Characterization of RNA‐binding motif 3 (RBM3) protein levels and nuclear architecture changes in aggressive and recurrent prostate cancer

Abstract

Background

The RNA‐binding motif protein 3 (RBM3) has been shown to be upregulated in several types of cancer, including prostate cancer (PCa), compared with normal tissues. Increased RBM3 nuclear expression has been linked to improved clinical outcomes.

Aims

Given that RBM3 has been hypothesized to play a role in critical nuclear functions such as chromatin remodeling, DNA damage response, and other posttranscriptional processes, we sought to (a) quantify RBM3 protein levels in archival PCa samples and (b) develop a nuclear morphometric model to determine if measures of RBM3 protein levels and nuclear features could be used to predict disease aggressiveness and biochemical recurrence.

Methods and results

This study utilized two tissue microarrays (TMAs) stained for RBM3 that included 80 total cases of PCa stratified by Gleason score. A software‐mediated image processing algorithm identified RBM3‐positive cancerous nuclei in the TMA samples and calculated 22 features quantifying RBM3 expression and nuclear architecture. Multivariate logistic regression (MLR) modeling was performed to determine if RBM3 levels and nuclear structural changes could predict PCa aggressiveness and biochemical recurrence (BCR). Leave‐one‐out cross validation (LOOCV) was used to provide insight on how the predictive capabilities of the feature set might behave with respect to an independent patient cohort to address issues such as model overfitting. RBM3 expression was found to be significantly downregulated in highly aggressive GS ≥8 PCa samples compared with other Gleason scores (P < .0001) and significantly downregulated in recurrent PCa samples compared with nonrecurrent samples (P = .0377). An 11‐feature nuclear morphometric MLR model accurately identified aggressive PCa, yielding a receiver operating characteristic area under the curve (ROC‐AUC) of 0.90 (P < .0001) in the raw data set and 0.77 (95% CI, 0.83‐0.97) for LOOCV testing. The same 11‐feature model was then used to predict recurrence, yielding an ROC‐AUC of 0.92 (P = .0004) in the raw data set and 0.76 (95% CI, 0.64‐0.87) for LOOCV testing.

Conclusions

The RBM3 biomarker alone is a strong prognostic marker for the prediction of aggressive PCa and biochemical recurrence. Further, RBM3 appears to be downregulated in aggressive and recurrent tumors.

Abbreviations

BCR

Biochemical recurrence

CSP

Cold‐shock protein

GS

Gleason score

IOD

Integrated Optical Density

LOOCV

Leave‐one‐out cross validation

MLR

Multivariate logistic regression

PCa

Prostate cancer

PSA

Prostate‐specific antigen

RBM3

RNA‐binding motif protein 3

ROC‐AUC

Receiver‐operating characteristic area under the curve

TMA

Tissue microarray

1. BACKGROUND

Prostate cancer (PCa) is the most common cancer in males in the United States and the second leading cause of cancer‐related deaths in men.1 Because of the heterogeneous nature of PCa progression, there is a growing need to identify biomarkers to aid in early prognostic risk stratification. While many men now present to the clinic with indolent, nonaggressive tumors due to widespread use of prostate‐specific antigen (PSA) screening, identifying patients whose tumors will behave aggressively and whose tumors may recur following treatment both remain critical concerns.2

Temperature can be an important element of the physical microenvironment that can regulate cellular stem cell‐like properties, especially in the context of cancer.3 Temperature effects are modulated at the cellular level by stress‐response pathways that include heat‐shock and cold‐shock proteins (CSP).4, 5, 6 In contrast to heat‐shock proteins (HSP), which are induced by increased temperatures, CSP are induced by low temperatures but are downregulated when temperatures are elevated.7 The RNA‐binding motif protein 3 (RBM3) is a glycine‐rich protein with an RNA recognition motif capable of binding to both RNA and DNA and is one of the earliest proteins synthesized in response to cold shock.4 The role of RBM3 as a putative cancer biomarker was originally unraveled using an antibody‐based discovery approach (www.proteinatlas.org).8, 9, 10 RBM3 is an evolutionarily conserved CSP that has been shown to regulate the translation machinery and facilitate protein synthesis during hypothermic stress and in brain development, where it functions as an RNA chaperone to maintain RNA stability.11, 12 Recent studies have suggested that RBM3 has a potential proto‐oncogenic function and has been found to be upregulated in various types of human malignancies, including breast, prostate, ovarian, pancreatic, colorectal, and skin.13, 14, 15, 16 Further, RBM3 has been specifically shown to have elevated nuclear expression in many cancer types, and accumulating evidence has associated this expression with prolonged times to biochemical recurrence (BCR) and clinical tumor progression in the PCa setting.17 In a recent study that used a gene set enrichment analysis to elucidate the role of RBM3 in underlying biological processes, increased RBM3 expression was associated with DNA‐dependent replication, chromatin remodeling, and DNA damage response mechanisms.18

It is well established that nuclear alterations are a hallmark of many types of cancers, including PCa.19 During cancer progression, alterations to the genetic, epigenetic, and tumor microenvironmental landscape result in significant changes to the nucleus of cancer cells, including size and shape changes, textural changes, and spatial changes in the context of the surrounding cells. Genetic and epigenetic variations during cancer are thought to drive large‐scale nuclear alterations during cancer progression. Specifically, it is believed that three‐dimensional chromatin organizational alterations and altered methylation patterns have important ramifications for altered gene expression profiles and DNA stability in cancer cell nuclei. Through connections mediated by the nuclear lamina and nuclear matrix, three‐dimensional chromatin changes can contribute to nuclear pleomorphism in cancer cells.

Given that RBM3 has been hypothesized to play a role in critical nuclear functions such as chromatin remodeling, DNA damage response, and other posttranscriptional processes, we sought to (a) further quantify RBM3 expression in archival PCa samples; (b) develop a nuclear morphometric model to determine if measures of RBM3 expression levels and nuclear features could be used to predict disease aggressiveness and BCR; and (c) assess the correlation between RBM3 expression and nuclear architectural features, thus providing an insight into how RBM3 is involved in nuclear alterations.

2. MATERIALS AND METHODS

2.1. Immunohistochemical tissue microarray preparation

Two tissue microarrays (TMAs) were obtained from the Prostate Cancer Biorepository Network (PCBN) at the Johns Hopkins Hospital (JHH; Baltimore, Maryland). Samples for TMAs were obtained from prostatectomy samples obtained at JHH from 2002 to 2011 following Institutional Review Board approval. Samples from tumors included in the TMAs are selected using the index lesion (highest grade/largest lesion). Each TMA sample includes four replicates of 0.60‐mm punched cores. The TMAs were prepared using a Beecher MT1 manual arrayer (Beecher Instruments, Silver Spring, Maryland) in the TMAJ pathology core facility at the PCBN and processed as previously described.20, 21 In brief, TMA samples were incubated with anti‐RBM3 antibody produced in rabbit (Sigma; Product Number HPA003624) at a 1:200 dilution for 1 hour at room temperature followed by an overnight incubation at 4°C. Secondary antirabbit antibody conjugated with horseradish peroxidase and adsorbed with IgG (KPL/SeraCare, Milford, Massachusetts) was incubated with the samples at a 1:200 dilution for 1 hour at room temperature. The slides were subsequently stained with hematoxylin and cover slipped with Vector Mount. All tissue processing procedures were uniformly performed for each TMA included in the study.

2.2. PCa cohort

The stained slides from each TMA were reviewed and graded by a series of expert uropathologists. Among the 80 cases in this cohort (40 cases from each TMA), 10 were scored as Gleason sum (GS) 3 + 3 disease, 20 were scored as GS 3 + 4, 20 were scored as GS 4 + 3 disease, and 30 were scored as GS ≥8 disease. Further, 20 out of 80 total cases exhibited BCR, which was defined as having a PSA level increase greater than 0.2 ng/mL following surgery. Patients with BCR had a mean follow‐up time of 8.9 years (median: 10 years; range: 3‐13 years). Patients that did not experience BCR had a mean follow‐up time of 5.3 years (median: 3 years; range: 1‐12 years). The complete and detailed clinicopathologic demographics of the TMA are found in Table 1.

Table 1.

Summary of the clinicopathologic data for the PCa cases in the tissue microarray (TMA) cohort

TOTAL PCa Cases in TMA Cohort: N = 80
Variable	N	%
Gleason score
3 + 3	10	12.5
3 + 4	40	50.0
4 + 3	20	25.0
≥8	10	12.5
TNM staging
2	28	35.0
3A	31	38.8
3B	8	10.0
Unknown	13	16.2
Biochemical recurrence
No recurrence	49	61.3
Biochemical recurrence	20	25.0
Increase in PSA > 0.2 ng/mL	12	15.0
Local recurrence of PCa	1	1.3
Distant metastasis	2	2.5
Both local recurrence and distant metastasis	2	2.5
Increase in PSA > 0.2 ng/mL after surgery, decrease in PSA through radiation treatment	3	3.7
Death from non‐PCa related cause	1	1.3
No follow‐up data	10	12.5

2.3. Nuclear morphometric analysis

The TMAs in this study were scanned with Aperio Scanning Microscope (Leica Biosystems, Wetzlar, Germany) at 20X, and the .tiff images were extracted using Aperio ImageScope Software with the no compression option. Each of the 0.6‐mm TMA core images at 20X magnification takes up 1670 × 1670 pixels and is considered to be sufficient for enclosing the prostate cells and glandular structure information. A customized image processing algorithm was created using semiquantitative methods supported by the ImagePro Premier 9.1 software package (Media Cybernetics, Rockville, Maryland) to segment and quantify RBM3‐stained nuclei in each TMA core image. The algorithm was trained to identify nuclear features from RBM3‐positive glandular epithelial cell nuclei by using band‐pass filtering for size, color, and intensity from each core in the 80‐case set. Thus, in total, with four cores per case and 80 cases in total, our algorithm was run on a total of 320 samples that included thousands of nuclei. In RBM3‐positive nuclei, 22 different nuclear features were determined and recorded, including measures of nuclear size and shape and RBM3 staining intensity (full descriptions of the parameters can be found in Supplementary Table 1). One software‐derived feature, Integrated Optical Density (IOD), was used as a measure of antibody staining intensity and was therefore used to estimate RBM3 protein content as previously performed.21, 22 Each of the features was averaged from each of the four cores per case to result in final data for each case. Using a subset of the most prognostic features (as determined using univariate logistic regression modeling), a multivariate logistic regression (MLR) model was generated to determine if measures of RBM3 staining intensity and nuclear structural changes could be used to distinguish more aggressive (GS 4 + 3 and GS ≥8 disease) PCa from less aggressive (GS 3 + 3 and GS 3 + 4) PCa. The same set of significant features was then used in a separate MLR model to determine if they could be used to identify patients who exhibited BCR against those that did not.

2.4. Statistical analysis

Prior to generating the MLR models to identify aggressive PCa and BCR cases, univariate logistic regression was first performed to identify which features were individually prognostic of aggressive PCa. Using the features that showed univariate significant prognostic capabilities (P < .05), the MLR models for aggressive PCa prediction and BCR prediction were then generated. Because of the small sample sizes, leave‐one‐out cross validation (LOOCV) was used to provide insight on how the predictive capabilities of the feature set might behave with respect to an independent patient cohort to address issues such as model overfitting.

Statistical significance was defined as P < .05. Nonparametric Mann‐Whitney U tests were used for all group comparison analyses. Correlations of RBM3 staining intensity with the measures of nuclear structural changes that were in the MLR model were evaluated using the Spearman rank correlation coefficients. GraphPad Prism 7 (GraphPad Software, La Jolla, California) was used to generate all graphs. STATA 13.0 (StataCorp LLC, College Station, Texas) was used for all univariate and multivariate modeling and analysis as well as for the LOOCV approach.

3. RESULTS

3.1. RBM3 is downregulated in aggressive PCa and biochemical recurrent patient samples

Two TMAs containing a total of 80 PCa patient samples were stained for the RBM3 protein. Nuclear features were quantified using the ImagePro Premier 9.1 software. Images of the TMA cores confirmed the nuclear localization of the RBM3 protein (Figure 1A). In order to quantify RBM3 staining intensity in TMA images, the IOD measurement was identified as the most favorable because of its quantification of staining luminance over a given pixel area, an approach previously used by our group.20 The IOD measure of staining intensity from TMA images was used as a measure of RBM3 protein expression. Interestingly, when comparing IOD levels across samples of various Gleason scoring, RBM3 showed a significant downregulation in samples with a GS ≥8 compared with GS 3 + 3 samples (P = .0018), GS 3 + 4 samples (P = .0001), and GS 4 + 3 samples (P = .0006) (Figure 1B). Similarly, significantly lower levels of RBM3 were observed in BCR cases compared with those that showed no disease recurrence (P = .0377).

Figure 1.

RBM3 is downregulated in aggressive and recurrent prostate cancer. Because RBM3 was previously found to be localized to the nucleus, we examined RBM3 expression in the nucleus of PCa from TMA samples. To quantify RBM3 expression, a measure of staining intensity called Integrated Optical Density (IOD) was used. A, Representative images of RBM3 staining in TMA images. Scale bar (blue) = 0.25 mm; Scale bar (black) = 0.15 mm. B, RBM3 expression as represented by IOD varies significantly across Gleason scores. C, Represents the same data as in (B) but with GS 3 + 3 and GS 3 + 4 cases and GS 4 + 3 and GS ≥8 cases grouped together. D, RBM3 expression is significantly downregulated in biochemical recurrent PCa cases, BCR(+), against cases that did not exhibit biochemical recurrence, BCR(−); * P < .05

3.2. MLR model distinguishes aggressive PCa phenotype

From the original set of 22 nuclear morphometric features that were extracted from the RBM3‐positive cancer nuclei, univariate logistic regression revealed that in addition to the IOD for the quantification of the RBM3 staining intensity, 10 of the nuclear architectural features were individually significant (P < .05) for predicting an aggressive PCa phenotype (GS 4 + 3 and GS ≥8 cases). The full results from the univariate logistic regression for each feature can be found in Table S1. Table 2 highlights the significant features and their descriptions with respect to quantifying a measure of nuclear architecture or staining intensity. These 11 features were then combined to generate an MLR model for the prediction of aggressive PCa. In the raw data set, this 11‐feature MLR model was significantly predictive for distinguishing aggressive PCa cases from less aggressive cases and yielded a receiver operating characteristic area under the curve (ROC‐AUC) of 0.90 (P < .0001; 95% CI, 0.83‐0.97) (Figure 2A). LOOCV using the 11 features for prediction of aggressive PCa confirmed the predictive capability of the feature set, yielding a ROC‐AUC of 0.77 (95% CI, 0.65‐0.89) (Figure 2B), showing that the 11 features generate a robust model for predicting aggressiveness. A plot of the predictive probabilities from the LOOCV testing for each case showed significant separations between the less aggressive and aggressive PCa cases (P < .0001) (Figure 2C).

Table 2.

Summary of the 10 features included in the multivariate logistic regression (MLR) model for predicting aggressiveness and biochemical recurrence (BCR)

Feature	Category	Description
Axis major	Nuclear size, shape	Length of major axis of ellipse
Axis minor	Nuclear size, shape	Length of minor axis of ellipse
Bound box height	Nuclear size, shape	Height of the smallest possible rectangular that can fully enclose the nucleus
Bound box width	Nuclear size, shape	Width of the smallest possible rectangular that can fully enclose the nucleus
Clumpiness	Texture	Pixel level measurement that measures the texture of the nucleus by quantifying the fraction of pixels deviating from the average
Diameter max	Nuclear size, shape	Length of longest line adjoining two points of region's outline and passing through the centroid
Diameter mean	Nuclear size, shape	Average length of all diameters measured at various intervals around boundary
Diameter Min	Nuclear size, shape	Length of shortest line adjoining two points of region's outline and passing through the centroid
Heterogeneity	Texture	Pixel‐level measure of nuclear disorder
Integrated optical density	Staining luminance	Quantifies staining intensity due to RBM3 proteins in cancer nuclei

Figure 2.

RBM3 and nuclear architectural features predict aggressive prostate cancer. An 11‐feature MLR model was used to predict aggressive (GS 4 + 3 and GS ≥8) PCa cases from less aggressive (GS 3 + 3 and GS 3 + 4) PCa cases. A, Receiver operating characteristic area under the curve (ROC‐AUC) = 0.90 in the raw data set (P < .0001, 95% CI, 0.83‐0.97). B, ROC‐AUC from the leave‐one‐out cross validation (LOOCV) testing shows the strength of the model for predicting aggressiveness; ROC‐AUC = 0.77 (95% CI, 0.65‐0.89). C, Plot of the predicted probability from the case in the ROC curve from LOOCV testing in (B). ** P < .0001

A file showing the STATA code and STATA outputs from the generation of the MLR model and LOOCV testing to predict aggressiveness can be found in Figure S1.

3.3. The 11‐feature MLR model further distinguishes BCR patient samples

We then sought to determine if the same set of 11 features could be used to identify patients with BCR. MLR modeling for prediction of BCR yielded an ROC‐AUC = 0.92 (P = .0004; 95% CI, 0.85‐0.98) (Figure 3A) in the raw data set. LOOCV using the 11 features for prediction of BCR confirmed the predictive capability of the feature set, yielding an ROC‐AUC of 0.76 (95% CI, 0.64‐0.87). The predictive probability plot of the LOOCV results further shows the significant separation in the BCR‐negative and BCR‐positive cases (P = .0008) (Figure 3C). Additionally, we evaluated the ability of the 11‐feature model to predict BCR compared with the currently used Gleason scoring system (Figure 3D). The ROC‐AUC of the 11‐feature model showed comparable performance with the Gleason sum (ROC‐AUC = 0.73) for predicting BCR cases (P = .70) (Figure 3D).

Figure 3.

The 11‐feature MLR model further predicts biochemical recurrence. The 11‐feature MLR model was then used to predict biochemical recurrence (BCR). A, ROC‐AUC = 0.92 from the raw data set (P = .0004; 95% CI, 0.85‐0.98). B, ROC‐AUC from LOOCV testing shows the strength of the model for predicting BCR; ROC‐AUC = 0.76 (95% CI, 0.64‐0.87). C, Plot of the predicted probability from the cases in the ROC curve from the LOOCV set in (B). * P = .0008. D, Comparison of the LOOCV model against the currently used Gleason sum shows comparable results for predicting BCR in this sample set (P = .70)

A file showing the STATA code and STATA outputs from the generation of the MLR model and LOOCV testing to predict BCR can be found in Figure S2. A full document showing all code can be found in Figure S3.

3.4. RBM3 protein staining intensity correlates significantly with features from the MLR model

We evaluated the correlation between RBM3 staining intensity (as represented by the IOD feature) and the other 10 univariately significant features from the aggressive and BCR MLR models. We found that RBM3 expression exhibited significant associations with the features with varying degrees of strength of correlation relationships. This included area (Spearman rho = 0.92; P < .0001), axis major (Spearman rho = 0.82; P < .0001) (Figure 4A), axis minor (Spearman rho = 0.91; P < .0001) (Figure 4B), bound box height (Spearman rho = 0.82; P < .0001) (Figure 4C), bound box width (Spearman rho = 0.86; P < .0001) (Figure 4D), clumpiness (Spearman rho = 0.64; P < .0001) (Figure 4E), maximum diameter (Spearman rho = 0.85; P < .0001) (Figure 4F), mean diameter (Spearman rho = 0.92; P < .0001) (Figure 4G), minimum diameter (Spearman rho = 0.88; P < .0001) (Figure 4H), and heterogeneity (Spearman rho = 0.40; P = .0003) (Figure 4I). With the strong positive correlations observed here, similar relationships between RBM3 staining and the model parameters were observed as seen in Figure 1B (ie, as RBM3 expression decreased in higher Gleason score tumors, similar decreases in maximum diameter, for example, also occurred).

Figure 4.

RBM3 staining intensity correlates significantly with nuclear architectural features. Integrated Optical Density (IOD), a measure to quantify the staining intensity of the RBM3 protein expression, correlates significantly with the other 10 features included in the 11‐feature model for predicting aggressiveness and BCR: IOD versus Area: Spearman rho = 0.92; P < .0001. IOD versus Axis Major: Spearman rho = 0.82; P < .0001. IOD versus Axis Minor: Spearman rho = 0.91; P < .0001. IOD versus Bound Box Height: Spearman rho = 0.82; P < .0001. IOD vs Bound Box Width: Spearman rho = 0.86; P < .0001. IOD versus Clumpiness: Spearman rho = 0.64; P < .0001. IOD versus Diameter Max: Spearman rho = 0.85; P < .0001. IOD versus Diameter Mean: Spearman rho = 0.92; P < .0001. IOD versus Diameter Min: Spearman rho = 0.88; P < .0001. IOD versus Heterogeneity: Spearman rho = 0.40; P = .0003

4. DISCUSSION

The RBM3 protein has previously been shown to be upregulated in a wide range of cancer types, including prostate, melanoma, colorectal, breast, and ovarian14, 16, 17, 18, 23, 24 as compared with benign tissues. Overexpression of the nuclear RBM3 protein has been associated with significantly improved survival.16, 17 In particular, in the context of PCa, patients with tumors expressing high nuclear levels of RBM3 have a significantly prolonged time to BCR and clinical progression.17 Our study attempts to further clarify the role of RBM3 in PCa by examining its protein expression in a cohort of archival PCa samples with varying Gleason scores and BCR statuses. Further, given the high levels of nuclear localization for RBM3, we sought to determine if RBM3 protein levels in conjunction with other nuclear features could stratify patients that had aggressive (GS 4 + 3 and GS ≥8) PCa and BCR based on changes in RBM3 levels and altered nuclear architectural features. While previous studies have focused on RBM3 expression using only staining intensity by immunohistochemical studies, this study provides the first evaluation of RBM3 in the context of nuclear morphometric changes within human PCa cells. We show that these features can be combined to predict PCa aggressiveness and BCR status.

Previous immunohistochemical analysis of RBM3 first uncovered upregulated RBM3 expression in prostatic intraepithelial neoplasia and invasive prostate tumor samples17 as compared with benign samples. While this effect was not quantified across various Gleason scores, the same study showed high nuclear expression was specifically associated with improved clinical outcomes (ie, increased time to BCR and disease progression) compared with a low‐expressing RBM3 group. Our study examined RBM3 levels across PCa samples of varying Gleason scores and interestingly found that significantly lower RBM3 levels were discovered in higher Gleason score tumors. Further, RBM3 levels were significantly downregulated in PCa that displayed BCR versus those that did not. In contrast with our results is a study that showed high RBM3 expression was associated with high Gleason score and advanced tumor stage.23 Notably, RBM3 expression was also related to early BCR. The discrepancy in these results may be attributable to the fact that these previous studies used a categorical descriptor of RBM3 staining intensity. Our study used a continuous variable derived from the image processing software, IOD, which may substantiate our results in comparison with categorical measures of staining intensity from the human eye. While our PCa cohort lacked longitudinal survival data, our results suggest that the high RBM3 nuclear expression in low Gleason score tumors, which are typically less aggressive and associated with longer times to BCR, has validity. This result is further substantiated by a 2013 study that showed decreased RBM3 expression in metastatic PCa samples, a finding that suggests that RBM3 upregulation may be important for the early oncogenic events leading to PCa while RBM3 downregulation may have ramifications for progression and metastasis.24 Interestingly, a recently published report details RBM3 downregualtion in aggressive esophageal tumors and were statistically linked to adverse tumor features.25 Given that our study further clarifies that RBM3 has a tumor inhibitory role with lower expression levels in high Gleason score PCa, there now seems to be a growing consensus that RBM3 could be a putative biomarker for PCa. While RBM3 has not previously been shown to be present in the serum or urine of PCa patients, perhaps RBM3 staining of a diagnostic biopsy or a surgery sample could provide insight into tumor behavior and aid clinicians into predicting tumor aggressiveness or potential for BCR.

Our data show that RBM3 staining, in addition to nuclear morphometric features, is predictive of aggressive PCa and BCR. The prediction of BCR using our novel 11‐feature model is slightly better (though not statistically significant) to the currently used Gleason score. We believe a potential application for nuclear morphometry modeling is to aid pathologists in their grading of PCa. Currently, Gleason scoring, which is based on a pathologist's subjective grading, is the best way to predict PCa progression. Even among specialists and expert uropathologists, discrepancies in determining tumor grade and Gleason scoring are significant as PCa is known to have high glandular and nuclear heterogeneity.26 Manual grading remains a laborious process, often resulting in poor score reproducibility between pathologists. A machine‐assisted approach would be advantageous in that it would select the same pixels, in the same image, even from consecutive cores if utilizing the same algorithm repeatedly. Further, even a consecutive sections might also show some variation in grading from a previous one. Computer‐aided approaches can ameliorate many of these challenges by providing an objective algorithm that can track both biomarker staining (not limited to RBM3), subvisual nuclear changes that a pathologist may not recognize at eye level, and score reproducibility.27 Recent advances in digital pathology, including the application of machine learning to pathology, has resulted in accurate diagnostic and prognostic predictions in the setting of PCa (extensively reviewed in ref 28). Further, our 11‐feature can add clinical value by providing a faster, more efficient method to predict tumor progression (both aggressiveness and potential BCR) and aid pathologists in their assessment of tissue samples. Our data show that it compares favorably with the currently used method and could be a feasible method to use clinically.

Our study was limited in that our overall sample size (N = 80), and the number of samples that exhibited BCR (N = 20) was small. This prevents large‐scale applicability of the results beyond the samples included in the study. Given the lack of an independent data set to test our model, this study warrants further external validation in a larger TMA set. However, data were derived from 320 cores (80‐case TMA set with four cores per case) with our algorithm quantifying thousands of nuclei per core, resulting in the generation of this model from a large amount of data. Additionally, we employed LOOCV to address the potential overfitting that may have occurred because of the small sample size of the cases and to provide a measure of internal validation. Our LOOCV results showed the strength of our 11‐feature model for predicting both aggressiveness and BCR. A decrease was observed in the ROC‐AUC from the raw data to the LOOCV results for both the aggressiveness and BCR models. We attribute this to the small sample sizes, but the LOOCV tests still exhibit strong predictive ability. In the case of the BCR model specifically, the model still compared favorably with the currently used Gleason sum for predicting which patients will experience BCR and adds the efficiency that a computational approach adds over the currently used manual grading system. In addition to the sample sizes, the lack of survival data with our samples limits comparisons with existing studies that examine RBM3 in the context of time to BCR or clinical progression. Further studies to assess our 11‐feature model predicting overall survival or time to BCR are warranted.

5. CONCLUSIONS

RBM3 protein staining, in addition to nuclear morphometric features, is prognostic of aggressive PCa and BCR. RBM3 shows reductions in expression levels in aggressive and biochemically recurrent PCa.

ETHICS STATEMENT

Tissue microarrays were obtained from the Prostate Cancer Biorepository Network (PCBN) at the Johns Hopkins Hospital (JHH; Baltimore, Maryland). Samples for TMAs were obtained from biopsy samples obtained at JHH from 2002 to 2011 following IRB approval. This study was conducted in accordance with the Declaration of Helsinki. Informed consent was obtained from patients prior to surgical sample collection.

CONFLICT OF INTEREST

All authors declare that they have no competing interests.

AUTHOR CONTRIBUTIONS

All authors had full access to the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Conception and design, G.Z. and R.W.V.; Acquisition of data, N.M.C., G.Z., and C.D.; Analysis and interpretation of data, N.M.C., G.Z., M.C.M., and R.W.V.; Drafting the manuscript, N.M.C., P.K., and R.W.V.; Final approval of the version to be published, N.M.C., G.Z., M.C.M., C.D., P.K., and R.W.V.

Supporting information

Supplementary Figure 1: Raw results from multivariate logistic regression model for aggressiveness with leave‐one‐out cross validation

Supplementary Figure 2: Raw results from multivariate logistic regression model for biochemical recurrence with leave‐one‐out cross validation

Supplementary Figure 3: Code used to generate all STATA modeling

Supplementary Table 1: Descriptions of the 22 nuclear parameters quantified by the algorithm

Supplementary Table 2: Raw results of univariate logistic regression analysis.

Carleton NM, Zhu G, Miller MC, Davis C, Kulkarni P, Veltri RW. Characterization of RNA‐binding motif 3 (RBM3) protein levels and nuclear architecture changes in aggressive and recurrent prostate cancer. Cancer Reports. 2020;3:e1237. 10.1002/cnr2.1237

DATA AVAILABILITY STATEMENT

Availability of Data and Materials: All data and analyses (including statistical software code) conducted during this study are available from the corresponding author on reasonable request.