A Novel MiRNA-Based Predictive Model for Biochemical Failure Following Post-Prostatectomy Salvage Radiation Therapy

Purpose To develop a microRNA (miRNA)-based predictive model for prostate cancer patients of 1) time to biochemical recurrence after radical prostatectomy and 2) biochemical recurrence after salvage radiation therapy following documented biochemical disease progression post-radical prostatectomy. Methods Forty three patients who had undergone salvage radiation therapy following biochemical failure after radical prostatectomy with greater than 4 years of follow-up data were identified. Formalin-fixed, paraffin-embedded tissue blocks were collected for all patients and total RNA was isolated from 1mm cores enriched for tumor (>70%). Eight hundred miRNAs were analyzed simultaneously using the nCounter human miRNA v2 assay (NanoString Technologies; Seattle, WA). Univariate and multivariate Cox proportion hazards regression models as well as receiver operating characteristics were used to identify statistically significant miRNAs that were predictive of biochemical recurrence. Results Eighty eight miRNAs were identified to be significantly (p<0.05) associated with biochemical failure post-prostatectomy by multivariate analysis and clustered into two groups that correlated with early (≤ 36 months) versus late recurrence (>36 months). Nine miRNAs were identified to be significantly (p<0.05) associated by multivariate analysis with biochemical failure after salvage radiation therapy. A new predictive model for biochemical recurrence after salvage radiation therapy was developed; this model consisted of miR-4516 and miR-601 together with, Gleason score, and lymph node status. The area under the ROC curve (AUC) was improved to 0.83 compared to that of 0.66 for Gleason score and lymph node status alone. Conclusion miRNA signatures can distinguish patients who fail soon after radical prostatectomy versus late failures, giving insight into which patients may need adjuvant therapy. Notably, two novel miRNAs (miR-4516 and miR-601) were identified that significantly improve prediction of biochemical failure post-salvage radiation therapy compared to clinico-histopathological factors, supporting the use of miRNAs within clinically used predictive models. Both findings warrant further validation studies.


Introduction
Prostate cancer (PCa) is one of the most common cancers worldwide and the most common cancer in men; however, treatment strategies remain highly controversial. Radical prostatectomy (RP) remains one of the more widely-used treatment options for men with early-stage PCa. Long-term data indicate that 30-40% of these patients experience biochemical failure after RP requiring salvage radiation therapy (RT); however, other studies have shown significantly different incidences due to different clinical prognostic characteristics of tumors [1][2][3]. The key clinical questions that are the focus of the current study are the identification of: 1) microRNAs (miRNAs) that predict biochemical recurrence after RP; 2) miRNAs that predict for biochemical recurrence after salvage radiation following failure after RP; and 3) miRNAs that can improve prediction of biochemical recurrence in combination with currently used clinicohistopathological factors, such as prostate-specific antigen (PSA), pathologic tumor (pT) and lymph node (pN) classification, resection status, and Gleason score. Multiple nomograms and classification models have been derived utilizing traditional clinico-histopathological parameters (CAPRA score [4][5][6], Partin table [7,8], D'Amico classification [9], and the three Stephenson Nomograms [10][11][12]) in an effort to establish prognosis, but these all have limitations in the context of salvage RT. Genetic markers to help guide decision making processes are also being developed for PCa as they have been for breast cancer, such as OncotypeDX [13] and a gene expression-based genomic classifier [14,15]. The goal of this study was to develop a miRNA signature that can add information to the existing clinical models and thereby help guide treatment decisions. miRNAs are small (*22 nucleotides), non-coding RNAs that regulate gene expression and are attractive candidates for biomarkers as they have been shown to play a vital role in tumorigenesis, and can be detected in clinical samples (biopsy, urine, and serum) allowing for non-invasive or minimally invasive molecular detection and prognosis of tumors [16]. Further, due to their small size they are stable in formalin-fixed paraffin-embedded (FFPE) tissues, which allows for discovery retrospectively in patient specimens [17]. Recently, numerous studies have been published demonstrating the value of studying miRNAs in the context of PCa (see ref [18] for an extensive review). These studies have used global profiling to examine miRNA signatures that are diagnostic or prognostic of biochemical failure post-prostatectomy [19][20][21][22][23][24]. However, these analyses have not accounted for treatment effects within the patient cohort. In the current study, we sought to examine a cohort of prostatectomy FFPE specimens from patients who all received salvage RT following biochemical failure post-prostatectomy using NanoString technology [25]. To our knowledge, no studies have been reported using the Nano-String nCounter system to measure miRNA expression in prostate tumor specimens, which directly measures 800 human transcripts by digital counts [25]. In addition, NanoString has shown to be a better technique than qRT-PCR for FFPE specimens and correlates more with its paired fresh-frozen counterpart than qRT-PCR [26]. NanoString technology also has been used in biofluids. For example, one previous study has examined a 20-gene mRNA panel using the NanoString system for urine samples collected from patients with PCa as a potential diagnostic tool [27]. In addition, a biomarker signature utilizing NanoString technology is FDA approved and in current clinical use for breast cancer representing its clinical utility [28]. Herein, we report on miRNAs that identify PCa patients at high risk of biochemical relapse after RP and miRNAs associated with high risk of biochemical relapse in patients undergoing salvage radiation after documented biochemical failure following prostatectomy. We also present a miRNA-and clinical-based predictive model of biochemical recurrence risk post-salvage radiation therapy. To our knowledge, this article represents the first such report on the prognostic value of miRNAs in the setting of biochemical failure and salvage RT post-RP.

Patient Cohort
FFPE blocks from 43 patients with PCa, who had undergone RP (1997-2009) followed by salvage RT between 2005 and 2011 at the University of Freiburg, Germany, were used in this study. Patients who received adjuvant RT (RT initiated less than 6 months after RP) were excluded from the analysis. Follow-up time after RT was defined to be greater than 4 years. A clinical database was established, which contained patient characteristics, tumor classification according to the International Union Against Cancer (UICC)/American Joint Committee on Cancer (AJCC) 2010 TNM system [29,30] due to the dates of the earliest radical prostatectomies in the study, tumor grading according to Gleason score without inclusion of tertiary score, radiation treatment details including toxicities, and follow-up details including PSA concentrations. There were no significant differences in terms of radiation techniques, planning target volume generation, doses and radiation volumes in the cohort. All patients were treated 1.8-2.0 Gy per fraction to a total dose of 66.6-74 Gy. Use of hormone therapy and whole pelvis radiation was at the discretion of the treating physician. Recurrence was defined as either biochemical failure (BF) with a rise of PSA 0.2 ng/ml at least twice as per the AUA recommendation [31] or clinical progression with local (prostatic fossa), regional (lymph nodes) or distant (metastasis) recurrence. Imaging tests (CT, PET/CT or bone scan) were performed when clinically indicated in cases of BF.
Ethics Statement. This study was approved by both the Medical Center University Freiburg and The Ohio State University institutional review boards with waived patient consent due to the archival nature of the study. In addition, data and tissue samples pertaining to this study were de-identified prior to analysis.

Sample Processing & RNA Isolation
All formalin-fixed, paraffin-embedded (FFPE) blocks were reviewed by the same pathologist specialized in genitourinary diseases and Gleason scores were reassessed for each core using the ISUP 2005 grading system [32]. Tissue samples from the Medical Center University of Freiburg pathology tissue bank were obtained, which included tissues from prostate tumors. Areas of interest were marked on an H&E stained slide and 1-mm diameter biopsy cores were punched from each area of interest. Cores from tumor were taken from the area with the highest Gleason Score and were enriched for tumor cells (>70%).
Total RNA was isolated using a combination of RecoverAll Total Nucleic Acid Isolation digestion buffer, AM1975, (Life Technologies; Carlsbad, CA) with Qiagen FFPE miRNeasy kit, Cat. # 217504 (Qiagen; Venlo, Limburg). In brief, cores were digested overnight using Recover-All Total Nucleic Acid Isolation digestion buffer and proteinase K leading to enhanced digestion of the core. Once cores were digested (next day), we followed the Qiagen FFPE miRNeasy manual starting at the 80°C for 15 minute incubation. RNA concentrations were measured using a NanoDrop 2000 spectrophotometer (Thermo Scientific; Waltham, MA).

miRNA Expression Analysis
For miRNA expression data generation, the NanoString human v2 array, which contains 800 miRNA probes, was used. Forty three tumor samples were analyzed. Expression analysis was conducted at The Ohio State University Nucleic Acid Core Facility (Columbus, OH). A total of 100ng RNA input was used per sample and conditions were set according to the manufacturer's recommended protocol (NanoString Technologies; Seattle, WA). miRNAs were quantified using the nCounter Digital Analyzer as counts. miRNAs were filtered out from downstream analysis if total counts were less than 32 across 90% of the samples. Four hundred and seventy seven miRs were left after the filtering. Data were normalized by the geometric mean of all targets using the nSolver software (NanoString Technologies; Seattle, WA) [33]. The data discussed in this publication have been deposited in NCBI's Gene Expression Omnibus [34] and are accessible through GEO Series accession number GSE65061.

Statistical Analysis
To compare the difference in tumor miRNA expression between late and early time to first biochemical recurrence post-RP (36 months as a cut-off) or between the time to the recurrence post-salvage RT 2-sample t-tests were used. For each miRNA, patients were dichotomized into high and low groups based on the median miRNA expression and the difference in the probabilities of the time to the recurrence (post-RP or post-salvage RT recurrence) was compared using the log-rank test for each miRNA. Multivariate analyses were performed using Cox proportional hazard regression models. ROC curve analysis was performed to determine the capability and cut-off level of variables that distinguished between the recurrence and nonrecurrence of post-salvage RT. All analyses were performed using SAS 9.3 (SAS, Inc; Cary, NC) or R 3.0.

miRNAs predictive of PSA recurrence post-prostatectomy
Univariate Analysis. A schematic of the study design is shown in Fig. 1. A clinical database was established for 43 PCa patients who underwent both RP and salvage RT (Table 1). Risk factors included 32.5% of patients who had a Gleason score of 8 or above, 25.6% of patients who had seminal vesicle invasion or extraprostatic extension, and 41.9% of patients who had positive margins. Salvage RT was started at a median PSA value of 0.39 ng/ml. Nineteen patients (44.2%) experienced BF after salvage RT at a median time of 27.1 months (range 0.0-64.1 months). The median follow-up times after RP and after salvage RT were 6.9 and 3.7 years, respectively.
Using these data, miRNA expression was correlated with time to first biochemical recurrence to determine if miRNAs can predict biochemical recurrence post-RP. Time to first recurrence was defined as the time from prostatectomy to the start date of salvage RT. Using tumor expression only, we identified 54 miRNAs that were significantly differentially expressed 1.5fold or greater between patients who had an early ( 36 months) versus late (> 36 months) recurrence (S1 Table). Thirty six months was used as the cut-off as it was near the median of time to recurrence. For another approach to identify miRNAs that correlated with time to first biochemical recurrence, we dichotomized patients according to median miRNA expression and compared the probabilities in the time to the first recurrences between the two groups (high vs low miRNA expression) using log-rank tests. One hundred and twenty three miRNAs were identified that could differentiate the two groups in the time to first biochemical recurrence (p-value < 0.05) (S2 Table).
Multivariate Analysis. In order to determine if miRNAs can be independent markers of biochemical recurrence post-RP, two multivariate Cox regression analyses were performed using all 123 miRNAs identified to be predictive of biochemical failure post-prostatectomy (pvalue < 0.05) by univariate log-rank analysis (S2 Table). The first multivariate analysis considered initial PSA value (continuous) and Gleason score; age and resection status were eliminated as these did not seem to affect the recurrence in univariate analysis. Upon analysis of the first recurrence using multivariable Cox regression analysis, 97 miRNAs had p-values < 0.05 (S3 Table). In the second multivariate analysis, different clinical factors were taken into consideration. Three clinical covariates were examined independently, D'Amico score, Stephenson score (categorical), and Stephenson score (continuous), to identify miRNAs associated with time to first recurrence. miRNAs were removed from further analysis if a significant p-value (< 0.05) was not obtained in any of the three analyses which resulted in 88 miRNAs that were statistically associated with time to first recurrence after multivariate analysis (S4 Table).  Patients were classified into two groups using unbiased hierarchical cluster analysis for those selected 88 miRNAs (Fig. 2A). The Kaplan-Meier method showed that the probabilities in the time to first biochemical failure were significantly different between the two groups of patients with distinguished miRNA signatures (log-rank, p = 0.005) (Fig. 2B).
miRNAs predictive of PSA recurrence post-salvage RT Univariate Analysis. miRNAs that could predict biochemical failure after salvage RT treatment are of utmost interest. At this time, no studies have examined miRNAs associated with clinical outcomes following post-RP salvage radiation. The recurrence after salvage radiation treatment (also referred to as the 2 nd recurrence) was defined as a rise in PSA to 0.2 ng/mL at least twice consecutively following the nadir [35]. In our patient cohort, 19 out of 43 patients experienced biochemical failure after salvage radiotherapy (Table 1). Further, using tumor miRNA expression only, we identified 4 miRNAs that were significantly expressed 1.5-fold or greater in patients who experienced a second recurrence when compared with those who did not (p-value < 0.05) (S5 Table). In the univariate log-rank analysis, 24 miRNAs were identified, which could predict patients who recurred a second time after salvage RT from those who did not recur (p-value < 0.05) (S6 Table). Multivariate Analysis. Similarly, in order to determine if miRNAs can independently predict biochemical recurrence post-salvage RT, multivariate Cox regression analyses were performed using the 24 miRNAs that were significantly associated (p-value < 0.05) with recurrence post-salvage RT by univariate log-rank analysis (S6 Table). Due to the small sample size and number of events, only two covariates were examined in the multivariate analysis. The multivariate analysis was performed using lymph node status and Gleason score (these factors were chosen as having the lowest p-values upon univariate analysis as shown in (Table 2)), leading to identification of 9 miRNAs that predicted biochemical recurrence post-salvage RT (Table 3).

Predictive effect of miR-601 + miR-4516 & Development of a miRNAbased Predictive Salvage RT Model
A primary objective of this study was to identify miRNAs that could be used independently as predictive biomarkers of salvage RT as well as to utilize in combination with other clinical factors to improve the sensitivity and selectivity of existing models. We then included lymph node status, Gleason score and the 9 miRNAs as the covariates in the Cox regression model and used a stepwise model selection strategy to capture miRNAs which could be good candidates in predicting the biochemical recurrence post-salvage RT. The two miRNAs, miR-601 and miR-4516 alone and together with Gleason score and lymph node status were selected by the model. An area under the curve (AUC) of receiver-operator characteristics (ROC) models using the nearest neighbor estimation method was utilized [78]. An AUC for lymph node and Gleason score was found to be 0.66, the addition of miR-601 and miR-4516 increased the AUC to 0.83 (Fig. 3). Interestingly, each miRNA alone (miR-601 AUC = 0.77 and miR-4516 AUC = 0.68) or together (miR-601 + miR-4516 AUC = 0.76) (Fig. 3) had a better predictive capability than the combination of positive lymph nodes together with Gleason score (AUC = 0.66) (Fig. 3). Using the risk score generated by this model (miR-4516, miR-601, Gleason score, and lymph node status), patients were classified as high or low risk groups based on median, the probabilities in recurrence post-salvage RT of the two groups were found to be significantly different (logrank, p<0.001) as shown in Fig. 4. To ensure that hormonal therapy did not confound the predictive model, patients that received hormonal treatment at any point were removed from the model and the model still retained its specificity and sensitivity with a log-rank p-value of 0.016 and AUC value of 0.83 (S1 Fig.). To better understand mechanism(s) by which these miRNAs are acting, mRNA targets of all novel miRNAs, including miR-601 and miR-4516 were identified using TargetScan [79] and microRNA.org [80] (S7-S10 Tables).

Discussion
Currently used models and nomograms of clinical response following RP are not able to clearly distinguish subgroups of patients with a more aggressive type of PCa from those with indolent disease. Further, there remains an ongoing debate of which patients should receive adjuvant Stephensen score (continuous)). This signature (A) appears to differentiate between patients with early recurrence (< 36 months) (red box) vs those with late recurrence (> 36 months). Kaplan-Meier plots were generated using the two cluster groups (B versus salvage RT post-RP. Thus, the identification of new biomarkers that are predictive of failure post-RP is essential as PSA can only accurately identify biochemical recurrence once failure has occurred. This study sought to use miRNAs as molecular biomarkers in conjunction with clinical factors to help identify patients who would respond to salvage RT as well as patients who may be excellent candidates for adjuvant RT. miRNAs are ideal biomarkers and capable of being detected in biopsies and biofluids which may make identification of miRNAs in PCa a clinically useful test. Very few studies have examined miRNAs globally in PCa, specifically in correlation to clinical outcome. The current study is the first to use Nanostring to profile miRNAs in PCa tissue and further elucidates the differential expression of miRNAs in PCa and correlates the expression with outcome data. There are a number of studies that have dealt with this question resulting in inconsistent results so far. One of the major problems in conducting global miRNA expression correlative studies is that a lot of the patient cohorts are heterogeneous in terms of treatments received (surgery, radiation, hormonal therapy). To our knowledge, there are no studies that have previously examined miRNAs in the context of post-prostatectomy radiation. This study addressed three different questions associated with PCa and miRNAs: 1) Which miRNA signatures can predict time to recurrence after RP?; 2) Which miRNAs can predict recurrence after salvage RT?; and 3) Which miRNAs can add predictive information beyond currently used clinico-histopathological factors?
There are some limitations to the present study. First of all the sample size was relatively small, secondly, <50% of patients had BF after salvage RT, and lastly, androgen deprivation therapy was prescribed at the discretion of the treating physician for a small subset of patients reflecting current treatment practices but imposing bias to the study cohort. Although hormonal therapy does not appear to be a confounding variable in the study, further validation using larger samples sizes as well as patients who received adjuvant and observation after prostatectomy will be required to determine the clinical usefulness of the identified predictive models. However, the predictive ability of the identified signature and regression model appears to be potentially useful relative to currently used clinical and histo-pathological criteria. In addition, further investigation of molecular mechanisms involving miRNAs will likely lead to the elucidation of prostate tumorigeneisis and progression as well as identification of putative targets for therapy.
In summary, a miRNA signature was identified that may be able to predict time to biochemical recurrence following RP and a logistic regression model was established which predicts biochemical failure following salvage RT. The 88-miRNA signature alone has the potential to predict time to first recurrence indicating that miRNAs identified by NanoString using FFPE specimens can serve as promising biomarkers of biochemical failure prediction. Currently, there are no routinely used clinical factors or models to predict failure post-salvage RT. We showed that miRNAs alone predicted biochemical recurrence post-salvage RT and adding two miRNAs, miR-601 and miR-4516, to lymph node status and Gleason score, greatly improved the predictive ability of these clinical factors alone. Additionally, this post-RT predictive model was superior to other clinical factors examined in this cohort of patients.  Table. Tumor-only miRNA expression comparisons between patients with early (36 months) and late (>36 months) biochemical recurrence. p-values were calculated using a repeated measured ANOVA model. Table. Tumor-only miRNA expression was used to predict time to first biochemical recurrence. Patients were divided into two groups by the median miRNA expression and a univariate log-rank test was performed to correlate expression with time to first biochemical recurrence. Significant p-value < 0.05. (DOCX) S3 Table. miRNAs that predict time to failure post-RP (first recurrence) in salvage radiation patients. Hazards ratios were generated using a multivariate Cox regression analysis (initial PSA and Gleason score). Only miRNAs with a significant p-value (<0.05) are shown. (DOCX) S4 Table. miRNAs that predict time to failure post-RP (first recurrence) in salvage radiation patients. Hazard ratios were generated using a multivariate Cox regression analysis (D'Amico, categorical Stephensen, or continuous Stephensen). Only miRNAs with a significant p-value (<0.05) are shown. (DOCX) S5 Table. Tumor-only miRNA expression comparisons between patients with biochemical recurrence post-salvage RT (second biochemical recurrence) versus no recurrence post-salvage radiation therapy. p-values were generated using a repeated measured ANOVA. (DOCX) S6 Table. Tumor-only miRNA expression was used to predict biochemical recurrence postsalvage RT (second biochemical recurrence). Patients were divided into two groups by the median miRNA expression and a univariate log-rank test was performed to correlate expression with second biochemical recurrence. Significant p-value < 0.05. (DOCX) S7 Table. Putative targets of novel miRNAs, miR-1193 and miR-626 associated with biochemical recurrence post-salvage RT (second biochemical recurrence). All targets of both miR-1193 and miR-626 are listed with scores according to both TargetScan and microRNA. org. (DOCX) S8 Table. Putative targets of miRNAs, miR-601 and miR-4516, within our predictive model of biochemical recurrence post-salvage RT (second biochemical recurrence). All targets of both miR-601 and miR-4516 are listed with scores according to both TargetScan and microRNA.org. miR-4516 was not listed in microRNA.org. (DOCX) S9 Table. Top 12 putative gene targets and function for miR-601. Twelve genes with the lowest scores and highest probability of being targeted by miR-601 according to TargetScan and microRNA.org are listed with their known gene functions. (DOCX) S10 Table. Top 10 putative gene targets and function for miR-4516. Ten genes with the lowest scores and highest probability of being targeted by miR-4516 according to TargetScan are listed with their known gene functions. (DOCX) S11 Table. Ingenuity Pathway Analysis (IPA) was performed on the 88-miRNA signature associated with time to first biochemical recurrence. Displayed are top networks, diseases and disorders, and molecular and cellular functions. (DOCX) S12 Table. Ingenuity Pathway Analysis (IPA) was performed on the 9-miRNAs associated with biochemical recurrence post-salvage radiation therapy. Displayed are top networks and diseases and disorders. (DOCX)