An expanded biomarker panel for the detection of prostate cancer from urine DNA

Background Prostate cancer diagnosis using the PSA test remains controversial because of overdiagnosis and overtreatment of potentially indolent cancers. There remains a need to increase the diagnostic lead time and to target treatment to patients with significant disease. One possible approach to overcome the limitations of PSA is to screen men for the molecular signature of early PCA, monitor the rate of disease progression and target treatment to patients who are likely to benefit from it. Such an approach requires a large panel of markers that define a molecular clock for PCA. We recently developed a panel of 19 markers for the non-invasive detection of PCA from urine DNA. It raised the possibility that additional methylation markers could be successfully analyzed from urine DNA, a prerequisite for increasing the diagnostic lead time and enabling disease monitoring. Methods We developed semi-quantitative polymerase chain reaction assays for 13 additional markers and determined their methylation status in 150 urine DNAs from 94 patients with elevated PSA. Eighty five samples were obtained following DRE and 65 samples were from first void. We combined the data of the 13 new markers with the previously reported 19 markers and calculated the sensitivity, specificity, negative and positive predictive values at every threshold from one to 32 positive markers. Results Using 10of32 positive markers as the threshold to recommend a biopsy yields a sensitivity of 81% (95% CI 0.68–0.93) and 93% (95% CI 0.84–1.02) and a specificity of 76% (95% CI 0.63–0.88) and 77% (95% CI 0.63–0.91) from DRE and FV DNA, respectively. The PPV was 71% and 77% and the NPV was 85% and 93% from DRE and FV, respectively. Conclusions This study shows that large marker panels can be analyzed from urine DNA without loss of sensitivity or specificity. Using 32 markers improved the stratification of patients undergoing screening for PCA particularly for patients below the 10of32 threshold. The results show the utility of larger biomarker panels for PCA diagnosis and suggest that the development of the panels needed to monitor disease progression could be successfully accomplished. Electronic supplementary material The online version of this article (10.1186/s40164-019-0137-x) contains supplementary material, which is available to authorized users.


Background
Over the last 2 decades, prostate cancer (PCA) screening using the prostate specific antigen (PSA) test exposed the complexities and challenges associated with early screening for a common cancer [1]. PCA is a heterogeneous disease with a long natural history and a wide range of prognostic outcomes. PSA screening advanced the PCA diagnosis by 5 to 7 years [2,3]. However, the improved lead time resulted in only a modest reduction in PCAspecific mortality and was not sufficient to diagnose all prostate cancers while curable [4].
Autopsy studies have shown that PCA can be histologically detected in approximately 50% of men over 80 years old, vastly exceeding its diagnostic and mortality rates [1,2,[5][6][7]. The large reservoir of slow growing prostate tumors contributes to the overdiagnosis and overtreatment of PCA when asymptomatic men are screened with PSA. It brought into question the value of early screening [1,3,[8][9][10][11]. Clinically significant disease which affects less than a quarter of PCA patients remains difficult to detect early [12,13]. Its diagnosis is further complicated by a highly variable rate of disease progression. A number of non-invasive diagnostic and prognostic tests have been developed to improve on PSA screening but none have achieved the accuracy needed to safely reduce overtreatment [14][15][16][17][18]. Their primary objective is to predict the likelihood of a significant cancer in patients with elevated PSA. This approach is limited by the PSA test itself which lacks the specificity and lead time needed for an early cancer screening test. Non-invasive molecular tests to increase the diagnostic lead time and enable monitoring for disease progression are still needed. A longer lead time will ensure that all patients are diagnosed while their disease is curable. Measuring the rate of disease progression will enable the targeting of treatment to patients with high risk tumors before the cancer becomes disseminated. These tests will likely require large panels of markers that can be detected in circulating DNA and correlated with disease grade and stage.
The most common cancer-specific aberration is the progressive accumulation of DNA methylation. It affects hundreds of CpG islands and can be assessed in liquid biopsies. We recently published a biomarker panel for the non-invasive diagnosis of PCA [19]. The panel detects the cumulative methylation of 19 markers in urine DNA recovered following a digital rectal exam (DRE) or from first void (FV) samples. The test measures the number of methylated markers in urine DNA and shows that the likelihood of a positive biopsy increases with increasing number of methylated markers. Using six methylated markers out of nineteen (6of19) to define a cancer case yielded a sensitivity of ≥ 89%, a specificity of 71%, a positive predictive value (PPV) of ≥ 71% and a negative predictive value (NPV) of ≥ 90%. The ability to analyze 19 methylation markers from circulating DNA without loss of specificity was unexpected. It raised the possibility that larger methylation panels could be successfully analyzed from urine DNA without loss of diagnostic accuracy.
We undertook the current study to determine the feasibility of analyzing additional markers in urine DNA. We selected 13 CpG islands with methylation ranging from 50 to 80% in primary prostate tumors (FRZB, GPR147, GPR62, GRASP, HOXA11as, HOXBAS3, HOXCrcAS3, HOXD3c, HOXD4rc, HOXD8rc, KLK10, RASSF1, SLC16A5rc) (authors unpublished data) and determined their methylation in 150 DNAs from 94 patients using semi-quantitative methylation-specific PCR. In this study, we report the methylation of the 13 new markers as well as the cumulative methylation of the 32 markers.

Patient cohort
Urine samples were collected under an IRB protocol approved by Western Institutional Review Board (WIRB, study # 1139453, Puyallup, WA) and were previously described [19]. Patients with positive biopsies were grouped based on CAPRA risk criteria into a low risk group (CAPRA score ≤ 2) and an increased risk group (CAPRA scores ≥ 3) [20]. Four DNA samples were lost during processing reducing the number of DNAs analyzed to 85 from DRE and 65 from FV.

Marker analysis
Marker assays were validated and urine DNAs were analyzed as described [19]. The MS-qPCR reactions were performed on the QuantStudio 6 (ABI) with the same bisulfite converted DNAs used for the first 19 markers [19]. Primer and probe sequences for all markers are shown in Additional file 1: Table S2. An assay for an imprinted gene (STIM1) was added to the primary amplification multiplexes to verify the recovery of amplifiable DNA from urine samples. The data for the additional 13 markers were collected after the first 19 markers were unblinded.

Data collection
The data were tabulated using the software provided with the QuantStudio 6. The data were further transformed by subtracting the Cq values from 32 (except for the 0 data points) to generate an increasing range of values from 0 (no amplification) to 15 (highest level of amplification, lowest Cq) as previously described [19]. The data were used directly for statistical analysis with no further manipulations.

Statistical analysis
Each subject in the study had at least one type of urine sample (DRE or FV) collected. Subject characteristics were summarized within the cases and within the controls, respectively. The cases are defined as the subjects with positive diagnosis of prostate cancer based on biopsy, and the controls are defined as those with negative diagnoses. Arithmetic mean and standard deviation were summarized for continuous characteristics, and frequency and percentage were calculated for categorical characteristics. Characteristics such as Gleason score and positive cores were available only for cases, and were summarized at both continuous and dichotomized levels.
The following statistical analyses were performed for both DRE and FV samples unless otherwise specified. Sensitivity and specificity associated with the presence of individual methylation markers were computed using the observed proportion of individuals with positive markers conditional upon diagnosis status, and their 95% confidence intervals were provided.
Similarly, sensitivity, specificity, negative and positive predictive value and their 95% confidence intervals were calculated for each of the possible number of positive markers among the 32 markers. The average methylation for each DNA sample was calculated by adding the values obtained for all 32 assays and dividing by 32. Box plots were generated by diagnosis status for both the number of positive markers and the average methylation levels of the 32 markers. ROC curves were plotted for the 32 markers based on either average methylation or the number of positive markers. We did not perform any modeling to further characterize the markers because of the small number of patients analyzed.
The average number of positive markers by patient grading group was compared using the Wilcoxon ranksum test. Grading groups of grade 0 (negative biopsy), 1 (positive biopsy, low risk group based on CAPRA), or 2 (positive biopsy, elevated risk group based on CAPRA) and the group combining grade 1 and 2 (all positive biopsies) were considered. Similar analysis was performed to compare the means of average methylation by grading group. Paired sample analysis was also performed using the paired t test. All statistical analyses were performed using R version 3.4.4 (https ://cran.r-proje ct.org), and the AUC calculations were performed with R package "ROCR" with version 1.0.7.

Patient characteristics
The patient cohort was described previously and included 52 patients with negative biopsies (controls) and 42 patients with positive biopsies (cases) [19]. The first void and DRE urine samples were collected prospectively up to 6 weeks prior to biopsy. Patient demographics are shown in Table 1. The mean Gleason score was 7.1 (SD = 1.3) and the mean number of positive cores was 4.2 (SD = 3.6). Clinical follow up was available for a subset of patients after the marker data was collected. Three patients who had a negative biopsy after urine collection were diagnosed with PCA within 2 years. They were included in the cancer group at the start of the statistical analysis.

DNA methylation in DRE and FV DNA
A presence (> 0) or absence of methylation (= 0) was used for the calculations of sensitivity and specificity regardless of the amount of methylation detected in urine DNA to limit any subjective interpretation of data to the analytical assay conditions. A low level of methylation may reflect a small number of methylated copies in urine DNA or may be an analytical error from incomplete deamination of genomic DNA (false positives). Including all positive markers regardless of level of methylation and without using cutoffs based on clinical data incorporates all potential false positives in the results of individual markers and in the cumulative urine methylation calculations and eliminates false negatives. The 13 markers were recovered with variable frequencies from both DRE and FV. The observed sensitivities of individual assays ranged from 19 to 87% while specificities ranged from 46 to 100%. Table 2 shows the sensitivity, specificity, positive and negative predictive values (PPV and NPV) for all 13 markers.

Cumulative methylation in DRE and FV urine DNA from biopsy positive and biopsy negative patients
Using the presence of methylation (> 0) to classify markers as positive, we calculated the number of positive markers for each DNA sample. The median number of methylated markers in cases was 16 in both DRE and FV (range: 3-31 in DRE and 6-31 in FV). The median number of methylated markers in controls was 5 in both DRE and FV (range: 0-17 for DRE and 0-18 for FV). Table 3 shows the sensitivity, specificity, PPV and NPV for the total number of methylated markers (nof32) at every threshold from 1 to 32. Using 10 out of 32 positive markers (10of32) as the threshold to refer a patient for biopsy yields observed sensitivities of 0.81 and 0.93 and specificities of 0.76 and 0.77 for DRE and FV, respectively. A specificity of ~ 70% is the target specificity for a PCA diagnostic test because the sampling errors of prostate biopsies can lead to a negative finding for up to a third of cancer patients [21][22][23]. The 10of32 threshold yields an estimated NPV of 0.85 and 0.93 and an estimated PPV of 0.71 and 0.77 for DRE and FV, respectively. The results of the 32 and 19 marker panels were comparable. Increasing the number of markers to 32 was not expected to improve the sensitivity or specificity of the 19 marker panel but it improved the stratification of patients, in  particular patients who did not meet the threshold for a positive diagnosis. The receiver operating characteristic curve (ROC) was calculated based on the number of methylated markers and the average methylation of all 32 markers. Figure 1 shows the ROC curves for the nof32, average methylation and PSA for the DRE and the FV data. The Area under the ROC curve (AUC) ranged between 0.87 in DRE to 0.91 in FV, a significant improvement over PSA (AUC = 0.688).

Comparison between the DRE and FV methylation results
Both DRE and FV urine samples were available for 32 patients with a negative biopsy and 26 patients with a positive biopsy. Paired sample analysis was performed to compare the methylation of individual markers in DRE and FV DNAs. Figure 2 shows the observed within subject mean difference in methylation levels for individual markers, total number of positive markers and average methylation. There was no significant difference in the recovery of the majority of markers between DRE and FV. Of the 32 markers, AOX1, GFRA2, and NEUROG3 were better recovered from DRE samples (p < 0.05), and GPR62 and HOXD3c were better recovered from FV DNA (p < 0.05). The observed differences for these five markers are small. They are likely due to the position of the underlying assays within the CpG island and the small number of samples analyzed.

Comparison of urine DNA methylation and Gleason score and tumor volume
The histopathological findings (Gleason score, # of positive cores, tumor volume) varied widely between patients. Some cancers were limited to a single focus in a single core (GS of 6 to 8, ≤ 1% core volume) while others showed widespread, poorly differentiated cancer in multiple cores (GS of 8 to 10, up to 12 positive cores and up to 100% tumor volume per core). Similarly, the number of methylated markers and the average methylation varied widely. Patients with positive biopsies were grouped based on UCSF-CAPRA risk scoring system into a low risk group (Group 1: CAPRA score of 1 and 2) and an elevated risk group (Group 2: CAPRA score ≥ 3) [20]. Patients in Group 2 have an intermediate risk (CAPRA score of 3-5) except for 5 patients diagnosed with higher grade tumors (CAPRA score 6-9). The minimum, mean and maximum values obtained for the number of methylated markers and average methylation for each group are shown in Additional file 2: Table S3. The mean number of methylated markers and the average methylation differed significantly between cases and controls for both DRE and FV (Wilcoxon p-values < 0.001 for both DNA types). Furthermore, both parameters differed significantly between Group 1 and 2 patients for DRE DNA (Wilcoxon p-values of 0.001 for average methylation and 0.005 for the number of methylated markers) but not for FV DNA (Wilcoxon p-values of 0.4578 and 0.647, respectively). Figure 3 shows the distribution of average urine DNA methylation and the number of positive markers for all 3 groups. Methylation of DRE DNA outperformed that of FV in identifying patients with higher Gleason score or higher overall tumor volume.

Discussion
We previously analyzed this patient cohort with 19 markers [19]. In this study, we determine the methylation of 13 additional markers and calculate the cumulative methylation of the full 32 marker panel. A single cutoff of 10of32 positive markers was used as the threshold needed to recommend a biopsy. However in a clinical setting, a personalized recommendation would be made based on the pattern of urine methylation for each patient. The 32-marker panel improved patient stratification over what was achieved with the 19-marker panel, especially for patients who did not meet the threshold needed to recommend a biopsy. Improving the stratification of patients undergoing PCA screening is important to decrease the frequency of repeat biopsies and to better predict the disease stage and grade in positive patients. Methylation affects hundreds if not thousands of CpG islands in prostate cancer. The markers were selected strictly based on the analytical conditions used for the bisulfite conversions. In urine DNA, some markers like HOXD3, HOXA7, GPR62 and KLK10 were detected with comparable frequency from all cancer patients and are candidates for biomarker panels used for the early detection of PCA. Others like HOXD8rc, CXCL14, SLC16A5rc, and GRASP were recovered more frequently from patients with higher Gleason grade and higher  This study was not performed to finalize the marker panel for a clinical test but to determine the feasibility of analyzing larger marker panels. Prostate cancer detected on biopsy is highly heterogeneous in volume and Gleason grade with prognostic outcomes that are difficult to predict. To identify accurate prognostic markers will require the elucidation of the epigenetic profile of urine DNA from thousands of patients, an n significantly larger than the few hundred patients needed to validate the 32 marker panel. It will be important to plan the clinical trials accordingly in order to reduce the time and cost associated with marker validation. For the 58 patients with both DRE and FV samples, paired sample analysis showed no significant difference in outcome between the 2 DNA sources. Overall, the methylation of FV DNA slightly outperformed that of DRE DNA at the 10of32 threshold, similar to what was obtained with the 19 markers. This supports its use for early PCA diagnostic tests. The small difference between DRE and FV could be attributed to the relatively small number of samples analyzed. The main advantage of using the FV samples for a screening test is the simplicity of obtaining multiple samples from each patient which reduces the sampling error rate and increases confidence in the diagnostic outcome.
The methylation of DRE DNA outperformed that of FV DNA in identifying patients with higher Gleason grade and larger volume tumors supporting its use for predictive and prognostic tests as well as tests aimed at monitoring disease progression. It is likely that the enrichment of cancer DNA following DRE contributed to the improved recovery of the assays used to determine marker methylation. The composition of the FV urine samples is dependent on the steady state release of tumor DNA into circulation and is subject to degradation and dilution. The FV DNA sampling error rate is expected to be higher than that of the DRE DNA. A larger marker panel, different markers or different assays for current