GENETIC TESTING AND MOLECULAR BIOMARKERS Volume 18, Number 3, 2014 ª Mary Ann Liebert, Inc. Pp. 156–163 DOI: 10.1089/gtmb.2013.0424
Diagnostic and Prognostic Scoring System for Prostate Cancer Using Urine and Plasma Biomarkers Wanlong Ma,1 Kevin Diep,1 Herbert A. Fritsche,2 Neal Shore,3 and Maher Albitar1
Aims: To avoid relying solely on serum prostate-specific antigen (sPSA) in screening for prostate cancer (PCa), we developed a scoring system for detecting PCa and the prediction of aggressiveness. We analyzed urine and plasma specimens from 121 patients with PCa or benign prostatic hyperplasia (BPH) for the levels of UAP1, PDLIM5, IMPDH2, HSPD1, PCA3, PSA, TMPRSS2, ERG, GAPDH, and B2M genes. Patient age, sPSA level, and polymerase chain reaction data were entered through multiple algorithms to determine models most useful for the detection of cancer and predicting aggressiveness. Results: In the first algorithm, we distinguished PCa from BPH (area under the receiver operating characteristic curve [AUROC] of 0.78). Another algorithm distinguished patients with the Gleason score (GS) of ‡ 7 from GS of < 7 cancer or BPH (AUROC of 0.88). By incorporating the two algorithms into a scoring system, 75% of the analyzed samples showed concordance between the two models (99% specificity and 68% sensitivity for predicting GS ‡ 7 in this group). Conclusion: A scoring system incorporating two algorithms using urine and plasma biomarkers highly predicts the presence of GS ‡ 7 PCa in 75% of patients. Our algorithms may assist with both biopsy indication and patient prognosis.
Introduction
P
rostate cancer (PCa) is the most common cancer and the second leading cause of cancer-related death in men in the United States, and its incidence has risen due to increased male longevity (Liu et al., 2012; Ribeiro da Silva et al., 2012; Siegel et al., 2012). Serum prostate-specific antigen (sPSA) measurement is the most current screening method in the United States, and an sPSA level of ‡ 4.0 ng/ mL has been a traditional threshold for a biopsy evaluation (Thompson et al., 2004). An elevated sPSA level may also be characteristic of benign prostatic hyperplasia (BPH) (Shariat et al., 2011). Relying on sPSA level alone leads to a 75% false-positive biopsy rate (Barry, 2001). When relying solely on sPSA testing, patients are potentially unnecessarily sometimes overdiagnosed and thus potentially overtreated, which has resulted in the U.S. Preventive Services Task Force recommendation of ‘‘D’’ (i.e., discouraged) for sPSA as a routine screening test (Moyer and U.S. Preventive Services Task Force, 2012). Gleason grading remains a mainstay prognostic predictor for PCa in clinical practice (Humphrey, 2004). Ample data support interventional strategies for Gleason score (GS) ‡ 7 PCa. However, for patients with lowrisk (GS < 7) PCa, these strategies may not be necessary and surveillance may be better suited (Carter et al., 2012). 1 2 3
Attempts have been made to improve the clinical utility of sPSA testing (Polascik et al., 1999). Free and complex sPSA, as well as isoforms of sPSA used as an adjunct to sPSA, show some improvement in specificity and sensitivity (Zhaohui et al., 2002; Rafi et al., 2003). Limited improvement has been shown with sPSA velocity and doubling time (Vickers and Brewster, 2012). Recently, additional biomarkers to detect PCa have been introduced. PCa gene 3 (PCA3) has demonstrated high prostate specificity (De Kok et al., 2002). PCA3 mRNA as a ratio of PSA mRNA in urine has shown improved diagnostic value (area under the receiver operating characteristic curve [AUROC], 66–72%) compared to PSA alone (AUROC, 54–63%), especially when the tested urine was collected post-digital rectal examination (DRE) (Prensner et al., 2012). TMPRSS2-ERG gene fusions are also prevalent in PCa (St John et al., 2012). In 2011, experimental data from post-DRE urinary analysis of TMPRSS2-ERG and PCA3 were combined with sPSA levels to develop a clinically practical multivariate algorithm that improved PCa prediction. However, fusion mRNA detection requires DRE and almost immediate urine processing, which limits the algorithm’s practical utility (Salami et al., 2013). RNA profiling of tumors has proven to be valuable in predicting clinical behavior of various tumors, including PCa (Cuzick et al., 2011; Cooperberg et al., 2013). Recently,
NeoGenomics Laboratories, Irvine, California. Health Discovery Corporation, Atlanta, Georgia. Carolina Urologic Research Center, Myrtle Beach, South Carolina.
156
PROSTATE BIOMARKERS CANCER SCORING SYSTEM
157
Cooperberg et al. demonstrated significant improvement in the accuracy of risk stratification when a scoring system is used that integrates the mRNA levels of 31 cell cycle progression genes as analyzed in paraffin-embedded PCa tissue. This risk stratification system was useful in men with clinically localized PCa, particularly when the cancer is determined histologically to be of low risk (Cooperberg et al., 2013). Using expression profiling, a 2009 study found that a set of four genes (UAP1, PDLIM5, IMPDH2, and HSPD1) were highly overexpressed in grades 3 and 4 PCa cells compared to noncancer prostate cells and benign prostate tissue (Guyon et al., 2009). In our study, we used urine and peripheral blood plasma to quantify these gene expression levels and those of additional genes known to be expressed in prostate tissue (i.e., PCA3, PSA, TMPRSS2, and ERG) and compared the levels for patients with PCa with those detected in patients with biopsy-proven benign prostate tissue. As a control, we quantified the levels of GAPDH and b-2 microglobulin (B2M). We also used mRNA levels from these genes in multivariate algorithms to develop a system to predict the presence or absence of cancer in patients with BPH and to determine cancer aggressiveness (GS ‡ 7).
TTGCATTCAGAAAGGAGCAGACT-3¢ (forward); 5¢-CAA CTGGTTCTGTAGGGTTCGTTT-3¢ (reverse); and VICTGG AGCAAAGGTGGTAGAMGBNFQ (probe). The primer probe set for HSPD1 produced a PCR product of 64 bp: 5¢AACCTGTGACCACCCCTGAA-3¢ (forward); 5¢-TCTTTGT CTCCGTTTGCAGAAA-3¢ (reverse); VICATTGCACAGGT TGCTACMGBNFQ (probe). The primer probe set for IMPDH2 was designed to encompass exons 10 and 11 and produced a PCR product of 74 bp: 5¢-CCACAGTCATGAT GGGCTCTC-3¢ (forward); 5¢-GGATCCCATCGGAAAAGA AGTA-3¢ (reverse); 6FAMACCACTGAGGCCCCTMGBN FQ (probe). The primer probe set for PSA produced a PCR product of 67 bp: 5¢-CCACTGCATCAGGAACAAAAG-3¢ (forward); 5¢-TGTGTCTTCAGGATGAAACAGG-3¢ (reverse); VICCGTGATCTTGCTGGGTMGBNNFQ (probe). B2M and GAPDH mRNA transcripts were measured as controls and purchased as Pre-Developed TaqMan Assay Reagents (Applied Biosystems). In all assays, an equal amount of plasma was used for RNA extraction, dissolved in equal amount of water, and equal amount of RNA solution was used in each assay. Similarly, for urine, all voided urine was used for concentration into 1 mL and RNA was extracted from total concentrate urine and dissolved in an equal amount of water. Equal amount of RNA solution was used in each assay.
Materials and Methods Study design and patients
We prospectively collected urine and blood samples from 141 men from 4 community urology practices: 61 patients who were biopsy positive for PCa, 60 patients with BPH who were biopsy negative for PCa, and 20 patients who recently underwent prostatectomy. Each site provided the histological GS of tumors for post-prostatectomy patients and those with biopsy-confirmed PCa. Gleason grading was performed according to the new modified system based on the 2005 consensus conference (Epstein et al., 2006). Patients who were receiving any therapy for their PCa or BPH were excluded. Inclusion criteria required the patients to be newly diagnosed and above the age of 18 years. All work was performed with IRB-approved protocol (Western IRP) with consent. Urine and plasma processing
Urine samples were concentrated by centrifugation using Amcion Ultra-15 Centrifugal Filter Units with 3 kDa membrane (Millipore, Billerica, MA). Total nucleic acid was extracted from urine and plasma using the NucliSENS extraction kit (BioMerieux, Durham, NC). Quantitative reverse transcription–polymerase chain reaction
Quantitative reverse transcription (RT)–real-time polymerase chain reaction (PCR) was performed using the RNA Ultrasense One-Step Quantitative RT-PCR System (Applied Biosystems, Foster City, CA) using a ViiA 7 Real-Time PCR System (Applied Biosystems) with the following thermocycler conditions: hold stage at 50C for 15 min, 95C for 2 min, followed by 45 cycles at 95C for 15 s and 60C for 30 s. The PDLIM5, PCA3, TMPRSS2, and ERG primers and probes were purchased as TaqMan Gene Expression Assays with assay IDs of Hs00935062_m1, Hs01371939_g1, Hs01120965_m1, and Hs01554629_m1, respectively (Applied Biosystems). The primer probe set for UAP1 produced a PCR product of 70 bp: 5¢-
Statistical analysis
Spearman’s rank correlation was used for all correlations. The p values are from testing the null hypothesis of zero correlation against the alternative hypothesis of non-zero correlation. Relationships among PCa, biomarkers, and age were investigated using Fisher’s exact test for categorical variables. Each biomarker and age was assessed for association with the presence or absence of PCa with univariate logistic regression. After dividing samples into a training set and a testing set, all biomarkers combined with age were analyzed as independent variables by multivariate logistic regression analysis to predict PCa using the following mathematical algorithms: logistic regression, support vector machine (SVM), least absolute shrinkage and selection operator (LASSO), boosting, bagging, random forest, classification and regression tree (CART), Matt (based on Wilcoxon rank sum and signed rank test), and conditional interference tree (CTREE). With multivariate regression models, sPSA, age, and different biomarker combinations were compared, and worthiness was judged by AUROC. A single model was selected based on the fewest variables yielding the most favorable AUROC. Cross validation was performed using bootstrapping. An algorithm based on the training set was used to validate the selected model first by using the testing set samples and then by using the training and testing set samples in combination. The specificity, sensitivity, positive predict value, and negative predict value were calculated using various cutoff points generated from the combined training and testing sets. The 95% confidence intervals were computed for sensitivity and specificity using binomial distribution. SAS version 9.1.3 (SAS Institute, Cary, NC) was used for most of the statistical analyses. Cutoff points were determined in the training sets based on the best separation and were validated in the testing set.
158
MA ET AL.
Table 1. Characteristics of Study Population Biopsy-confirmed PCa (n = 61) Age, median (range) Race White Hispanic Black Asian Histological Gleason grade No. of patients (%) 3+3 3 + 4/3 + 5 4 + 3/5 + 3 4 + 4/4 + 5/5 + 4/5 + 5 PSA (ng/mL)
BPH (n = 60)
Post-prostatectomy (n = 20)
66 (45–84)
63 (45–84)
67 (50–77)
51 6 3 1
47 10 2 1
17 2 1 0
(83%) (10%) (5%) (2%)
(78%) (17%) (3%) (2%)
p-Value 0.21 0.73
(85%) (10%) (5%) (0%) 0.26
30 13 9 9 5.7
(49%) (21%) (15%) (15%) (1.5–283)
4 11 3 2 0.01
4.4 (0.5–14.1)
(21%) (53%) (16%) (10%) (0–6.0)
< 0.001
PCa, prostate cancer; BPH, benign prostatic hyperplasia; PSA, prostate-specific antigen.
Results Patient characteristics
Patients with biopsy-confirmed PCa and patients with BPH were of similar age (median 66 and 63 years, respectively) ( p = 0.21) and race (Table 1). As expected, sPSA levels were significantly different between these two groups ( p < 0.001), with median 5.7 ng/mL in the cancer group and 4.4 ng/mL in the BPH group (Table 1). Patients in the post-prostatectomy (control) group were of similar age and race, and sPSA levels were also significantly lower (median 0.01 ng/mL) than those for both the cancer and BPH groups. Histological GS was similar for the cancer and post-prostatectomy groups. Univariate comparisons of post-prostatectomy, PCa, and BPH groups
Univariate analysis showed a significant difference ( p < 0.05) between the post-prostatectomy and cancer groups in PDLIM5 ( p = 0.005), UAP1 ( p = 0.001), PCA3 ( p < 0.0001), and TMPRSS ( p = 0.009) in urine, and in HSPD ( p = 0.01), IMPDH2 ( p = 0.003), UAP1 ( p = 0.02), and ERG ( p = 0.02) in plasma. There was a significant difference between the postprostatectomy and BPH groups in HSPD1 ( p = 0.004),
IMPDH2 ( p = 0.002), PDLMI5 ( p = 0.0003), UAP1 ( p = 0.0003), PCA3 ( p < 0.0001), and TMPRSS ( p = 0.0006) in urine, and in HSPD ( p = 0.006), IMPDH2 ( p = 0.002), and UAP1 ( p = 0.03) in plasma. Univariate analysis showed a significant difference between the PCa and BPH and PCa groups only in HSPD1 ( p = 0.05), IMPDH2 ( p = 0.01), and PDLIM5 ( p = 0.05) in urine, and in ERG ( p = 0.0003) in plasma. Multivariate comparison and an algorithm to distinguish PCa from BPH
To distinguish PCa from BPH and use as many relevant variables as possible, we explored the value of mathematical algorithms incorporating all variables. Our training set included 70 patients (35 with biopsy-confirmed PCa and 35 with BPH), and the testing set included 51 patients (26 with biopsyconfirmed PCa and 25 with BPH) (Table 2). Training set data were used (approximately two thirds for model creation and one third for testing) before validation of the model using the testing group. The following variables were used to develop the algorithm: UAP1, PDLIM5, IMPDH2, HSPD1, PCA3, TMPRSS2, ERG, GAPDH, B2M, age, and sPSA level. We used multiple mathematical algorithms based on machine learning (details in the Statistical Analysis section).
Table 2. Multivariate Comparison Using Multiple Algorithms to Predict PCa from BPH in the Testing Set (Based on 100 Iterations Testing) Mathematical algorithm Logistic regression LASSO SVM Boosting Bagging Random forest CART Matta CTREE
Mean AUROC
AUROC standard deviation
Mean error rate
Error standard deviation
0.773 0.726 0.672 0.667 0.643 0.642 0.609 0.586 0.54
0.067 0.072 0.082 0.084 0.089 0.079 0.081 0.061 0.049
0.269 0.322 0.365 0.387 0.392 0.397 0.397 0.415 0.444
0.01 0.01 0.012 0.01 0.012 0.011 0.01 0.008 0.006
a Based on the Wilcoxon rank sum and signed rank test. AUROC, area under the receiver operating characteristic curve; LASSO, least absolute shrinkage and selection operator; SVM, support vector machine; CART, classification and regression tree; CTREE, conditional interference tree.
PROSTATE BIOMARKERS CANCER SCORING SYSTEM
159
FIG. 1. Means and 95% confidence intervals (CIs) of (A) area under the receiver operating characteristic curve (AUROC) and (B) error rate of the various algorithms to distinguish prostate cancer (PCa) from benign prostatic hyperplasia (BPH) using only the training set. Logistic regression produced the best mean AUROC and the least mean error rate. (C) Relative contribution of six variables used in the Boosting algorithm. (D) Determining the cutoff point for distinguishing cancer from BPH using all 121 samples. The solid line represents cutoff points of 0.565; dotted lines at 0.55–0.58. The solid line cutoff point of 0.565 reflects least mean error rate at 0.25. p, plasma; u, urine. Figure 1A and B and Table 2 show that logistic regression produced the algorithm with the best mean AUROC (0.77) and the least mean error rate (0.27). Figure 1C shows the contributions of the six variables used (plasma ERG, sPSA, and urine PCA3, IMPDH2, PDLIM5, and HSPD1). Feature elimination was used to eliminate noncontributory variables.
Similar results were obtained for the testing set (Fig. 2). Logistic regression produced a mean AUROC of 0.78. When all 121 samples were considered and each group was tested 100 times selecting random samples each time, AUROC was 0.70–0.85 (Fig. 1A). The logistic regression algorithm suggested a cutoff point of 0.565 (Fig. 1D), least mean error rate
FIG. 2. Means and 95% CIs of (A) AUROC and (B) error rate of the various algorithms to distinguish PCa from BPH using only the testing set.
160
MA ET AL.
FIG. 3. Means and 95% CIs of (A) AUROC and (B) error rate of the various algorithms to distinguish aggressive PCa (Gleason score [GS] ‡ 7) from the combined BPH and indolent PCa (GS < 7) groups using only the testing set. Logistic regression produced the best mean AUROC and the least mean error rate. (C) Relative contribution of six variables used in the Random Forest algorithm. (D) AUROC curve in distinguishing aggressive PCa (GS ‡ 7) from the combined BPH and indolent PCa (GS < 7) group. of 0.25, and specificity and sensitivity of 88% and 67%, respectively. For all patients, using sPSA alone with a cutoff point of 4.0 ng/mL, the specificity was 62% and sensitivity was 56%. Using an sPSA cutoff point of > 14.1 ng/mL, we obtained 100% specificity, but only 18% sensitivity. Multivariate comparison and an algorithm to determine aggressive PCa
In the modified Gleason grading system, a score of < 7 represents an indolent cancer with a very small risk of mortality; for these patients, the risk of dying 10–15 years after diagnosis is the same regardless of treatment (Humphrey, 2004). Therefore, we combined data of our patients with GS < 7 PCa with those of patients with BPH and explored our biomarkers’ potential to distinguish patients with GS ‡ 7 (aggressive PCa; n = 32) from the others (GS < 7 and BPH; n = 89). The whole data set was partitioned randomly into a training set (n = 69) of 18 patients with GS ‡ 7 PCa and 51 patients with BPH/GS < 7 and a testing set (n = 52) of 14 patients with GS ‡ 7 PCa and 38 patients with BPH/GS < 7 cancer. Mathematical models for this second algorithm were created in the same manner as the algorithm to distinguish PCa
from BPH using the training set. Figure 3A and B and Table 3 show the mean AUROC and the mean error rate for each algorithm. Again, logistic regression showed the most informative model, with mean AUROC of 0.87 in the training set (based on 100 iterations testing) and mean AUROC of
Table 3. Statistical Analysis of Training Set Showing Logistic Regression as Best Algorithm to Distinguish Aggressive PCa (GSa ‡ 7) from BPH/Indolent Cancer (Based on 100 Iterations Testing) Mathematical algorithm Logistic regression LASSO Boosting Random forest Matt Bagging SVM CART CTREE a
GS, Gleason score.
Mean AUROC
AUROC standard deviation
0.828 0.824 0.797 0.738 0.725 0.713 0.699 0.649 0.617
0.094 0.094 0.093 0.107 0.089 0.113 0.105 0.084 0.128
PROSTATE BIOMARKERS CANCER SCORING SYSTEM
161
0.88 in the testing set. When data for all samples were combined and tested, AUROC was 0.88 (Fig. 3D). With this model, four variables were adequate for developing this second algorithm: sPSA, plasma UAP1, ERG, and urine PDLIM5 (Fig. 3C). Based on mean AUROC, we selected 0.61 as a cutoff point, resulting in specificity and sensitivity of 99% and 47%, respectively (Table 4). The number of patients with aggressive cancer was relatively small (n = 32), but mean AUROC of 0.88 was within one standard deviation (SD) (0.73–0.92 – 1 SD) based on 50 iterations testing. Combined algorithm to distinguish aggressive PCa from BPH/indolent cancer
Our two independent models use different variables and different algorithms. When using both models together, 91 of 121 patients (75%) had concordant results (i.e., the first algorithm confirmed the presence or absence of cancer/BPH). There was a higher level of confidence in diagnosing aggressive PCa (99% specificity and 68% sensitivity), indicating a strong predictive value (Table 4 and Fig. 4). The 30 other patients (25%) had discordant results and their cancers should be evaluated using the first algorithm, that is, considered as cancer or no cancer (88% specificity and 67% sensitivity). This algorithm should be considered only to predict the presence or absence of PCa and not to reliably classify the aggressiveness of the disease. Discussion
The search of biomarkers for early detection of PCa has proved difficult and complicated (Liong et al., 2012). Detecting PCa is important, but it is also important to determine the aggressiveness of the cancer. Serum PSA testing—even its variants—is clearly not specific for cancer, cannot distinguish between cancer and BPH (Casadio et al., 2013), and provides limited prognostic information. However, sPSA
Table 4. Comparison of Two Algorithms and the Combined Model Showing Higher Level of Confidence in Diagnosing Aggressive PCa Algorithm type
Estimated value
95% CI lower limit
95% CI upper limit
PCa vs. BPH with cutoff of 0.565 Specificity 0.88 0.77 0.95 Sensitivity 0.67 0.54 0.78 PPV 0.85 0.72 0.93 NPV 0.73 0.61 0.82 Aggressive PCa (GS ‡ 7) vs. BPH/GS < 7 with cutoff of 0.61 Specificity 0.99 0.93 1.00 Sensitivity 0.47 0.30 0.65 PPV 0.94 0.68 1.00 NPV 0.84 0.75 0.90 Combined model for predicting aggressive PCa (GS ‡ 7) vs. BPH/GS < 7 Specificity 0.99 0.91 1.00 Sensitivity 0.68 0.45 0.85 PPV 0.94 0.68 1.00 NPV 0.91 0.81 0.96 CI, confidence interval; PPV, positive predict value; NPV, negative predict value.
FIG. 4. Schematic presentation of individual PCa and BPH patients (depicted as rectangles) with biopsy-confirmed cancer and as predicted two algorithms and by combined model. Red rectangles represent PCa patients as predicted by first algorithm and those with aggressive PCa predicted by second algorithm. Red rectangles in bottom row (combined model) depict patients with aggressive PCa as predicted by the combined algorithms. Patients negative for cancer are shown in green. FP, false positive; FN, false negative; Dis, patients with discordant results with the two algorithms.
protein testing has proved useful when compared with the numerous studied biomarkers (Detchokul and Frauman, 2011). Although testing PCA3:PSA mRNA ratio in urine has been suggested as a replacement for sPSA testing, most data suggest that it is better in predicting the presence of cancer only after DRE (Loeb and Partin, 2011) and does not provide prognostic information (Auprich et al., 2011; Augustin et al., 2013). The fusion TMPRSS2-ERG gene, methylation of GSTP1, EZH2, and DNMT3A2 (Yegnasubramanian et al., 2004; Kobayashi et al., 2011), other biomarkers, and circulating tumor cells (Diamond et al., 2012) also have failed to provide a reliable clinically useful means for predicting cancer in patients with BPH because almost all patients with PCa also have BPH (Orsted and Bojesen, 2013). A combination of biomarkers in urine and plasma may provide a more precise means for early PCa detection in the background of BPH and may help to predict aggressive cancer (GS ‡ 7). Numerous studies of hematologic disease show that the more aggressive the cancer, the more likely its protein or mRNA is in circulation due to higher turnover of aggressive tumor cells when compared with indolent tumor. Products of cell turnover can be detected in plasma and urine as acellular particles or debris. Our study suggests that most of our tested biomarkers (PDLIM5, UAP1, PCA3, TMPRSS2, HSPD1, IMPDH2, and ERG) reflect prostate-specific changes and, more importantly, that both plasma and urine can reflect their specificity. Clear and significant differences in mRNA levels of these biomarkers were detected in the plasma or urine of patients with cancer or BPH compared with the levels for postprostatectomy patients. We believe our selected genes are uniquely suited for the detection of PCa. The UAP1, PDLIM5, IMPDH2, and HSPD1 genes were selected based on comparing expression profile of laser microdissected PCa cells with normal tissue from 87 patients with advanced GS. Recursive feature elimination algorithms (SVM) were used
162
to identify differentially overexpressed genes. These 4 genes were selected from 63 overexpressed genes by the recursive feature elimination algorithm to be the most reliable in distinguishing cancer from noncancer. These genes were validated when used in a mathematical equation to classify tissues as cancer versus noncancer in an independently collected set of formalin-fixed and fresh-frozen prostate tissues. The PDLIM5 gene is involved in the protein kinase C pathway and important in signaling. IMPDH2 is important for guanine nucleotide biosynthesis and apoptosis. HSPD1 is heat shock protein involved in apoptosis. UAP1 is associated with androgen response, male infertility, and cancer (Guyon et al., 2009). However, when we compared BPH and PCa using plasma and urine, borderline significant differences were noted in HSPD1, IMPDH2, and PDLIM5 in urine, but the ERG level was significantly different for BPH ( p = 0.0003). With the exception of plasma ERG level, the differences for biomarkers between BPH and cancer were minimal. This reflects the difficulty in distinguishing between PCa and BPH, which appears to be different from normal prostate tissue. Although these biomarker differences individually may distinguish between BPH and PCa, they were not strong enough for clinical use. Logistic regression showed the best predictive values and the least mean error rates when we tested multiple algorithms using patient age, sPSA, and all possible combinations of biomarkers. The best combination for distinguishing cancer from BPH was plasma ERG, sPSA, and urine PCA3, IMPDH2, PDLIM5, and HSPD1 (AUROC of 0.78). However, when we combined patients with GS < 7 cancer with patients with BPH and used the same biomarkers in various algorithms in the same manner, distinguishing between the two new groups was significantly better (AUROC of 0.88), and the only variables needed were plasma UPA1, plasma ERG, urine PDLIM5, and sPSA protein. This shows that distinguishing patients with more aggressive cancer from patients with indolent PCa/BPH is significantly better because indolent cancer is not that significantly different biologically from BPH. Our result augments the consideration that GS < 7 tumors may be reclassified as not cancer (Carter et al. 2012).9 The two algorithms for predicting the presence or absence of cancer and for distinguishing BPH/indolent cancer from aggressive PCa are fairly independent of each other. Although the number of patients studied is small, concordant results with the two algorithms show a more highly reliable distinction between the presence of aggressive cancer and BPH/indolent cancer. Seventy-five percent of patients showed positive or negative results with both algorithms with very high specificity and sensitivity (99% and 68%, respectively); 25% were discordant and the first algorithm could distinguish cancer from BPH in these patients with specificity and sensitivity of 88% and 67%, respectively. Our preliminary data provide a scoring system incorporating two algorithms to use as a standard for PCa screening, to determine the necessity for a patient to undergo a biopsy, and to determine the accuracy of any biopsy results. Further validation with a larger number of patients is needed before these algorithms are adopted in routine practice. More importantly, an extended follow-up is needed on patients in this study since the model’s specificity and sensitivity are based on biopsy results, which could be falsely negative and represent a change in Gleason grading after prostatectomy.
MA ET AL.
Although more validation of these algorithms is needed, our findings are very encouraging. Developing a clinically useful test to distinguish patients who have aggressive PCa that needs intervention from those who do not need intervention may not only alleviate apprehension about a PCa diagnosis but may also reduce potential harm resulting from unnecessary aggressive treatment of patients who do not need therapy. Furthermore, this test is based on testing plasma and urine, both of which are readily accessible, thus reducing the need for biopsies. Furthermore, the potential to reduce or eliminate unnecessary biopsies and the ability to utilize a more accurate blood/urine-based test may reduce the cost of dealing with prostate problems in the aging population, which increasingly is becoming a serious issue for the society as a whole and the medical community. Acknowledgment
The authors thank Ms. Catherine Coffin for editing and formatting the article. Author Disclosure Statement
W.M., K.D., and M.A. are employed by a commercial diagnostic company that offers testing for prostate cancer. H.A.F. and M.A. are consultant to Health Discovery Corporation, which own a patent on some of the genes used in this study. N.S. has no potential conflict of interest. References
Augustin H, Mayrhofer K, Pummer K, et al. (2013) Relationship between prostate cancer gene 3 (PCA3) and characteristics of tumor aggressiveness. Prostate 73:203–210. Auprich M, Bjartell A, Chun FK, et al. (2011) Contemporary role of prostate cancer antigen 3 in the management of prostate cancer. Eur Urol 60:1045–1054. Barry MJ (2001) Prostate-specific-antigen testing for early diagnosis of prostate cancer. N Engl J Med 344:1373– 1377. Carter HB, Partin AW, Walsh PC, et al. (2012) Gleason Score 6 adenocarcinoma: Should it be labeled as cancer? J Clin Oncol 30:4294–4296. Casadio V, Calistri D, Salvi S, et al. (2013) Urine cell-free DNA integrity as a marker for early prostate cancer diagnosis: a pilot study. Biomed Res Int 2013:270457. Cooperberg MR, Simko JP, Cowan JE, et al. (2013) Validation of a cell-cycle progression gene panel to improve risk stratification in a contemporary prostatectomy cohort. J Clin Oncol 31:1428–1434. Cuzick J, Swanson GP, Fisher G, et al. (2011). Prognostic value of an RNA expression signature derived from cell cycle proliferation genes in patients with prostate cancer: a retrospective study. Lancet Oncol 12:245–255. De Kok JB, Verhaegh GW, Roelofs RW, et al. (2002) DD3(PCA3), a very sensitive and specific marker to detect prostate tumors. Cancer Res 62:2695–2698. Detchokul S, Frauman AG (2011) Recent developments in prostate cancer biomarker research: therapeutic implications. Br J Clin Pharmacol 71:157–174. Diamond E, Lee GY, Akhtar NH, et al. (2012) Isolation and characterization of circulating tumor cells in prostate cancer. Front Oncol 2:131–142. Epstein JI, Allsbrook WC Jr, Amin MB, et al. (2006) Update on the Gleason grading system for prostate cancer: Results of an
PROSTATE BIOMARKERS CANCER SCORING SYSTEM
163
international consensus conference of urologic pathologists. Adv Anat Pathol 13:57–59. Guyon I, Fritsche H, Choppa P, et al. (2009) A four-gene expression signature for prostate cancer cells consisting of UAP1, PDLIM5, IMPDH2, and HSPD1. UroToday Int J 2:3834–3844. Humphrey PA (2004) Gleason grading and prognostic factors in carcinoma of the prostate. Mod Pathol 17:292–306. Kobayashi Y, Absher DM, Gulzar ZG, et al. (2011) DNA methylation profiling reveals novel biomarkers and important roles for DNA methyltransferases in prostate cancer. Genome Res 21:1017–1027. Liong ML, Lim CR, Yang H, et al. (2012) Blood-based biomarkers of aggressive prostate cancer. PLoS One 7:e45802–e45808. Liu Y, Hegde P, Zhang F, et al. (2012) Prostate cancer—a biomarker perspective. Front Endocrinol (Lausanne) 3:72–81. Loeb S, Partin AW (2011) Review of the literature: PCA3 for prostate cancer risk assessment and prognostication. Rev Urol 13:e191–e195. Moyer VA, U.S. Preventive Services Task Force (2012) Screening for prostate cancer: U.S. Preventive Services Task Force recommendation statement. Ann Intern Med 157:120–134. Orsted DD, Bojesen SE (2013) The link between benign prostatic hyperplasia and prostate cancer. Nat Rev Urol 10:49–54. Polascik TJ, Oesterling JE, Partin AW (1999) Prostate specific antigen: a decade of discovery—what we have learned and where we are going. J Urol 162:293–306. Prensner JR, Rubin MA, Wei JT, et al. (2012) Beyond PSA: The next generation of prostate cancer biomarkers. Sci Transl Med 4:127–138.rv3. Rafi T, Sattar A, Asif N, et al. (2003) The comparison of percent free PSA with total PSA in the diagnosis of prostate cancer. JPMA 53–56. Ribeiro da Silva MN, Tobias-Machado M, Lima-Pompeo AC, et al. (2012) Prostate cancer: promising biomarkers related to aggressive disease. Actas Urol Esp 36:484–490.
Salami SS, Schmidt F, Laxman B, et al. (2013) Combining urinary detection of TMPRSS2:ERG and PCA3 with serum PSA to predict diagnosis of prostate cancer. Urol Oncol 31:566–571. Shariat SF, Scherr DS, Gupta A, et al. (2011) Emerging biomarkers for prostate cancer diagnosis, staging, and prognosis. Arch Esp Urol 64:681–694. Siegel R, Naishadham D, Jemal A (2012) Cancer statistics, 2012. CA Cancer J Clin 62:10–29. St John J, Powell K, Conley-Lacomb MK, et al. (2012) TMPRSS2-ERG fusion gene expression in prostate tumor cells and its clinical and biological significance in prostate cancer progression. J Cancer Sci Ther 4:94–101. Thompson IM, Pauler DK, Goodman PJ, et al. (2004) Prevalence of prostate cancer among men with a prostate-specific antigen level < or = 4.0 ng per milliliter. N Engl J Med 350:2239–2246. Vickers AJ, Brewster SF (2012) PSA velocity and doubling time in diagnosis and prognosis of prostate cancer. Br J Med Surg Urol 5:162–168. Yegnasubramanian S, Kowalski J, Gonzalgo ML, et al. (2004) Hypermethylation of CpG islands in primary and metastatic human prostate cancer. Cancer Res 64:1975–1986. Zhaohui C, Jun Z, Jianmin W, et al. (2002) The application of complex PSA and its relative indexes in the detection of prostate cancer. Chinese-German J Clin Oncol 1:233–237.
Address correspondence to: Maher Albitar, MD NeoGenomics Laboratories 5 Jenner, Suite 100 Irvine, CA 92618 E-mail: [email protected]
Diagnostic and Prognostic Scoring System for Prostate Cancer Using Urine and Plasma Biomarkers Wanlong Ma,1 Kevin Diep,1 Herbert A. Fritsche,2 Neal Shore,3 and Maher Albitar1
Aims: To avoid relying solely on serum prostate-specific antigen (sPSA) in screening for prostate cancer (PCa), we developed a scoring system for detecting PCa and the prediction of aggressiveness. We analyzed urine and plasma specimens from 121 patients with PCa or benign prostatic hyperplasia (BPH) for the levels of UAP1, PDLIM5, IMPDH2, HSPD1, PCA3, PSA, TMPRSS2, ERG, GAPDH, and B2M genes. Patient age, sPSA level, and polymerase chain reaction data were entered through multiple algorithms to determine models most useful for the detection of cancer and predicting aggressiveness. Results: In the first algorithm, we distinguished PCa from BPH (area under the receiver operating characteristic curve [AUROC] of 0.78). Another algorithm distinguished patients with the Gleason score (GS) of ‡ 7 from GS of < 7 cancer or BPH (AUROC of 0.88). By incorporating the two algorithms into a scoring system, 75% of the analyzed samples showed concordance between the two models (99% specificity and 68% sensitivity for predicting GS ‡ 7 in this group). Conclusion: A scoring system incorporating two algorithms using urine and plasma biomarkers highly predicts the presence of GS ‡ 7 PCa in 75% of patients. Our algorithms may assist with both biopsy indication and patient prognosis.
Introduction
P
rostate cancer (PCa) is the most common cancer and the second leading cause of cancer-related death in men in the United States, and its incidence has risen due to increased male longevity (Liu et al., 2012; Ribeiro da Silva et al., 2012; Siegel et al., 2012). Serum prostate-specific antigen (sPSA) measurement is the most current screening method in the United States, and an sPSA level of ‡ 4.0 ng/ mL has been a traditional threshold for a biopsy evaluation (Thompson et al., 2004). An elevated sPSA level may also be characteristic of benign prostatic hyperplasia (BPH) (Shariat et al., 2011). Relying on sPSA level alone leads to a 75% false-positive biopsy rate (Barry, 2001). When relying solely on sPSA testing, patients are potentially unnecessarily sometimes overdiagnosed and thus potentially overtreated, which has resulted in the U.S. Preventive Services Task Force recommendation of ‘‘D’’ (i.e., discouraged) for sPSA as a routine screening test (Moyer and U.S. Preventive Services Task Force, 2012). Gleason grading remains a mainstay prognostic predictor for PCa in clinical practice (Humphrey, 2004). Ample data support interventional strategies for Gleason score (GS) ‡ 7 PCa. However, for patients with lowrisk (GS < 7) PCa, these strategies may not be necessary and surveillance may be better suited (Carter et al., 2012). 1 2 3
Attempts have been made to improve the clinical utility of sPSA testing (Polascik et al., 1999). Free and complex sPSA, as well as isoforms of sPSA used as an adjunct to sPSA, show some improvement in specificity and sensitivity (Zhaohui et al., 2002; Rafi et al., 2003). Limited improvement has been shown with sPSA velocity and doubling time (Vickers and Brewster, 2012). Recently, additional biomarkers to detect PCa have been introduced. PCa gene 3 (PCA3) has demonstrated high prostate specificity (De Kok et al., 2002). PCA3 mRNA as a ratio of PSA mRNA in urine has shown improved diagnostic value (area under the receiver operating characteristic curve [AUROC], 66–72%) compared to PSA alone (AUROC, 54–63%), especially when the tested urine was collected post-digital rectal examination (DRE) (Prensner et al., 2012). TMPRSS2-ERG gene fusions are also prevalent in PCa (St John et al., 2012). In 2011, experimental data from post-DRE urinary analysis of TMPRSS2-ERG and PCA3 were combined with sPSA levels to develop a clinically practical multivariate algorithm that improved PCa prediction. However, fusion mRNA detection requires DRE and almost immediate urine processing, which limits the algorithm’s practical utility (Salami et al., 2013). RNA profiling of tumors has proven to be valuable in predicting clinical behavior of various tumors, including PCa (Cuzick et al., 2011; Cooperberg et al., 2013). Recently,
NeoGenomics Laboratories, Irvine, California. Health Discovery Corporation, Atlanta, Georgia. Carolina Urologic Research Center, Myrtle Beach, South Carolina.
156
PROSTATE BIOMARKERS CANCER SCORING SYSTEM
157
Cooperberg et al. demonstrated significant improvement in the accuracy of risk stratification when a scoring system is used that integrates the mRNA levels of 31 cell cycle progression genes as analyzed in paraffin-embedded PCa tissue. This risk stratification system was useful in men with clinically localized PCa, particularly when the cancer is determined histologically to be of low risk (Cooperberg et al., 2013). Using expression profiling, a 2009 study found that a set of four genes (UAP1, PDLIM5, IMPDH2, and HSPD1) were highly overexpressed in grades 3 and 4 PCa cells compared to noncancer prostate cells and benign prostate tissue (Guyon et al., 2009). In our study, we used urine and peripheral blood plasma to quantify these gene expression levels and those of additional genes known to be expressed in prostate tissue (i.e., PCA3, PSA, TMPRSS2, and ERG) and compared the levels for patients with PCa with those detected in patients with biopsy-proven benign prostate tissue. As a control, we quantified the levels of GAPDH and b-2 microglobulin (B2M). We also used mRNA levels from these genes in multivariate algorithms to develop a system to predict the presence or absence of cancer in patients with BPH and to determine cancer aggressiveness (GS ‡ 7).
TTGCATTCAGAAAGGAGCAGACT-3¢ (forward); 5¢-CAA CTGGTTCTGTAGGGTTCGTTT-3¢ (reverse); and VICTGG AGCAAAGGTGGTAGAMGBNFQ (probe). The primer probe set for HSPD1 produced a PCR product of 64 bp: 5¢AACCTGTGACCACCCCTGAA-3¢ (forward); 5¢-TCTTTGT CTCCGTTTGCAGAAA-3¢ (reverse); VICATTGCACAGGT TGCTACMGBNFQ (probe). The primer probe set for IMPDH2 was designed to encompass exons 10 and 11 and produced a PCR product of 74 bp: 5¢-CCACAGTCATGAT GGGCTCTC-3¢ (forward); 5¢-GGATCCCATCGGAAAAGA AGTA-3¢ (reverse); 6FAMACCACTGAGGCCCCTMGBN FQ (probe). The primer probe set for PSA produced a PCR product of 67 bp: 5¢-CCACTGCATCAGGAACAAAAG-3¢ (forward); 5¢-TGTGTCTTCAGGATGAAACAGG-3¢ (reverse); VICCGTGATCTTGCTGGGTMGBNNFQ (probe). B2M and GAPDH mRNA transcripts were measured as controls and purchased as Pre-Developed TaqMan Assay Reagents (Applied Biosystems). In all assays, an equal amount of plasma was used for RNA extraction, dissolved in equal amount of water, and equal amount of RNA solution was used in each assay. Similarly, for urine, all voided urine was used for concentration into 1 mL and RNA was extracted from total concentrate urine and dissolved in an equal amount of water. Equal amount of RNA solution was used in each assay.
Materials and Methods Study design and patients
We prospectively collected urine and blood samples from 141 men from 4 community urology practices: 61 patients who were biopsy positive for PCa, 60 patients with BPH who were biopsy negative for PCa, and 20 patients who recently underwent prostatectomy. Each site provided the histological GS of tumors for post-prostatectomy patients and those with biopsy-confirmed PCa. Gleason grading was performed according to the new modified system based on the 2005 consensus conference (Epstein et al., 2006). Patients who were receiving any therapy for their PCa or BPH were excluded. Inclusion criteria required the patients to be newly diagnosed and above the age of 18 years. All work was performed with IRB-approved protocol (Western IRP) with consent. Urine and plasma processing
Urine samples were concentrated by centrifugation using Amcion Ultra-15 Centrifugal Filter Units with 3 kDa membrane (Millipore, Billerica, MA). Total nucleic acid was extracted from urine and plasma using the NucliSENS extraction kit (BioMerieux, Durham, NC). Quantitative reverse transcription–polymerase chain reaction
Quantitative reverse transcription (RT)–real-time polymerase chain reaction (PCR) was performed using the RNA Ultrasense One-Step Quantitative RT-PCR System (Applied Biosystems, Foster City, CA) using a ViiA 7 Real-Time PCR System (Applied Biosystems) with the following thermocycler conditions: hold stage at 50C for 15 min, 95C for 2 min, followed by 45 cycles at 95C for 15 s and 60C for 30 s. The PDLIM5, PCA3, TMPRSS2, and ERG primers and probes were purchased as TaqMan Gene Expression Assays with assay IDs of Hs00935062_m1, Hs01371939_g1, Hs01120965_m1, and Hs01554629_m1, respectively (Applied Biosystems). The primer probe set for UAP1 produced a PCR product of 70 bp: 5¢-
Statistical analysis
Spearman’s rank correlation was used for all correlations. The p values are from testing the null hypothesis of zero correlation against the alternative hypothesis of non-zero correlation. Relationships among PCa, biomarkers, and age were investigated using Fisher’s exact test for categorical variables. Each biomarker and age was assessed for association with the presence or absence of PCa with univariate logistic regression. After dividing samples into a training set and a testing set, all biomarkers combined with age were analyzed as independent variables by multivariate logistic regression analysis to predict PCa using the following mathematical algorithms: logistic regression, support vector machine (SVM), least absolute shrinkage and selection operator (LASSO), boosting, bagging, random forest, classification and regression tree (CART), Matt (based on Wilcoxon rank sum and signed rank test), and conditional interference tree (CTREE). With multivariate regression models, sPSA, age, and different biomarker combinations were compared, and worthiness was judged by AUROC. A single model was selected based on the fewest variables yielding the most favorable AUROC. Cross validation was performed using bootstrapping. An algorithm based on the training set was used to validate the selected model first by using the testing set samples and then by using the training and testing set samples in combination. The specificity, sensitivity, positive predict value, and negative predict value were calculated using various cutoff points generated from the combined training and testing sets. The 95% confidence intervals were computed for sensitivity and specificity using binomial distribution. SAS version 9.1.3 (SAS Institute, Cary, NC) was used for most of the statistical analyses. Cutoff points were determined in the training sets based on the best separation and were validated in the testing set.
158
MA ET AL.
Table 1. Characteristics of Study Population Biopsy-confirmed PCa (n = 61) Age, median (range) Race White Hispanic Black Asian Histological Gleason grade No. of patients (%) 3+3 3 + 4/3 + 5 4 + 3/5 + 3 4 + 4/4 + 5/5 + 4/5 + 5 PSA (ng/mL)
BPH (n = 60)
Post-prostatectomy (n = 20)
66 (45–84)
63 (45–84)
67 (50–77)
51 6 3 1
47 10 2 1
17 2 1 0
(83%) (10%) (5%) (2%)
(78%) (17%) (3%) (2%)
p-Value 0.21 0.73
(85%) (10%) (5%) (0%) 0.26
30 13 9 9 5.7
(49%) (21%) (15%) (15%) (1.5–283)
4 11 3 2 0.01
4.4 (0.5–14.1)
(21%) (53%) (16%) (10%) (0–6.0)
< 0.001
PCa, prostate cancer; BPH, benign prostatic hyperplasia; PSA, prostate-specific antigen.
Results Patient characteristics
Patients with biopsy-confirmed PCa and patients with BPH were of similar age (median 66 and 63 years, respectively) ( p = 0.21) and race (Table 1). As expected, sPSA levels were significantly different between these two groups ( p < 0.001), with median 5.7 ng/mL in the cancer group and 4.4 ng/mL in the BPH group (Table 1). Patients in the post-prostatectomy (control) group were of similar age and race, and sPSA levels were also significantly lower (median 0.01 ng/mL) than those for both the cancer and BPH groups. Histological GS was similar for the cancer and post-prostatectomy groups. Univariate comparisons of post-prostatectomy, PCa, and BPH groups
Univariate analysis showed a significant difference ( p < 0.05) between the post-prostatectomy and cancer groups in PDLIM5 ( p = 0.005), UAP1 ( p = 0.001), PCA3 ( p < 0.0001), and TMPRSS ( p = 0.009) in urine, and in HSPD ( p = 0.01), IMPDH2 ( p = 0.003), UAP1 ( p = 0.02), and ERG ( p = 0.02) in plasma. There was a significant difference between the postprostatectomy and BPH groups in HSPD1 ( p = 0.004),
IMPDH2 ( p = 0.002), PDLMI5 ( p = 0.0003), UAP1 ( p = 0.0003), PCA3 ( p < 0.0001), and TMPRSS ( p = 0.0006) in urine, and in HSPD ( p = 0.006), IMPDH2 ( p = 0.002), and UAP1 ( p = 0.03) in plasma. Univariate analysis showed a significant difference between the PCa and BPH and PCa groups only in HSPD1 ( p = 0.05), IMPDH2 ( p = 0.01), and PDLIM5 ( p = 0.05) in urine, and in ERG ( p = 0.0003) in plasma. Multivariate comparison and an algorithm to distinguish PCa from BPH
To distinguish PCa from BPH and use as many relevant variables as possible, we explored the value of mathematical algorithms incorporating all variables. Our training set included 70 patients (35 with biopsy-confirmed PCa and 35 with BPH), and the testing set included 51 patients (26 with biopsyconfirmed PCa and 25 with BPH) (Table 2). Training set data were used (approximately two thirds for model creation and one third for testing) before validation of the model using the testing group. The following variables were used to develop the algorithm: UAP1, PDLIM5, IMPDH2, HSPD1, PCA3, TMPRSS2, ERG, GAPDH, B2M, age, and sPSA level. We used multiple mathematical algorithms based on machine learning (details in the Statistical Analysis section).
Table 2. Multivariate Comparison Using Multiple Algorithms to Predict PCa from BPH in the Testing Set (Based on 100 Iterations Testing) Mathematical algorithm Logistic regression LASSO SVM Boosting Bagging Random forest CART Matta CTREE
Mean AUROC
AUROC standard deviation
Mean error rate
Error standard deviation
0.773 0.726 0.672 0.667 0.643 0.642 0.609 0.586 0.54
0.067 0.072 0.082 0.084 0.089 0.079 0.081 0.061 0.049
0.269 0.322 0.365 0.387 0.392 0.397 0.397 0.415 0.444
0.01 0.01 0.012 0.01 0.012 0.011 0.01 0.008 0.006
a Based on the Wilcoxon rank sum and signed rank test. AUROC, area under the receiver operating characteristic curve; LASSO, least absolute shrinkage and selection operator; SVM, support vector machine; CART, classification and regression tree; CTREE, conditional interference tree.
PROSTATE BIOMARKERS CANCER SCORING SYSTEM
159
FIG. 1. Means and 95% confidence intervals (CIs) of (A) area under the receiver operating characteristic curve (AUROC) and (B) error rate of the various algorithms to distinguish prostate cancer (PCa) from benign prostatic hyperplasia (BPH) using only the training set. Logistic regression produced the best mean AUROC and the least mean error rate. (C) Relative contribution of six variables used in the Boosting algorithm. (D) Determining the cutoff point for distinguishing cancer from BPH using all 121 samples. The solid line represents cutoff points of 0.565; dotted lines at 0.55–0.58. The solid line cutoff point of 0.565 reflects least mean error rate at 0.25. p, plasma; u, urine. Figure 1A and B and Table 2 show that logistic regression produced the algorithm with the best mean AUROC (0.77) and the least mean error rate (0.27). Figure 1C shows the contributions of the six variables used (plasma ERG, sPSA, and urine PCA3, IMPDH2, PDLIM5, and HSPD1). Feature elimination was used to eliminate noncontributory variables.
Similar results were obtained for the testing set (Fig. 2). Logistic regression produced a mean AUROC of 0.78. When all 121 samples were considered and each group was tested 100 times selecting random samples each time, AUROC was 0.70–0.85 (Fig. 1A). The logistic regression algorithm suggested a cutoff point of 0.565 (Fig. 1D), least mean error rate
FIG. 2. Means and 95% CIs of (A) AUROC and (B) error rate of the various algorithms to distinguish PCa from BPH using only the testing set.
160
MA ET AL.
FIG. 3. Means and 95% CIs of (A) AUROC and (B) error rate of the various algorithms to distinguish aggressive PCa (Gleason score [GS] ‡ 7) from the combined BPH and indolent PCa (GS < 7) groups using only the testing set. Logistic regression produced the best mean AUROC and the least mean error rate. (C) Relative contribution of six variables used in the Random Forest algorithm. (D) AUROC curve in distinguishing aggressive PCa (GS ‡ 7) from the combined BPH and indolent PCa (GS < 7) group. of 0.25, and specificity and sensitivity of 88% and 67%, respectively. For all patients, using sPSA alone with a cutoff point of 4.0 ng/mL, the specificity was 62% and sensitivity was 56%. Using an sPSA cutoff point of > 14.1 ng/mL, we obtained 100% specificity, but only 18% sensitivity. Multivariate comparison and an algorithm to determine aggressive PCa
In the modified Gleason grading system, a score of < 7 represents an indolent cancer with a very small risk of mortality; for these patients, the risk of dying 10–15 years after diagnosis is the same regardless of treatment (Humphrey, 2004). Therefore, we combined data of our patients with GS < 7 PCa with those of patients with BPH and explored our biomarkers’ potential to distinguish patients with GS ‡ 7 (aggressive PCa; n = 32) from the others (GS < 7 and BPH; n = 89). The whole data set was partitioned randomly into a training set (n = 69) of 18 patients with GS ‡ 7 PCa and 51 patients with BPH/GS < 7 and a testing set (n = 52) of 14 patients with GS ‡ 7 PCa and 38 patients with BPH/GS < 7 cancer. Mathematical models for this second algorithm were created in the same manner as the algorithm to distinguish PCa
from BPH using the training set. Figure 3A and B and Table 3 show the mean AUROC and the mean error rate for each algorithm. Again, logistic regression showed the most informative model, with mean AUROC of 0.87 in the training set (based on 100 iterations testing) and mean AUROC of
Table 3. Statistical Analysis of Training Set Showing Logistic Regression as Best Algorithm to Distinguish Aggressive PCa (GSa ‡ 7) from BPH/Indolent Cancer (Based on 100 Iterations Testing) Mathematical algorithm Logistic regression LASSO Boosting Random forest Matt Bagging SVM CART CTREE a
GS, Gleason score.
Mean AUROC
AUROC standard deviation
0.828 0.824 0.797 0.738 0.725 0.713 0.699 0.649 0.617
0.094 0.094 0.093 0.107 0.089 0.113 0.105 0.084 0.128
PROSTATE BIOMARKERS CANCER SCORING SYSTEM
161
0.88 in the testing set. When data for all samples were combined and tested, AUROC was 0.88 (Fig. 3D). With this model, four variables were adequate for developing this second algorithm: sPSA, plasma UAP1, ERG, and urine PDLIM5 (Fig. 3C). Based on mean AUROC, we selected 0.61 as a cutoff point, resulting in specificity and sensitivity of 99% and 47%, respectively (Table 4). The number of patients with aggressive cancer was relatively small (n = 32), but mean AUROC of 0.88 was within one standard deviation (SD) (0.73–0.92 – 1 SD) based on 50 iterations testing. Combined algorithm to distinguish aggressive PCa from BPH/indolent cancer
Our two independent models use different variables and different algorithms. When using both models together, 91 of 121 patients (75%) had concordant results (i.e., the first algorithm confirmed the presence or absence of cancer/BPH). There was a higher level of confidence in diagnosing aggressive PCa (99% specificity and 68% sensitivity), indicating a strong predictive value (Table 4 and Fig. 4). The 30 other patients (25%) had discordant results and their cancers should be evaluated using the first algorithm, that is, considered as cancer or no cancer (88% specificity and 67% sensitivity). This algorithm should be considered only to predict the presence or absence of PCa and not to reliably classify the aggressiveness of the disease. Discussion
The search of biomarkers for early detection of PCa has proved difficult and complicated (Liong et al., 2012). Detecting PCa is important, but it is also important to determine the aggressiveness of the cancer. Serum PSA testing—even its variants—is clearly not specific for cancer, cannot distinguish between cancer and BPH (Casadio et al., 2013), and provides limited prognostic information. However, sPSA
Table 4. Comparison of Two Algorithms and the Combined Model Showing Higher Level of Confidence in Diagnosing Aggressive PCa Algorithm type
Estimated value
95% CI lower limit
95% CI upper limit
PCa vs. BPH with cutoff of 0.565 Specificity 0.88 0.77 0.95 Sensitivity 0.67 0.54 0.78 PPV 0.85 0.72 0.93 NPV 0.73 0.61 0.82 Aggressive PCa (GS ‡ 7) vs. BPH/GS < 7 with cutoff of 0.61 Specificity 0.99 0.93 1.00 Sensitivity 0.47 0.30 0.65 PPV 0.94 0.68 1.00 NPV 0.84 0.75 0.90 Combined model for predicting aggressive PCa (GS ‡ 7) vs. BPH/GS < 7 Specificity 0.99 0.91 1.00 Sensitivity 0.68 0.45 0.85 PPV 0.94 0.68 1.00 NPV 0.91 0.81 0.96 CI, confidence interval; PPV, positive predict value; NPV, negative predict value.
FIG. 4. Schematic presentation of individual PCa and BPH patients (depicted as rectangles) with biopsy-confirmed cancer and as predicted two algorithms and by combined model. Red rectangles represent PCa patients as predicted by first algorithm and those with aggressive PCa predicted by second algorithm. Red rectangles in bottom row (combined model) depict patients with aggressive PCa as predicted by the combined algorithms. Patients negative for cancer are shown in green. FP, false positive; FN, false negative; Dis, patients with discordant results with the two algorithms.
protein testing has proved useful when compared with the numerous studied biomarkers (Detchokul and Frauman, 2011). Although testing PCA3:PSA mRNA ratio in urine has been suggested as a replacement for sPSA testing, most data suggest that it is better in predicting the presence of cancer only after DRE (Loeb and Partin, 2011) and does not provide prognostic information (Auprich et al., 2011; Augustin et al., 2013). The fusion TMPRSS2-ERG gene, methylation of GSTP1, EZH2, and DNMT3A2 (Yegnasubramanian et al., 2004; Kobayashi et al., 2011), other biomarkers, and circulating tumor cells (Diamond et al., 2012) also have failed to provide a reliable clinically useful means for predicting cancer in patients with BPH because almost all patients with PCa also have BPH (Orsted and Bojesen, 2013). A combination of biomarkers in urine and plasma may provide a more precise means for early PCa detection in the background of BPH and may help to predict aggressive cancer (GS ‡ 7). Numerous studies of hematologic disease show that the more aggressive the cancer, the more likely its protein or mRNA is in circulation due to higher turnover of aggressive tumor cells when compared with indolent tumor. Products of cell turnover can be detected in plasma and urine as acellular particles or debris. Our study suggests that most of our tested biomarkers (PDLIM5, UAP1, PCA3, TMPRSS2, HSPD1, IMPDH2, and ERG) reflect prostate-specific changes and, more importantly, that both plasma and urine can reflect their specificity. Clear and significant differences in mRNA levels of these biomarkers were detected in the plasma or urine of patients with cancer or BPH compared with the levels for postprostatectomy patients. We believe our selected genes are uniquely suited for the detection of PCa. The UAP1, PDLIM5, IMPDH2, and HSPD1 genes were selected based on comparing expression profile of laser microdissected PCa cells with normal tissue from 87 patients with advanced GS. Recursive feature elimination algorithms (SVM) were used
162
to identify differentially overexpressed genes. These 4 genes were selected from 63 overexpressed genes by the recursive feature elimination algorithm to be the most reliable in distinguishing cancer from noncancer. These genes were validated when used in a mathematical equation to classify tissues as cancer versus noncancer in an independently collected set of formalin-fixed and fresh-frozen prostate tissues. The PDLIM5 gene is involved in the protein kinase C pathway and important in signaling. IMPDH2 is important for guanine nucleotide biosynthesis and apoptosis. HSPD1 is heat shock protein involved in apoptosis. UAP1 is associated with androgen response, male infertility, and cancer (Guyon et al., 2009). However, when we compared BPH and PCa using plasma and urine, borderline significant differences were noted in HSPD1, IMPDH2, and PDLIM5 in urine, but the ERG level was significantly different for BPH ( p = 0.0003). With the exception of plasma ERG level, the differences for biomarkers between BPH and cancer were minimal. This reflects the difficulty in distinguishing between PCa and BPH, which appears to be different from normal prostate tissue. Although these biomarker differences individually may distinguish between BPH and PCa, they were not strong enough for clinical use. Logistic regression showed the best predictive values and the least mean error rates when we tested multiple algorithms using patient age, sPSA, and all possible combinations of biomarkers. The best combination for distinguishing cancer from BPH was plasma ERG, sPSA, and urine PCA3, IMPDH2, PDLIM5, and HSPD1 (AUROC of 0.78). However, when we combined patients with GS < 7 cancer with patients with BPH and used the same biomarkers in various algorithms in the same manner, distinguishing between the two new groups was significantly better (AUROC of 0.88), and the only variables needed were plasma UPA1, plasma ERG, urine PDLIM5, and sPSA protein. This shows that distinguishing patients with more aggressive cancer from patients with indolent PCa/BPH is significantly better because indolent cancer is not that significantly different biologically from BPH. Our result augments the consideration that GS < 7 tumors may be reclassified as not cancer (Carter et al. 2012).9 The two algorithms for predicting the presence or absence of cancer and for distinguishing BPH/indolent cancer from aggressive PCa are fairly independent of each other. Although the number of patients studied is small, concordant results with the two algorithms show a more highly reliable distinction between the presence of aggressive cancer and BPH/indolent cancer. Seventy-five percent of patients showed positive or negative results with both algorithms with very high specificity and sensitivity (99% and 68%, respectively); 25% were discordant and the first algorithm could distinguish cancer from BPH in these patients with specificity and sensitivity of 88% and 67%, respectively. Our preliminary data provide a scoring system incorporating two algorithms to use as a standard for PCa screening, to determine the necessity for a patient to undergo a biopsy, and to determine the accuracy of any biopsy results. Further validation with a larger number of patients is needed before these algorithms are adopted in routine practice. More importantly, an extended follow-up is needed on patients in this study since the model’s specificity and sensitivity are based on biopsy results, which could be falsely negative and represent a change in Gleason grading after prostatectomy.
MA ET AL.
Although more validation of these algorithms is needed, our findings are very encouraging. Developing a clinically useful test to distinguish patients who have aggressive PCa that needs intervention from those who do not need intervention may not only alleviate apprehension about a PCa diagnosis but may also reduce potential harm resulting from unnecessary aggressive treatment of patients who do not need therapy. Furthermore, this test is based on testing plasma and urine, both of which are readily accessible, thus reducing the need for biopsies. Furthermore, the potential to reduce or eliminate unnecessary biopsies and the ability to utilize a more accurate blood/urine-based test may reduce the cost of dealing with prostate problems in the aging population, which increasingly is becoming a serious issue for the society as a whole and the medical community. Acknowledgment
The authors thank Ms. Catherine Coffin for editing and formatting the article. Author Disclosure Statement
W.M., K.D., and M.A. are employed by a commercial diagnostic company that offers testing for prostate cancer. H.A.F. and M.A. are consultant to Health Discovery Corporation, which own a patent on some of the genes used in this study. N.S. has no potential conflict of interest. References
Augustin H, Mayrhofer K, Pummer K, et al. (2013) Relationship between prostate cancer gene 3 (PCA3) and characteristics of tumor aggressiveness. Prostate 73:203–210. Auprich M, Bjartell A, Chun FK, et al. (2011) Contemporary role of prostate cancer antigen 3 in the management of prostate cancer. Eur Urol 60:1045–1054. Barry MJ (2001) Prostate-specific-antigen testing for early diagnosis of prostate cancer. N Engl J Med 344:1373– 1377. Carter HB, Partin AW, Walsh PC, et al. (2012) Gleason Score 6 adenocarcinoma: Should it be labeled as cancer? J Clin Oncol 30:4294–4296. Casadio V, Calistri D, Salvi S, et al. (2013) Urine cell-free DNA integrity as a marker for early prostate cancer diagnosis: a pilot study. Biomed Res Int 2013:270457. Cooperberg MR, Simko JP, Cowan JE, et al. (2013) Validation of a cell-cycle progression gene panel to improve risk stratification in a contemporary prostatectomy cohort. J Clin Oncol 31:1428–1434. Cuzick J, Swanson GP, Fisher G, et al. (2011). Prognostic value of an RNA expression signature derived from cell cycle proliferation genes in patients with prostate cancer: a retrospective study. Lancet Oncol 12:245–255. De Kok JB, Verhaegh GW, Roelofs RW, et al. (2002) DD3(PCA3), a very sensitive and specific marker to detect prostate tumors. Cancer Res 62:2695–2698. Detchokul S, Frauman AG (2011) Recent developments in prostate cancer biomarker research: therapeutic implications. Br J Clin Pharmacol 71:157–174. Diamond E, Lee GY, Akhtar NH, et al. (2012) Isolation and characterization of circulating tumor cells in prostate cancer. Front Oncol 2:131–142. Epstein JI, Allsbrook WC Jr, Amin MB, et al. (2006) Update on the Gleason grading system for prostate cancer: Results of an
PROSTATE BIOMARKERS CANCER SCORING SYSTEM
163
international consensus conference of urologic pathologists. Adv Anat Pathol 13:57–59. Guyon I, Fritsche H, Choppa P, et al. (2009) A four-gene expression signature for prostate cancer cells consisting of UAP1, PDLIM5, IMPDH2, and HSPD1. UroToday Int J 2:3834–3844. Humphrey PA (2004) Gleason grading and prognostic factors in carcinoma of the prostate. Mod Pathol 17:292–306. Kobayashi Y, Absher DM, Gulzar ZG, et al. (2011) DNA methylation profiling reveals novel biomarkers and important roles for DNA methyltransferases in prostate cancer. Genome Res 21:1017–1027. Liong ML, Lim CR, Yang H, et al. (2012) Blood-based biomarkers of aggressive prostate cancer. PLoS One 7:e45802–e45808. Liu Y, Hegde P, Zhang F, et al. (2012) Prostate cancer—a biomarker perspective. Front Endocrinol (Lausanne) 3:72–81. Loeb S, Partin AW (2011) Review of the literature: PCA3 for prostate cancer risk assessment and prognostication. Rev Urol 13:e191–e195. Moyer VA, U.S. Preventive Services Task Force (2012) Screening for prostate cancer: U.S. Preventive Services Task Force recommendation statement. Ann Intern Med 157:120–134. Orsted DD, Bojesen SE (2013) The link between benign prostatic hyperplasia and prostate cancer. Nat Rev Urol 10:49–54. Polascik TJ, Oesterling JE, Partin AW (1999) Prostate specific antigen: a decade of discovery—what we have learned and where we are going. J Urol 162:293–306. Prensner JR, Rubin MA, Wei JT, et al. (2012) Beyond PSA: The next generation of prostate cancer biomarkers. Sci Transl Med 4:127–138.rv3. Rafi T, Sattar A, Asif N, et al. (2003) The comparison of percent free PSA with total PSA in the diagnosis of prostate cancer. JPMA 53–56. Ribeiro da Silva MN, Tobias-Machado M, Lima-Pompeo AC, et al. (2012) Prostate cancer: promising biomarkers related to aggressive disease. Actas Urol Esp 36:484–490.
Salami SS, Schmidt F, Laxman B, et al. (2013) Combining urinary detection of TMPRSS2:ERG and PCA3 with serum PSA to predict diagnosis of prostate cancer. Urol Oncol 31:566–571. Shariat SF, Scherr DS, Gupta A, et al. (2011) Emerging biomarkers for prostate cancer diagnosis, staging, and prognosis. Arch Esp Urol 64:681–694. Siegel R, Naishadham D, Jemal A (2012) Cancer statistics, 2012. CA Cancer J Clin 62:10–29. St John J, Powell K, Conley-Lacomb MK, et al. (2012) TMPRSS2-ERG fusion gene expression in prostate tumor cells and its clinical and biological significance in prostate cancer progression. J Cancer Sci Ther 4:94–101. Thompson IM, Pauler DK, Goodman PJ, et al. (2004) Prevalence of prostate cancer among men with a prostate-specific antigen level < or = 4.0 ng per milliliter. N Engl J Med 350:2239–2246. Vickers AJ, Brewster SF (2012) PSA velocity and doubling time in diagnosis and prognosis of prostate cancer. Br J Med Surg Urol 5:162–168. Yegnasubramanian S, Kowalski J, Gonzalgo ML, et al. (2004) Hypermethylation of CpG islands in primary and metastatic human prostate cancer. Cancer Res 64:1975–1986. Zhaohui C, Jun Z, Jianmin W, et al. (2002) The application of complex PSA and its relative indexes in the detection of prostate cancer. Chinese-German J Clin Oncol 1:233–237.
Address correspondence to: Maher Albitar, MD NeoGenomics Laboratories 5 Jenner, Suite 100 Irvine, CA 92618 E-mail: [email protected]
