Comparison of prostate cancer diagnostic models based on PCA3 and TMPRSS2:ERG RNA biomarkers from post-DRE and non-DRE urine specimens

Abstract

Purpose

Prostate cancer associated 3 (PCA3) and TMPRSS2:ERG fusion (T:E) are established urinary RNA biomarkers for detecting clinically significant prostate cancer (csPCa, Grade Group ≥ 2). Large within-subject paired studies directly comparing post-DRE and non-DRE urine collection are lacking. Using a multicenter cohort with samples collected from the same patients, we evaluated the effect of urine collection method on PCA3 and T: E diagnostic performance.

Methods

In a multicenter prospective study paired post-DRE and non-DRE whole-urine samples were collected before systematic biopsy. PCA3, T:E, and reference PSA RNAs were quantified in 373 post-DRE and 355 non-DRE eligible specimens. Logistic regression models combining PCA3, T:E, PSA density, and age were developed for post-DRE (PD2) and non-DRE (ND2) cohorts and evaluated by cross-validated receiver operating characteristic (ROC) analysis and decision curve analysis.

Results

PD2 outperformed ND2 in discrimination (AUC 0.78 versus 0.73) and net benefit. At ~ 95% sensitivity, PD2 specificity was 32% versus 23% for ND2. In the paired-sample analysis (n = 333), adding PCA3 and T:E to clinical variables improved AUC from 0.70 to 0.80 for post-DRE urine (p = 0.003), while gains were non-significant for non-DRE urine. Post-DRE specimens showed markedly higher RNA yields (median ~ 20-fold higher PCA3 and T:E copy numbers).

Conclusions

Post-DRE whole urine collection increased biomarker RNA yield and significantly improved csPCa diagnostic accuracy versus non-DRE whole urine. PCA3 and T:E provided meaningful additive value only in post-DRE specimens, supporting its preferred use. As biopsies were performed without MRI guidance, future studies in MRI-integrated pathways would further characterize clinical utility.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00345-026-06396-z.

Introduction

Prostate cancer (PCa) mortality is strongly related to tumour histological grade, as assessed by Gleason score, Grade Group (GG), or ISUP grade [1–3]. Clinically significant prostate cancer (csPCa, defined as GG ≥ 2) is associated with substantially greater risk of metastasis and mortality [1, 4]. Routine prostate-specific antigen (PSA) testing with or without digital rectal exam (DRE) lacks specificity for csPCa, while DRE itself has considerable inter-examiner variability [5–9]. Consequently, risk calculators, biomarkers, and multiparametric magnetic resonance imaging (mpMRI) are recommended as reflex tests to aid in biopsy decisions [10].

Urinary RNA biomarkers prostate cancer associated 3 (PCA3) and transmembrane serine protease 2 (TMPRSS2) gene and ETS transcription factor ERG (ERG) gene fusion (T:E) are used in non-invasive tests such as MyProstateScore (MPS), and ExoDx EPI [11, 12]. PCA3 is a long non-coding RNA overexpressed in prostate cancer cells versus healthy prostate cells. T:E results from a chromosomal rearrangement between TMPRSS2 and ERG producing a fusion mRNA detectable in urine that is specific to malignant prostate tissue. MPS uses whole post-DRE urine and combines PCA3, T:E, and serum PSA, whereas EPI measures PCA3 and ERG in an exosome fraction from non-DRE urine. According to the published data, the post-DRE MPS had slightly better discrimination than non-DRE EPI, but methodological differences between tests and cohorts preclude a direct comparison of urine-type effects.

We developed a test detecting PCA3 and T:E in whole urine specimens and compared the performance of logistic regression models using paired post-DRE (PD2) and non-DRE (ND2) specimens. To our knowledge, this is the first detailed, within-subject paired study with a substantial sample size to directly compare PCA3 and T:E for discrimination of csPCa.

Materials and methods

Study cohort

This prospective, multicenter study (May 2017 - June 2018) enrolled patients referred for prostate biopsy at four Lithuanian sites: National Cancer Institute (Vilnius), Lithuanian University of Health Sciences Medical Academy (Kaunas), Republican Siauliai County Hospital (Siauliai), and Vilnius University Hospital Santaros Klinikos (Vilnius).

Paired post-DRE and non-DRE whole urine specimens were obtained from each participant on the same day, prior to either primary or repeat prostate biopsy. After laboratory quality control, 373 post-DRE and 355 non-DRE specimens were eligible and used for the analyses unless stated otherwise. Based on biopsy findings patients were divided into two groups: a negative biopsy or a GG = 1 PCa and a clinically significant prostate cancer (csPCa, defined as GG ≥ 2) group.

Paired-sample analysis

We performed within-subject paired analyses using the subset of patients who provided both a valid post-DRE and a valid non-DRE urine sample collected on the same day (n = 333). Pairing was exact at the patient level: both specimens originated from the same individual and were required to meet laboratory quality-control criteria. Patients with only one valid specimen were excluded from paired analyses. Whenever the paired-sample subset was used, this is noted in the text. Specifically, the paired-sample subset was used to compare biomarker-based model discrimination and assess the additive value of PCA3 and/or T:E by adding the biomarkers to a clinical model (PSA density (PSAD) + age) in paired ROC (receiver operating characteristic) analyses.

Description of the laboratory analysis

The quantification of PCa biomarkers followed a proprietary protocol (Diagnolita, Vilnius, Lithuania). This protocol involves the collection of urine samples and the detection of biomarkers using the following kits: a urine collection and stabilization kit (#SK001), a nucleic acid purification kit (#PK001), and a kit for the quantification of targeted PCa-specific RNA biomarkers – PCA3 and T:E, along with the PSA mRNA for normalization (#QK001). The biomarkers were detected in the whole urine. Analytical validation of biomarker detection was done.

Model development for prediction of clinically significant prostate cancer

Logistic regression models were developed to predict the risk of clinically significant prostate cancer (csPCa) separately for post-DRE and non-DRE urine cohorts. Predictor variables were: log2(x + 1) transformed PCA3 and T:E scores, PSA density (PSAD) or PSA, and age (untransformed). Our primary Diagnolita test models, designated PD2 (post-DRE urine) and ND2 (non-DRE urine), incorporated PCA3 + T:E + PSAD + age.

For comparative analysis, MP (MP + for post-DRE, MP- for non-DRE) and EP (EP + for post-DRE, EP- for non-DRE) models were fitted to mirror variable combinations of MPS and EPI tests: PCA3 + T:E + PSA and PCA3 + T:E, respectively. Some comparisons involved established PCa risk calculators: the European Randomized Study of Screening for Prostate Cancer Risk Calculator (ERSPC-RC) and the Prostate Cancer Prevention Trial Risk Calculator version 2.0 (PCPTRC2).

Estimation of the additive value of biomarkers

The analysis of the additive value of biomarkers was conducted on a paired-sample subset (n = 333). The area under the ROC curve (AUC) values of nested models were compared by adding either one or both biomarkers to the clinical characteristic-only model "PSAD + age". The analysis was extended by comparing copy numbers and biomarker scores in post-DRE and non-DRE urine data.

Statistical analysis

Statistical analyses and figure generation were conducted in R (version 4.2.2). Two-tailed tests were employed, and a p-value of ≤ 0.05 was considered statistically significant.

The performance of Diagnolita test PD2/ND2 models was compared to that of the MP+/MP- and EP+/EP- models, along with PCa risk calculators and clinical characteristics. The comparisons involved Akaike Information Criteria (AIC) differences from the best fitted model (∆AIC), ROC characteristics, and decision curve analysis. The presented diagnostic performance characteristics were obtained using cross-validation (CV). Representative CV iterations with the ROC characteristics closest to the average from all iterations were used for presenting results or generating plots. Cutoffs for biomarker models were chosen to yield rounded sensitivities approximating 0.95.

Further details on laboratory analysis, model development, assessment of biomarker additive value, and statistical methods are given in the Supplementary Methods.

Results

Study cohort

We prospectively enrolled patients considered for prostate biopsy at 4 healthcare centers (Supplementary Table 2). In total, 409 post-DRE and 415 non-DRE samples were obtained. After excluding cases with missing clinical data and wrong sample collection (ineligible urine volume), 384 post-DRE and 384 non-DRE specimens remained suitable for RNA quantification. The non-informative rate (samples failing to meet RNA quality criteria, primarily PSA mRNA < 1000 copies) was significantly lower in the post-DRE cohort at 2.9% (11/384) compared to 7.6% for the non-DRE cohort (29/384). The final cohorts included 373 post-DRE and 355 non-DRE specimens (Supplementary Fig. 1).

Patients were divided into two groups: negative biopsy or GG = 1 prostate cancer versus GG ≥ 2 prostate cancer (clinically significant prostate cancer or csPCa). The csPCa group was older, had higher PSA and PSA density (PSAD) values, and a greater proportion of abnormal DRE results, while prostate volume was smaller (Supplementary Tables 3–4).

Model development for prediction of clinically significant prostate cancer

The PD2 (post-DRE) and ND2 (non-DRE) logistic regression models were built to predict csPCa using PCA3 and T:E biomarker scores, along with clinical variables, including PSAD and age. The minimum sample size requirements were met for both the post-DRE and non-DRE urine cohorts (Supplementary Table 1). Cross-validation revealed slight overfitting for both model types, as indicated by AUC values (Supplementary Table 5).

Performance characteristics of the developed models

The PD2 and ND2 models showed significantly higher predicted probabilities of csPCa in the GG ≥ 2 PCa group compared to the GG = 1 PCa and negative biopsy groups (Supplementary Fig. 2). PD2 showed the best overall accuracy: cross-validated AUC = 0.78 (95% CI 0.72–0.83), and a superior model fit (∆AIC versus the next best model, MP+ = 10.8). At cutoffs yielding ~ 95% sensitivity, PD2 specificity was 0.32 (versus 0.30 for MP+, 0.26 for EP+, 0.06 for ERSPC-RC and 0.14 for PCPTRC2), corresponding to 26% potentially spared biopsies (Table 1; Fig. 1 and Supplementary Table 6).

Table 1.

Performance characteristics of urine biomarker-based models and ERSPC-RC for detection of clinically significant prostate cancer (GG ≥ 2)

Model (post-DRE)	n a	∆AICb	AUC	p-value vs. PD2c	Cutoff d	Sensitivity	Specificity	NPV	PPV	Spared biopsies
PD2	373	0.0	0.78 (0.72–0.83)	NA	9.5	0.95 (0.90–0.99)	0.32 (0.27–0.38)	0.95 (0.90–0.99)	0.31 (0.26–0.37)	26%
MP+	373	10.8	0.75 (0.70–0.81)	0.053	10.0	0.95 (0.90–0.99)	0.30 (0.25–0.35)	0.94 (0.90–0.99)	0.31 (0.25–0.36)	24%
EP+	373	23.3	0.73 (0.67–0.79)	0.011	9.0	0.95 (0.90–0.99)	0.26 (0.21–0.31)	0.94 (0.88–0.99)	0.29 (0.24–0.35)	21%
ERSPC-RCe	332	40.3	0.70 (0.63–0.76)	0.039	1.0	0.96 (0.92–1.00.92.00)	0.06 (0.03–0.09)	0.83 (0.66–1.00.66.00)	0.25 (0.20–0.30)	5%

Model (non-DRE)	na	∆AICb	AUC	p-value vs. ND2c	Cutoff d	Sensitivity	Specificity	NPV	PPV	Spared biopsies
ND2	355	0.0	0.73 (0.67–0.79)	NA	11.5	0.94 (0.90–0.99)	0.23 (0.18–0.28)	0.92 (0.86–0.99)	0.29 (0.24–0.35)	18%
MP-	355	11.1	0.69 (0.63–0.76)	0.007	12.0	0.94 (0.90–0.99)	0.15 (0.11–0.20)	0.89 (0.80–0.98)	0.28 (0.23–0.32)	13%
EP-	355	20.8	0.66 (0.59–0.72)	0.004	13.5	0.94 (0.90–0.99)	0.17 (0.12–0.21)	0.90 (0.81–0.98)	0.28 (0.23–0.33)	14%
ERSPC-RCe	314	11.8	0.70 (0.64–0.77)	0.532	2.0	0.95 (0.90–1.00.90.00)	0.14 (0.10–0.18)	0.89 (0.79–0.99)	0.27 (0.22–0.32)	12%

AIC Akaike information criterion, AUC area under the receiver-operating characteristic curve, NPV negative predictive value, PPV positive predictive value, PD2 post-DRE urine Diagnolita test model, ND2 non-DRE urine Diagnolita test model, MP MPS-like model, EP EPI-like model, ERSPC-RC European Randomized Study of Screening for Prostate Cancer Risk Calculator

^aOnly samples meeting the data requirements for ERSPC-RC were included in ERSPC-RC comparisons, resulting in a smaller sample size for those rows

^b∆AIC was calculated as AIC − min(AIC) among candidate models (post-DRE or non-DRE). The optimal model has min(AIC) and thus ΔAIC = 0. AIC for ERSPC-RC was calculated by fitting a logistic regression model where the independent variable is a predicted clinically significant PCa probability from the calculator

^cMedian p-values are reported from 200 repeated cross-validation iterations. In each iteration AUC value of Diagnolita test model (PD2 or ND2) was compared to AUC value of alternative model using DeLong test

^dCutoffs were chosen to yield rounded sensitivities approximating 0.95. Reported sensitivity, specificity, NPV and PPV are the corresponding cross-validated estimates. "Spared biopsies" indicates the proportion of biopsies that would be avoided at the shown cutoff

^eERSPC-RC performance characteristics are uncorrected (no cross-validation performed) because no new model was trained for ERSPC-RC

Fig. 1.

Representative ROC curves for prediction of clinically significant prostate cancer in a post-DRE and b non-DRE urine cohorts. Solid green: Diagnolita test PD2/ND2 models, orange dashed: MP+/MP- models, violet dashed: EP+/EP- models, pink long-dash: ERSPC-RC, yellow long-dash: PSA. Biomarker-based model curves were taken from the single CV iteration (of 200 repeats) whose AUC, sensitivity, specificity and proportion of spared biopsies were closest to the mean across iterations. AUCs are shown with 95% confidence intervals. Biomarker-based model ROC curves are CV-corrected whereas ERSPC-RC and PSA curves are uncorrected

In the non-DRE cohort ND2 ranked highest (AUC = 0.73; 95% CI 0.67–0.79). While the ND2 model AUC was significantly greater than most alternatives, it was not statistically significantly different from ERSPC-RC (Table 1). At ~ 95% sensitivity cutoff values, the specificities were 0.23 for ND2, 0.15 for MP-, 0.17 for EP-, 0.14 for ERSPC-RC, and 0.14 for PCPTRC2 (Table 1; Fig. 1 and Supplementary Table 6).

To mitigate patient selection effects, we performed a paired-sample analysis in the subset of patients with both eligible specimens, which revealed that post-DRE models had a statistically significantly higher AUC than non-DRE models (Supplementary Table 7). Model performance in the PSA gray zone (PSA 3–20 ng/mL) and in the PSA gray zone with negative DRE was broadly similar to the main results (Supplementary Tables 8–9).

Generalizability was assessed via internal-external (leave-one-clinic-out) validation (Supplementary Table 10). AUCs were higher in Kaunas (0.82) than in Vilnius (PD2 0.75; ND2 0.68). Sensitivity was high across sites (0.88–1.00), whereas specificity varied: PD2 0.34–0.35 and ND2 0.31 (Vilnius) versus 0.12 (Kaunas).

Net benefit and calibration

Net benefit of the Diagnolita test PD2/ND2 models was contrasted against alternative MP and EP models and ERSPC-RC (Fig. 2). In the post-DRE urine dataset, PD2 model outperformed ERSPC-RC and showed the highest net benefit across the upper range of clinically relevant thresholds for csPCa, starting from 14% (Fig. 2a). In the non-DRE dataset, the biomarker-based models mostly exceeded ERSPC-RC except at high thresholds for MP- and EP- models (Fig. 2b). Overall, PD2 provided greater net benefit than ND2 model (Fig. 2), and in the paired-sample analysis PD2 consistently outperformed ND2 across a broad range of clinically relevant thresholds (Supplementary Fig. 3).

Fig. 2.

Decision curve analysis of the PD2/ND2 models (green curves), MP+/MP- models (orange curves), EP+/EP- models (violet curves), and ERSPC-RC (pink curves) compared to standards of biopsying all the patients (gray curves) or biopsying none of the patients (black horizontal lines). All comparisons were done by estimating the net benefit for csPCa prognosis. a Post-DRE urine cohort. b Non-DRE urine cohort

Calibration plots indicated that PD2 was reasonably well calibrated (Supplementary Fig. 4), whereas ND2 aligned with observed outcomes up to ~ 50% predicted risk but tended to overestimate risk at higher predicted probabilities (Supplementary Fig. 5).

The additive value of biomarkers

The additive value of PCA3 and/or T:E was assessed by adding them to the clinical variable-only model "PSAD + age". In the post-DRE urine group, the inclusion of T:E or both biomarkers led to significantly higher AUC values. In contrast, in the non-DRE urine cohort, the additional value of the biomarkers was minimal and not statistically significant (Supplementary Table 11). Similar results were observed in the full cohorts (Supplementary Table 12).

The better performance of the post-DRE urine biomarker-based models might be attributed to higher prostate RNA content. The paired-sample analysis showed that all RNA copy numbers were much higher after prostate massage (Supplementary Table 13). Interestingly, the effect size differed based on the RNA examined, with PCA3 and T:E showing the larger increase compared to PSA. This increase was reflected in the biomarker scores, which were significantly higher in post-DRE urine samples, particularly within the csPCa group (Supplementary Tables 13–15). However, the effect size varied considerably between patients, and about a third of patients exhibited decreased biomarker scores after massage (Supplementary Table 13).

Discussion

Most urine-based PCa tests use post-DRE samples, with the most common detected biomarkers being PCA3 and TMPRSS2:ERG (T:E) [13–16]. Non-DRE collection is more convenient, avoids the discomfort of DRE, and can be performed at home. Nevertheless, our paired-sample analysis showed that the post-DRE model (PD2) outperformed the non-DRE model (ND2): PD2 had a higher AUC (0.80 vs. 0.73 for ND2; p < 0.05). At high sensitivity levels, the post-DRE urine test had higher specificity than non-DRE alternative (0.37 versus 0.24; Supplementary Table 7). The impact of both PCA3 and T:E biomarker scores on the statistical model’s predictions was greater when post-DRE urine specimens were collected (Supplementary Tables 11–12). This within-subject paired comparison addresses a crucial gap in the literature, as previous studies of MPS (post-DRE) and EPI (non-DRE) tests utilized different cohorts and methodologies that prevented direct comparisons [11, 12, 16].

The post-DRE MPS test combined PCA3 and T:E with serum PSA and reported an AUC of 0.77 for csPCa detection [16]. The analogous MP+ model in our study had AUC = 0.75 (Table 1). Tosoian and colleagues validated MPS in a large group of 1525 patients [12]. The resulting sensitivity and specificity values in the mentioned study and from our data (MP+ model) were 0.97/0.95 and 0.33/0.30, respectively (Table 1). The main factor behind these differences could be that patient characteristics vary a lot based on the population [17]. Another reason could be a slightly higher accuracy of biopsies in the MPS validation study. As detailed, 99% of patients underwent ≥ 12 core biopsy versus 65% in our study (Supplementary Table 3) [12]. Regarding non-DRE testing, the exosome-based EPI test reported AUCs of 0.70 and 0.67 in American and European cohorts, respectively [18, 19]. In our cohort, a similar whole urine-based EP- model had an AUC of 0.66 (Table 1). The similarity of our MP + and EP- results to published MPS and EPI data, despite different cohorts and sample processing, suggests that specimen type (post-DRE vs. non-DRE) is a key factor of performance. Our paired-sample analyses confirmed this: post-DRE models consistently outperformed their non-DRE counterparts, showing significantly higher AUCs and greater specificity at cutoffs yielding ~ 95% sensitivity (Supplementary Table 7).

The higher additive value of biomarkers detected from the post-DRE urine could be related to the enrichment of prostate RNA content. Prostate massage increases prostate cell shedding and exosome release into urine, with our data showing a median 15-fold increase in PSA mRNA, 20-fold in PCA3, and 22-fold in TE copy numbers (Supplementary Table 13). The sample type effect was especially noticeable when detecting the lower abundance T:E biomarker. In non-DRE samples, T:E levels were often below the LOQ of 80 copies (Supplementary Tables 13–14). The worse performance of the PCA3 biomarker in the non-DRE urine cohort cannot be attributed to differences in detected copy numbers, as those numbers were actually above the LOQ. Another explanation could be related to the fact that urinary PCA3 is detected even in men under the age of 45 with no known prostate cancer risk factors [20]. It’s not entirely accurate to consider PCA3 as a prostate cancer-specific biomarker since healthy prostates have enough PCA3 RNA to generate valid PCA3 scores. In non-DRE urine, the variability of how many healthy cells were shed, could have a large impact on the resulting PCA3 scores. On the other hand, the prostate massage would help to improve the signal to noise ratio by increasing the content from shed PCa cells and from exosomes [13, 21, 22]. The high exosomal fraction in post-DRE urine may indicate a higher RNA fraction with PCa origin, as tumor cells excrete more exosomes than healthy cells [23, 24].

Our findings have practical implications for biomarker implementation. The superior diagnostic performance of post-DRE urine supports its use in urological clinic settings where DRE is already part of the standard evaluation. In this context, prostate massage adds minimal procedural burden while substantially enhancing biomarker signal, particularly for the lower-abundance T:E transcript. There is also a clear practical advantage to post-DRE sampling: it has a lower non-informative rate due to insufficient RNA (2.9% versus 7.6%). Therefore, post-DRE sampling is preferred unless a patient declines DRE. The non-DRE biomarker-based test accuracy could be improved by extending the panel beyond the established PCA3 and T:E biomarkers. In this way, the MPS2 test involving 18 biomarkers was validated [25]. Additionally, integrating urinary biomarkers with mpMRI could optimize patient selection for imaging and/or biopsy. Our post-DRE model (PD2) plus mpMRI showed promising results, achieving 99% sensitivity while reducing unnecessary biopsies by 25% [26]. Future studies to better evaluate model utility and generalizability are needed.

Our study has several limitations that need to be acknowledged. First, compared with other studies [11, 12, 16, 18], our cohort size was smaller, which reduced the statistical power to detect differences in AUC values. Second, the population was ethnically homogeneous (Lithuanian), and performance in more diverse populations may differ. Third, and most critically, this study was conducted prior to the widespread adoption of mpMRI at the participating centers (data collection: May 2017 - June 2018). Biopsies were performed using systematic protocols without MRI guidance or fusion targeting, which was the standard of care at the time. This pre-MRI reference standard differs from contemporary clinical pathways, and cancer detection rates and model performance thresholds reported here should not be directly extrapolated to mpMRI-guided cohorts. We note, however, that this is a limitation of the historical reference standard rather than of the biomarkers themselves, and that post-DRE PD2 model has since been shown to provide complementary value in an mpMRI-integrated pathway [26]. Fourth, there was no central pathology review. Fifth, we did not have a separate cohort for external validation. To mitigate potential over-optimism, we conducted internal validation and checked model generalizability by running clinic-level internal-external validation (Supplementary Table 10). Nevertheless, the model’s generalizability will need to be assessed more thoroughly in future external validation studies.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

M.L., S.P., J.B., and M.U were involved in manuscript writing and editing. M.L. was involved in data analysis. L.K, L.D., I. ZV. were involved in sample purification, biomarker detection and biomarker data curation. M.B., A.B., R.M., A. Z., A. U., L. J., F. J., and M. J. were involved in sample collection, patient data collection and management. J.G. was involved in supervision of the study. All authors read and approved the final manuscript.

Funding

The research was partly funded by using European Union Funds according to the priority axis 1 “Strengthening Research and Development and Innovation” implement No. J05-LVPA-K-01 “Intelektas. Bendri mokslo – verslo projektai”. The funded project had code No. J05-LVPA-K-01-0032 and was carried out from 2016-09-15 to 2018-12-14.

Data availability

The data supporting the findings of this study are available within the paper and its supplementary Information.

Declarations

Conflict of interest

M. Lopeta, S. Paliksa, and J. Belevicius are employees of JSC Diagnolita. M. Ukanis, L. Kasiliauskiene, L. Dziuge, I. Zutautaite-Varela, and J. Gagilas are former employees of JSC Diagnolita. The remaining authors declare no conflict of interest.

Ethical approval

The study was performed in accordance with the Declaration of Helsinki. Biomedical research protocol No. 2015-PVD was approved by the Lithuanian Bioethics Committee, approval No. L-15-06/1.

Informed consent

Informed consent was obtained from all individual participants included in the study.