European Urology

1. Introduction

Development of biomarkers by genomic and proteomic high-throughput platforms has garnered great expectations of improving cancer screening, early detection, staging, and prognosis. Recently, the urinary prostate cancer gene 3 (PCA3) assay has shown promising results for prostate cancer (PCa) detection. This assay measures PCA3–messenger ribonucleic acid (mRNA) and prostate-specific antigen (PSA)–mRNA concentrations in post–digital rectal examination (post-DRE) urine [1]. PCA3 is highly overexpressed (median: 66-fold) in malignant prostate tissue compared with benign and normal tissues [2]. Several studies demonstrated superior sensitivity and specificity of the PCA3 assay score over that of PSA level [3], [4], and [5]. These findings could translate into improved identification of men at risk of harboring PCa and reduction in the number of unnecessary biopsies. Consequently, a urinary PCA3 assay was developed and was made available for clinical use as a Conformité Européenne (CE)–marked product [1].

Beyond univariate and highly discriminatory ability, improvement of sensitivity and specificity, and confirmation of its independent predictor status, it is mandatory for a novel marker to increase the combined multivariate predictive accuracy (PA) of established risk factors. Furthermore, the increase in multivariate PA should not only be statistically significant but should also address a significant number of individuals. If these criteria are met, the novel marker may be considered clinically meaningful, and its application in clinical practice can be justified [6], and [7]. PCA3 has never been tested in a multivariate biopsy nomogram setting. To address this void, we tested several cut-off thresholds of urinary PCA3 assay scores in the largest reported PCA3 biopsy data set to date; we applied stringent analytic methods in addition to testing the multivariate independent status of PCA3; and we quantified the increment in PA related to its inclusion to established risk factors in risk models for biopsy outcome.

2. Materials and methods

3. Results

Patient characteristics are shown in Table 1. PCa was detected in 316 men (39.1%). Median PCA3 score was 25.9 (0.2–366.9). Mean and median PCA3 scores were significantly higher in the PCa versus the biopsy negative group (56.5 and 37.4 vs 34.6 and 19.5, respectively) (p < 0.001). Mean and median ages were 65 yr and 66 yr, respectively (p > 0.05). PSA levels ranged from 0.1 ng/ml to 48.5 ng/ml (mean: 7.4; median: 6.3). Most patients exhibited a normal DRE (n = 586, 72.4%) and a history of previous biopsy sessions (n = 569, 70.3%). High PSA level, suspicious DRE results, low prostate volume, previous biopsy sessions, and high PCA3 assay score were significantly associated with PCa at biopsy (all p < 0.001).

In the PCA3 assay score cut-off analysis for PA, a score ≤17 or >17 was identified as the most statistically significant cut-off (Table 1). In all subsequent analyses, a continuously coded PCA3 assay score as well as PCA3 cut-off thresholds of 17, 24, and 35 were explored. Table 2 shows the univariate and multivariate LRMs predicting PCa at biopsy. Predictor variables were age, DRE result, PSA level, prostate volume, history of previous biopsy and/or biopsies, and PCA3 assay score. In univariate analyses (Table 2a), all variables were statistically significant predictors of PCa at biopsy (p ≤ 0.02). In univariate analyses, PA estimates for age, DRE, PSA, prostate volume, and history of previous biopsy were 0.582, 0.576, 0.527, 0.619, and 0.523, respectively. Regardless of its PCA3 coding, PCA3 demonstrated the highest PA: The continuously coded PCA3 demonstrated the highest PA (0.679) followed by the PCA3 assay score cut-off threshold 17 (PCA3-17; 0.633), PCA3-24 (0.624) and PCA3-35 (0.620). The highest odds ratio of 3.6 (95% confidence interval [CI]: 2.58–5.02) was recorded for PCA3-17. In multivariate models (Table 2b), all variables were independent predictors of PCa (p ≤ 0.02). The combined multivariate PA of the base model was 0.679. Addition of PCA3, regardless of its coding, improved bootstrap-adjusted multivariate PA of the base model from 2.3% to 4.6%. The highest increment in PA resulted from inclusion of the PCA3-17, for which a statistically significant 4.6% gain in PA, from 0.679 to 0.725 (p = 0.04), was recorded.

Fig. 1a shows the nomogram of regression coefficient–based PCA3-17 demonstrating significantly higher accuracy compared with the nomogram devised from base predictor variables. Fig. 1b and c show calibration plots for the PCA3-17 nomogram and the base nomogram. On these calibration plots, the predicted probability of the nomogram is represented on the x-axis and the observed rate of biopsy-proven PCa is represented on the y-axis; close agreement with the 45° line indicates near-perfect prediction. Specifically, in the low- and high-risk prediction ranges, the performance of the PCA3-17 nomogram exhibits virtually perfect agreement, whereas the base model underestimates the risk of PCa.

Create Platinum Slide Series

Fig. 1 (a) Prostate cancer gene 3 (PCA3) nomogram predicting cancer on prostate biopsy; (b) local regression nonparametric smoothing plots showing the calibration of the PCA3 nomogram; (c) local regression nonparametric smoothing plots showing the calibration of the base nomogram. Instructions for physicians: To obtain nomogram-predicted probability of prostate cancer, locate patient values at each axis. Draw a vertical line to the “Point” axis to determine how many points are attributed for each variable value. Sum the points for all variables. Locate the sum on the “Total Points” line to be able to assess the individual probability of cancer on prostate biopsy on the “Probability of prostate cancer at biopsy” line. Instructions for readers: Perfect predictions correspond to the 45° line. Points estimated below the 45° line correspond to nomogram overprediction, whereas points situated above the 45° line correspond to nomogram underprediction. A nonparametric, smoothed curve indicates the relationship between predicted probability and observed frequency of prostate cancer on initial biopsy. Vertical lines indicate the frequency distribution of predicted probabilities. PSA = prostate-specific antigen; DRE = digital rectal examination.

Table 3a and b show the effect of applying nomogram-derived probabilities of PCa in the study population. If, for example, a nomogram-predicted probability cut-off of 20% is applied, patients whose assay scores fall below that cut-off would be qualified by the nomogram as being at low risk; patients whose assay scores are above that cut-off would be qualified as high risk of having PCa. The base nomogram correctly classified only 16.8% of patients below the threshold who did not harbor biopsy-confirmed PCa. Conversely, the PCA3-17 nomogram correctly classified substantially more (30.4%) patients who did not harbor biopsy-confirmed PCa below the threshold. At the other end of the scale, the base nomogram incorrectly classified 83.2% of patients with a negative biopsy above the threshold, falsely indicating high risk of PCa. Again, the PCA3-17 nomogram classified substantially fewer patients (69.5%) with a negative biopsy above the nomogram threshold, exposing fewer men to prostate biopsy if this novel nomogram had been used. Finally, at the nomogram-predicted probability cut-off of 20%, these figures translate into an equally high sensitivity of 95% versus 92% for the base nomogram versus the PCA3-17 nomogram. Conversely, regarding specificity, the PCA3-17 nomogram outperforms the base nomogram: At the cut-off threshold of 20% risk of harboring PCa, the difference in specificity was 16.8% for the base nomogram versus 30.4% for the PCA3-17 nomogram. Accordingly, negative and positive predictive values were in favor of the PCA3-17 nomogram.

4. Discussion

In this study, we used the most stringent methodologic criteria suggested by Kattan, for which, in addition to demonstrating its independent predictor status, the candidate marker should enhance the overall PA of established predictors [6], and [7]. We added this methodology to the standard univariate and multivariate tests of the candidate marker PCA3. In univariate analyses predicting PCa at biopsy, all forms of PCA3 coding represented a statistically significant predictor (all p < 0.001), and they outperformed all other tested risk factors. Regardless of PCA3 coding, bootstrap-corrected PA of PCA3 assays was the highest among tested risk factors. The continuously coded PCA3 demonstrated the highest PA (0.68), followed by PCA3-17 (0.63), PCA3-24 (0.62), and PCA3-35 (0.62) (Table 2a). These findings match previously published results [4], [5], and [12]. Marks et al demonstrated an AUC for PCA3 versus PSA of 0.68 versus 0.52 in 226 men with PSA levels ≥2.5 ng/ml [4]. Similar results were found for PSA ranges <2.5 ng/ml, for PSA ranges >10 ng/ml, and in initial and repeat biopsy settings by Deras et al [5]. In their study, the AUC of PCA3 versus PSA in an initial and repeat biopsy setting was 0.70 and 0.68 versus 0.62 and 0.55, respectively. Furthermore, Haese et al reported an AUC for PCA3 of 0.66 versus 0.58 for percent of free PSA [12]. In multivariate analyses, regardless of whether PCA3 was used as a continuously coded variable or whether PCA3-17, PCA3-24, or PCA3-35 were used, PCA3 invariably represented an independent risk factor of PCa (all p < 0.001). The combined multivariate bootstrap-corrected PA of established risk factors age, DRE, PSA, prostate volume, and history of previous biopsy (ie, base model) for predicting PCa was 0.68. When PCA3 was added to the multivariate model, PA improved between 2% and 5%. PCA3-17 led to the highest increase in PA (5%, p = 0.04). This emphasizes the clinical potential of PCA3 assays to improve PCa prediction (Table 2b). Similar findings have been demonstrated by others [5], and [13]. Deras et al tested a multivariate model consisting of PSA level, prostate volume, and DRE result [5]. They found an increment in AUC from 0.67 for the base model versus 0.75 for the base model with PCA3 score. In contrast to our study, they did not investigate the incremental value of different PCA3 variants in multivariate models. In fact, we further carried on analyses and constructed a regression-based nomogram to predict PCa risk on an individual basis (Fig. 1a). This novel PCA3-17 nomogram was 0.73 accurate after internal validation, and it compares favorably with other recent biopsy nomograms [14], [15], and [16]. Ankerst et al tested urinary PCA3 in the Prostate Cancer Prevention Trial (PCPT) risk calculator [13]. Compared with our results and previous diagnostic prediction tools, which range between 0.73 and 0.76, they demonstrated a lower overall AUC of 0.70 after addition of urinary PCA3 assay to the six PCPT risk-calculator base predictors (PSA level, DRE result, family history of PCa, biopsy history, age, and African American race) which was significantly higher than the established PCPT risk calculator (0.70 vs 0.65; p < 0.05). Unlike our analyses, neither their AUC comparison between the multivariate model, including PCA3 versus PCA3 assay alone (0.70 vs 0.67), nor PCA3 assay alone versus PSA demonstrated statistically significantly differences (p > 0.05). Moreover, clinical decision-making aids should be simple to access, and many risk factors make their use impractical in busy clinical routine. The PCPT risk calculator including PCA3 assay relies on seven risk factors versus five risk factors in our study. Computational access is needed to apply the PCPT risk calculator to future patients. Conversely, our nomogram is easy to access in a paper-based format. Last, as correctly stated by Ankerst et al, their nomogram applies primarily to screening patients as opposed to referral patients in our cohort.

Taken together, our findings demonstrate that, following most stringent criteria, PCA3 does fulfill the characteristics of an independent and informative marker; thus PCA3 can be termed a novel diagnostic marker of PCa [6], and [7]. Moreover, it is important to note that within our cohort, commonly applied standard risk factors of PCa (age, DRE result, prostate volume, and biopsy history) failed to achieve satisfactory PA (0.68). This finding corroborates previous recent studies that show standard risk factors of PCa lose their PA in men at risk for PCa, [14], and [17], and this finding further corroborates the need for investigation of novel markers to improve PCa prediction. Our results suggest that the PCA3 assay is capable of diminishing this loss in PA. Specifically, addition of the PCA3 assay is related to an increase in PA of 5%. Recently, Shariat et al [18] added nine different, experimental, novel diagnostic markers to a multivariate model and demonstrated a gain in PA of 15% (72% to 87%, p < 0.001). Therefore, the increment of +5% related to one single marker (PCA3) is clearly remarkable and clinically significant. Furthermore, it may be expected that, in the context of the ongoing PSA screening debate with decreasing cancer-specific mortality [19] and growing PCa awareness, physicians will be faced with increasing numbers of men who are potential candidates for prostatic evaluation. Klotz et al recently calculated for the year 2007 that if the PSA threshold had been lowered to 2.5 ng/ml 2.74 million American men between 50–70 yr would have faced a biopsy indication [20]. Assuming a 30% detection rate, roughly 1.97 million men would have been left with elevated PSA levels. Such numbers demonstrate the clinical value of a novel, highly specific, PCa detection marker such as PCA3 and its potential as an individual risk stratification tool to select men for prostatic evaluation. Therefore, health and/or economic expenses are reduced as well as patient concerns (anxiety, discomfort, pain, and complications associated with prostate biopsies) [12]. Finally, in the PCa-staging scenario for which discrimination between significant and insignificant disease is under debate [21], PCA3 assay may be valuable in identifying patients for active surveillance or definitive therapy. This hypothesis is supported by work suggesting a correlation between tumor volume and PCA3 assay score [22].

After confirmation of PCA3 as a novel marker, we assessed risk factors of PCa of our cohort to construct a novel nomogram. Within this process, we determined 17 to be the most significant PCA3 cut-off value, and we tested variants of the risk factor PCA3 coded linearly or PCA3-17 and compared those to published PCA3 assay thresholds of 24 and 35 [4], and [11]. Our findings differ from results reported by Marks et al, who suggested that PCA3-35 is the most efficacious [4]. These differences may be explained by the relatively small sample size available to the authors; our findings were based on a study cohort that was four times larger. Nevertheless, we may hypothesize that, analogous to PSA levels, the PCA3 assay score represents a continuum of risk that is expressed in our univariate analysis: The continuously coded PCA3 score had the highest PA (0.68), which was as high as the base multivariate model; however, after controlling for other factors, the effect of continuously coded PCA3 disappeared. Instead, the multivariate results suggested a cut-off of 17 as the most informative PCA3 threshold to use for nomogram construction. The advantage of considering PCA3 is further corroborated by graphical exploration and nomogram probability cut-off analysis between the novel nomogram versus the base model (Fig. 1b and c, Table 3). Obviously, the increase of 5% in PA related to inclusion of the PCA3 assay improved detection of PCa in those men at lowest and highest risk of PCa, respectively. Using a cut-off probability of 20% (Table 3), for example, the PCA3 nomogram increased the rate of true-negative individuals from 17% to 30% and decreased the number of false-positive individuals from 83% to 70%. This is further corroborated by equally high sensitivity and substantially higher specificity of the PCA3-17 nomogram versus the base nomogram according to increasing nomogram-predicted PCa risk probability analyses (Table 3). Taken together, our PCA3 nomogram seems reliable to select men with PCa more accurately but also to identify those for whom further work may be spared.

Several limitations may apply to our findings. Although this study involves one of the largest PCA3 biopsy-verified patient cohorts to date, our findings are still based on a relatively small sample size. It needs to be acknowledged that statistical models such as nomograms depend heavily on their development data. Therefore, PCA3 discrepancies in PA between continuously coded PCA3 and PCA3-17 between univariate and multivariate analyses may be biased by the limited number of investigated subjects and must be corroborated in further validation studies. Furthermore, due to missing data, we were unable to compare performance of other established models such as the PCPT risk calculator or the initial biopsy nomogram by Nam et al to determine the most accurate model out of a wide array of published biopsy nomograms [16], and [23]. It may be argued that the prostate volume risk factor [24] does not represent a readily available risk factor because TRUS may be necessary; however, for those scheduled for repeat biopsy, TRUS-derived prostate volume is available. Actually, this may represent a strength of the PCA3 nomogram because it is applicable not only to the initial biopsy setting but also to the repeat biopsy setting. Nevertheless, further studies are warranted to replicate our findings and to validate externally the performance of this novel risk-stratification tool.

5. Conclusions

In conclusion, PCA3 was identified as a statistically independent and informative novel marker that is capable of increasing the PA of multivariate biopsy models. We constructed a novel PCA3-based individual risk stratification tool to identify men at risk of harboring PCa. It may assist patients and clinicians in deciding whether further prostatic evaluations are necessary.