Development and validation of a novel multivariate risk score to guide biopsy decision for the diagnosis of clinically significant prostate cancer

Abstract Objectives Selecting patients suspected of having prostate cancer (PCa) for a prostate biopsy remains a challenge. Prostate‐specific antigen (PSA)‐based testing is hampered by its low specificity that often leads to negative biopsy results or detection of clinically insignificant cancers, especially in the 2‐10 ng/mL range. The objective was to evaluate a novel diagnostic test called Proclarix incorporating thrombospondin‐1 and cathepsin D alongside total and free PSA as well as age for predicting clinically significant PCa. Patients and methods The test was developed following a retrospective study design using biobanked samples of 955 men from two reference centres. A multivariate approach was used for model development followed by validation to discriminate significant (grade group ≥2) from insignificant or no cancer at biopsy. The test specificity, positive predictive value (PPV) and negative predictive value (NPV) at a fixed sensitivity of 90% were compared to percent free PSA (%fPSA) alone. The number of avoidable prostate biopsies deemed to be representative of clinical utility was also assessed. Results In the targeted patient population, the test displayed increased diagnostic accuracy compared to %fPSA alone. Application of the established model on 955 patients at a fixed sensitivity of 90% for significant disease resulted in a specificity of 43%, NPV of 95% and a PPV of 25%. This is in comparison to a specificity of 17%, NPV of 89% and PPV of 19% for %fPSA alone and had the potential to reduce the total number of biopsies needed to identify clinically significant cancer. Further, the test score correlated with significance of cancer assessed on prostate biopsy. Conclusions The Proclarix test can be used as an aid in the decision‐making process if to biopsy men in this challenging patient population. The use of the test could reduce the number of biopsies performed avoiding invasive procedures, anxiety, discomfort, pain and complications.


| INTRODUC TI ON
Selecting patients suspected of having clinically significant prostate cancer (PCa) (ISUP grade group GG ≥ 2) for a prostate biopsy remains a challenge despite the growing number of available diagnostic tools. The challenge to orient biopsy decision making is especially present when evaluating patients within the "prostate-specific antigen (PSA) grey zone" of total PSA (tPSA) 4-10 ng/ mL 1 or 2-10 ng/mL, 2,3 a growing cohort of patients with an ageing population, where test performance can vary. The decision to biopsy is further compounded when considering the influence of prostate volume, 4 family history, 5 prior biopsy status 6 and digital rectal exam (DRE) status. 7 The diagnostic tools currently available are generally PSA based, using other forms of PSA, different molecular markers or mathematical combinations of such markers. 8 While they represent an improvement, their performance levels vary depending on the validation cohort used and the intended target population in terms of PSA range, prostate volume and DRE characteristics. PSA is historically and currently the most frequently used marker but it is not cancer-specific 9 and its low specificity leads to over diagnosis. 10 In consequence, depending on the specific cohort,only about 25% up to a maximum of 60% of men with tPSA values in the 4-10 ng/mL range have a positive biopsy. 11 Lower cut-offs, for example, 2 ng/ mL or age-specific cut-offs can improve sensitivity, but sacrifice specificity. 12 The ratio of free PSA (fPSA) to tPSA (percent free PSA [%fPSA]) has also been shown to improve test performance 11 but has limitations as %fPSA both increases with age as well as with prostate size and yields improved results only in patients with small prostates (<40 mL). 13,14 We have previously shown in one of the most challenging subsets of subjects presenting with a tPSA of 2-10 ng/mL, prostate volume ≥35 mL, no prior history of PCa and a normal DRE, that the combined measurement of two novel glycoproteins thrombospondin-1 (THBS1) and cathepsin D (CTSD) can improve the identification of clinically significant PCa. 15 Based on these results we have developed a new test named Proclarix. This test incorporates THBS1 and CTSD with patient age, tPSA and %fPSA values into a dedicated algorithm and provides a risk score that corresponds to the probability of detecting clinically significant PCa on biopsy.
The purpose of this study was to validate the performance of Proclarix including the 5-parameter multivariate logistic regression algorithm, as compared to %fPSA alone in discriminating no cancer and GG < 2 versus GG ≥ 2.

| Study design
This was a prospectively planned, retrospective, blinded, 2-center study using biobanked samples to establish and validate the use of the test in identifying clinically significant PCa. Samples were blinded to experimenters and were only unblinded once measurements were complete. A formal sample size calculation was not performed, however, to select predictors to forecast a binary outcome from k variables, it is proposed that there would need to be k × 10 to k × 20 cases in the smaller group, 16 whereas Steyerberg et al 17 recommends k × 50 cases and mentions a lower limit is k × 10 cases. Considering five biomarkers, a sample size of 100 results for the smaller group (positive cases) is deemed sufficient to meet the proposed requirements.

| Study population
The study population consisted of biobanked samples from two reduce the number of biopsies performed avoiding invasive procedures, anxiety, discomfort, pain and complications.

| Assay methods
The Proclarix test is comprised of two quantitative Enzyme-linked Immunosorbent Assays (ELISA) that measure the concentration of THBS1 and CTSD in human serum. A dedicated software integrates the values for THBS1 and CTSD, age, tPSA and fPSA (from third party manufacturers) to calculate a risk score. The ELISAs along with the software form the CE marked Proclarix test. The ELISA kits used in this study were derived from two different manufacturing lots to which samples were randomly assigned. If during measuring samples were outside of the measuring range of the assay, samples were re-diluted and remeasured so that all subjects in the study population had a test result. Serum tPSA and fPSA were re-analysed for Hamburg samples using the ADVIA Centaur immunoassay system (Siemens Healthcare) to calculate %fPSA. Available PSA values of samples from Innsbruck were obtained from an Elecsys system (Roche Diagnostics). Due to known variations for tPSA and fPSA measurements between analysis kits from different manufacturers, 18

| Model development
An initial feature selection was performed on the dataset to obtain the best possible model to predict the outcome of clinically significant PCa (GG ≥ 2). This resulted in a final mathematical biomarker model incorporating THBS1, CTSD, tPSA, %fPSA and age that was subsequently established using all 955 samples. The "risk score" is derived from the regression analysis and is represented as a percentage scale from 0%-100% indicating the risk of significant PCa. The cut-off was set at a sensitivity of 90%, no further model recalibration was performed.

| Model assessment
The mathematical biomarker model was validated using a split sample approach, where the complete dataset has been divided into a training and validation dataset in a 100:70 ratio with balanced prevalence. Logistic regression parameter estimates and a cut-off for the linear predictor, referring to 90% sensitivity selected for a 10% false negative rate, as obtained from training cohort's analyses, have been applied to samples in the validation set in order to assess the model's suitability. In total, the model validation was based upon 1000 independent sets of training and validation sets. In order to evaluate the performance of the developed model, 19

| Study population
Patient characteristics are summarised in Table 1

| Model assessment: Split sample approach
Following development of the biomarker model, a split sample training-validation approach was used to yield reliable performance predictions. The median specificity based upon 1000 independent sets of training and validation resulted at 89% sensitivity (derived from cut-off at 90% from training set) in 42% specificity, NPV of 95% and PPV of 25%, respectively, for Proclarix. This is in comparison to %fPSA alone at 90% sensitivity which displayed a specificity of 17%, 89% NPV and 19% PPV respectively (Table 2 and Figure 1). As the test displayed a median specificity of 42% in 1000 independent validations, the biomarker model was shown to be validated with respect to its suitability in predicting clinically significant PCa. Values for training and validation for both the biomarker model and %fPSA were similar, suggesting limited overfitting of the model ( Table 2).
The Proclarix biomarker model results in a risk score that showed a significant increase across groups (no PCA, GG < 2 and GG ≥ 2) (Kruskal-Wallis P < .001) and could thus differentiate aggressiveness of clinically significant PCa detected on biopsy (Figure 2).
In comparison, using %fPSA as the decision support test, 13 Gleason 3 + 4 (GG 2), two 4 + 4 and two 3 + 5 (GG 4) were missed. The number of avoided biopsies would be more than double in the biomarker model (37%) compared to %fPSA (16%), while keeping the number of missed cancers constant.

CO N FLI C T O F I NTE R E S T S
Some of the authors have received/held stock options (SG) and salaries (BG) and founder shares (RS) from ProteoMediX. SK and TS are advisors to ProteoMediX. TS and RS are inventors of the patent application WO2018011212, and SG and RS have the patent application 32 WO2009138392.

FU N D I N G I N FO R M ATI O N
Funding for this work was provided by ProteoMediX AG.