percent.54-60 However, there are clearly ethical constraints to taking prostate tissue samples in asymptomatic men in order to exclude an undiagnosed disease. In one of the studies, controls were selected from the same group of men referred to prostate cancer centers who had either a PSA value ≥4.0ng/ml or an abnormal DRE and who had no biopsy evidence of prostate cancer.195 The results of the clinical validity evaluation of the 5-SNP panel in this study were similar to those of the other studies in which this panel was evaluated.189-191 In all of the studies, it seems unlikely that the index test result affected the decision to undertake prostate biopsy, or the interpretation of histopathological examination of biopsy specimens. However, since all of the studies were conducted in research contexts, it is not clear that decisionmaking incorporated the same clinical data as would have been available in routine practice.

The execution of the genotyping component of the index test was adequately described in all but one198 of the studies (see section on analytic validity). Almost all of the studies related to participants of European origin, and those that did not adjusted for ethnicity or conducted analyses restricted to participants of European origin. This is likely to have limited the risk of bias resulting from population stratification; that is, the presence within a population of subgroups among which allele (or genotype, or haplotype) frequencies and disease risks differ.215-218 However, some of the other variables included in risk scores may have been prone to differential error because of the retrospective case-control design used in all but the PLCO Trial,193,203 the PHS,197 and the San Antonio cohort.196

By combining the results of the NOS185 evaluation and the QUADAS186 criteria for the individual studies, all studies of the 5-SNP panel were found to have a moderate risk of bias. Based on three selected domains in the NOS185 (selection of controls, comparability of cases and controls, method of ascertainment of cases and controls), along with limited data about genotyping methods and quality control, lack of specification of which candidate nongenetic variables were initially examined or considered for inclusion in the risk models, and lack of information about how these variables were assessed, the overall risk of bias of was assessed as being at least ‘moderate’. Using the same approach, the assessments of the other 14 panels were based on single studies, reported in eleven articles,191-201 and these were also all considered to have at least a moderate risk of bias.

Rating the Body of Evidence

Four domains were considered in the assessment of overall strength of evidence (SOE) for the SNP panels identified. These were risk of bias (internal validity of the studies), the consistency of findings, directness (how closely the tests were applied in a way which resembles routine practice), and precision (whether the estimates allow clinically useful conclusions).

For the domain of internal validity, all studies were assessed as having at least a moderate risk of bias. For the domain of consistency, it is impossible to assess results for panels evaluated in single studies only. For the Focus 5 panel, where there were several studies, the data did not permit development of an ROC curve, and therefore consistency could also not be assessed quantitatively. For models containing the five SNPs included in the Focus 5 panel, but with diverse other variables included, the AUC ranged between 63 percent and 73 percent.188,193,195 Compared with the models that did not include the SNPs, the 5 SNPs increased the AUC by 1 to 3 percent.

For the domain of directness, all studies were conducted in a research context, no panel being applied in a setting that might be considered close to routine clinical practice. As well as presenting difficulty in assessing generalizability to a ‘typical’ clinical approach, this meant that none of the tests were explicitly evaluated in a medical test framework. Specifically, the case-control design meant there was no meaningful comparison of any SNP panel against a routine clinical alternative ‘test’. Finally, the assessment of the precision domain requires a clear idea of clinically meaningful differences between levels of sensitivity, specificity, AUC, and other accuracy metrics (i.e., how much difference in one of these would make a ‘real’ difference in clinical or patient decisionmaking). This area of evaluation appears to be underdeveloped in the clinical literature, and the studies evaluated shed no light on this aspect. We were therefore unable to offer a valid assessment of this domain.

We are unable to assess the extent of publication bias in this review. We contacted a comprehensive list of companies we considered most likely to be developing SNP panels for commercial application, and received no responses. It is possible that unpublished data exist to support the clinical validity of one or more of the SNP panels reviewed here, or of other SNP panels which were not identified in this report. If so, this review's conclusions would be unduly negative. However, this would be an unlikely scenario, since publication bias is usually considered to lead to selective reporting of studies with systematically larger effect sizes than is actually the case.219 Only papers published in English were included. There is no empirical evidence of the effects of language restriction on genetic risk prediction studies. Although there is some empirical evidence of systematic differences in effect sizes of genetic associations reported in studies in Asian populations published in English and in Chinese, it is not clear that these differences are due to publication bias.220 Moreover, there is evidence of considerable overlap of publications in English and Chinese medical journals on the same studies.221 In the literature on randomized controlled trials, restriction to English language publications does not appear to bias estimates of effectiveness of conventional interventions.222

Overall, it is unlikely that any of the biases identified would be sufficient to alter the interpretation of the findings from (at best) inadequacy of evidence to clearly positive supporting evidence for any of the SNP panels reviewed.

For characteristics of included studies see Tables 810. The Focus 5 test is reported in Table 11. Summary of SNPs and other variables included in test panels is reported in Table 12. Table 13 reports genetic variants tested for by deCODE ProstateCancer and Table 14 reports case studies on the Newcastle-Ottawa Scale.