A Novel Artificial Intelligence–Powered Method for Prediction of Early Recurrence of Prostate Cancer After Prostatectomy and Cancer Drivers

PURPOSE

To develop a novel artificial intelligence (AI)–powered method for the prediction of prostate cancer (PCa) early recurrence and identification of driver regions in PCa of all Gleason Grade Group (GGG).

MATERIALS AND METHODS

Deep convolutional neural networks were used to develop the AI model. The AI model was trained on The Cancer Genome Atlas Prostatic Adenocarcinoma (TCGA-PRAD) whole slide images (WSI) and data set (n = 243) to predict 3-year biochemical recurrence after radical prostatectomy (RP) and was subsequently validated on WSI from patients with PCa (n = 173) from the University of Wisconsin-Madison.

RESULTS

Our AI-powered platform can extract visual and subvisual morphologic features from WSI to identify driver regions predictive of early recurrence of PCa (regions of interest [ROIs]) after RP. The ROIs were ranked with AI-morphometric scores, which were prognostic for 3-year biochemical recurrence (area under the curve [AUC], 0.78), which is significantly better than the GGG overall (AUC, 0.62). The AI-morphometric scores also showed high accuracy in the prediction of recurrence for low- or intermediate-risk PCa—AUC, 0.76, 0.84, and 0.81 for GGG1, GGG2, and GGG3, respectively. These patients could benefit the most from timely adjuvant therapy after RP. The predictive value of the high-scored ROIs was validated by known PCa biomarkers studied. With this focused biomarker analysis, a potentially new STING pathway–related PCa biomarker—TMEM173—was identified.

CONCLUSION

Our study introduces a novel approach for identifying patients with PCa at risk for early recurrence regardless of their GGG status and for identifying cancer drivers for focused evolution-aware novel biomarker discovery.

INTRODUCTION

Significant advances in targeting genomic drivers have resulted in durable clinical benefits in some tumor types. In a majority of tumors, however, we are yet to find unique targets in clones driving disease recurrence and resistance. Features at the cellular and molecular levels in prostate cancer (PCa) pathogenesis and progression in each patient are divergent and complex.^1-3 This complicates the development of clinically relevant biomarkers and the timing and sequence determination for drug selection strategies to dynamically combat resistance mechanisms in the progressing PCa.⁴ The molecular signals from a handful of cells driving tumor progression are often lost in the milieu of the whole tissue section homogenate. In addition, it is increasingly realized that underlying pathobiology of the tumor cells and their spatial interactions with the tumor microenvironment (TME) are likely drivers of disease outcomes.^5,6 In addition, the variability in immune responses has emphasized the importance of incorporating critical TME features such as stromal structures, tumor infiltrating lymphocytes, and angiogenesis for better understanding cancer evolution and prediction of response.^7,8 The areas within and around heterogenous tumor that drive progression may exhibit a unique cancer cell and TME profile with distinct features that are yet to be fully elucidated.

CONTEXT

• Key Objective

• Can artificial intelligence (AI)–powered method identify driver regions in prostate cancer (PCa) for the prediction of early recurrence and targeted biomarker discovery?

• Knowledge Generated

• Our AI-powered platform can extract visual and subvisual morphologic features from hematoxylin and eosin–stained whole slide images to identify driver regions in PCa (regions of interest) after prostatectomy. The regions of interest were ranked with AI-morphometric scores, predicting 3-year biochemical recurrence with high accuracy (area under the curve, 0.78) and enabling targeted biomarker analysis and discovery. A potentially new STING pathway–related PCa biomarker—TMEM173—was identified.

• Relevance

• The capability and high accuracy of our AI-based approach in distinguishing aggressive from indolent PCa, even within PCa with low or intermediate Gleason grades, have the potential to improve PCa risk stratification and clinical management and discover new biomarkers and targets.

These complex processes in PCa progression are reflected in subtle morphologic variations that are not fully captured by the Gleason Grading System (GGS) and are not apparent to the human eye. With the development of digital whole slide image (WSI) and computational pathology, especially application of artificial intelligence (AI), extraction of subvisual morphometric features using deep convolutional neural networks (DCNNs) from tissue samples is now a realistic possibility. The task of predicting outcomes from WSI is particularly challenging because of the large size of these images (approximately 100,000 × 100,000 pixels) and the fact that the morphologic features associated with outcomes, being unknown, could be present in any part of the imaged tissue. A few early success in this field includes predicting microsatellite instability status in colorectal cancer, gastric cancer,^9,10 and a few others.^11,12 Such morphologic features in PCa have eluded detection thus far.

These limitations led to an intense search for biomarkers that can be used to guide clinical management of PCa using molecular biology and bioinformatics.^6,13 PCa-related molecular markers, such as Ki-67, p53, MYC, PTEN, Rb, AR, and ERG, are well studied for disease stratification.^14-20 On the basis of the results of the whole tissue genomic analysis, a few sets of genomic markers included in the scoring systems of DECIPHER, POLARIS, Oncotype Dx, etc have been identified, which add to the GGS for PCa prognosis and have recently entered clinical trials. Most of these genome-based scores, however, marginally improved the prognostic accuracy of GGS thus far. Proteogenomic biomarker identification has suffered from the current method of analysis, where driver genes/proteins in a handful of cells in the tissue are often lost in the milieu of the whole tissue extract.

Data have shown that the application of newer therapies sooner in the PCa disease course have improved patient outcomes. For example, PROSPER,²¹ SPARTAN,²² and ARAMIS²³ trials adding enzalutamide, apalutamide, or darolutamide, respectively, to standard androgen deprivation therapy showed improvements in multiple clinical outcomes including overall survival, time to metastases, time to prostate-specific antigen (PSA) progression, and time to chemotherapy. Moving to an earlier state of PCa, the TITAN²⁴ trial examined the addition of apalutamide for metastatic castration-sensitive PCa and showed an advantage for the addition of the apalutamide. The STAMPEDE²⁵ trials in a subset analysis showed improved outcomes with the addition of radiation in patients with low volume metastatic disease. Therefore, it is of paramount importance to be able to stratify patients with PCa for targeted therapy in a timely fashion to improve patients' survival.

We report here that DCNNs can extract morphologic features from WSI tiles to identify driver regions in PCa for outcome prediction and targeted biomarker discovery.

This study was conducted under a research protocol approved by the institutional review board at the University of Wisconsin-Madison (ID# 2017-0670).

MATERIALS AND METHODS

Materials and Methods for Identifying Regions of Interest and Outcome Prediction

Training data set.

The model was trained using hematoxylin and eosin (H&E)–stained prostate radical prostatectomy (RP) WSI and outcome data from the National Institute of Health Cancer Genomic Atlas Prostatic Adenocarcinoma (TCGA-PRAD) data set.²⁶ The data set contains H&E WSIs and associated clinical and outcome data for 500 patients,—of which 243 patients were selected, who had biochemical recurrence (BCR) in < 3 years (n = 92) or did not have BCR for at least 3 years (n = 151; Data Supplement). None of the 243 patients had any peri-RP treatment (Fig 1).

FIG 1.

Study workflow. AI, artificial intelligence; AI-MS, AI-morphometric score; PRAD, prostatic adenocarcinoma; ROI, region of interest; RP, radical prostatectomy; TCGA, The Cancer Genome Atlas; WSI, whole slide image.

Validation data set.

The model was independently validated on 173 RP patients selected from the University of Wisconsin (UW)-Madison pathology archive and clinical database, which included patients who had BCR in < 3 years (n = 78) or did not have BCR for at least 3 years (n = 95; Fig 1). None of the patients had any peri-RP treatment regardless of margin status (Data Supplement) Patients' demographic and clinical information is shown in the Data Supplement. One representative H&E-stained slide from each of the 173 RP samples with the most aggressive tumor was scanned at 40× magnification (Leica Biosystems, Buffalo Grove, IL).

Our model architecture involved the following steps (Fig 2):

1. Input: The model took as input H&E WSIs and known clinical outcomes.

2. Tiling: The slides were divided into tiles of size (256 × 256 pixels)—the number of tiles varied from 10,000 to 100,000 tiles.

3. Feature extraction: A multiscale CNN Feature Extractor was trained using randomly sampled tiles of 64 × 64 at 40×, 256 × 256 at 40×, and 1,024 × 1,024 at 5× resolution to capture the nuclear detail, glandular context, and TME elements (see the Data Supplement for details).

4. Feature ranking module: A score was assigned to each image tile cluster on the basis of its correlation with patient outcomes. A high score (around 1) represented the tile that will show up mainly in patients with adverse outcomes, whereas a low score (around 0) represented good outcomes. These scores were generated using a set of weights that are learnt from known patient outcomes obtained retrospectively. The tile level scores were combined to a slide level morphometric score to predict patient disease outcomes. To train this model, we used the attention mechanism,²⁷ which uses a weighted average of tiles (low dimensional embeddings) where weights are determined by a neural network (see the Data Supplement for details).

5. Region of interest (ROI) profiling: The model finally outputs a heatmap per slide with patch level scores (as shown in Fig 2). High-ranked ROIs (defined by morphometric score [MS] 0.8-1, red in heatmap, PCa-r) and relatively lower ranked ROIs (defined by MS < 0.8, nonred in heatmap, PCa-nr) from the PCa area were selected for biomarker analysis.

6. Outcome prediction: The morphology scores were combined with known clinical features such as Gleason Grade, TNM staging, and margin status to come up with a combined classifier for patient outcome prediction.

FIG 2.

Platform workflow for outcome prediction. AUC, area under the curve; H&E, hematoxylin and eosin; ROI, region of interest; WSI, whole slide image.

Statistical analysis.

Univariate logistic regression analysis was conducted to evaluate whether the AI-morphometric score (AI-MS) predicts BCR as a binary outcome (BCR < 3 years v BCR > 3 years). The predictive power of the AI-MS was quantified by calculating the odds ratio (OR), which was reported along with the two-sided 95% CI, and the area under the curve (AUC) of the corresponding receiver operating characteristic curve. The AI-MS was correlated with the 3-year BCR rate—the null hypothesis is that the AUC is at most 0.6, which is expected using the Gleason Grade Group (GGG). This hypothesis was tested against the alternative hypothesis that the AUC is > 0.6. With the proposed validation sample size of 173, and assuming a 3-year BCR rate of 30%, an AUC of 0.75 or greater was detected with 90% power at the one-sided 0.05 significance level. The sample size calculation was performed in PASS 2020. The analyses described above were conducted using SAS software (SAS Institute Inc, Cary, NC) version 9.4.

Materials and Methods for Biomarker Analysis in the Ranked ROIs

Twenty RP cases from the validation cohort of sample size 173 were selected for analysis. The 20 cases were selected on the basis of PCa grade and the quality of the tissue sections and blocks. The case selection process is shown in the Data Supplement. Case information is presented in the Data Supplement.

Biomarkers and immune cells were selected for analysis on the basis of critical literature review and our pilot RNA and protein expression analysis in the ranked ROIs of three RP cases using GeoMx RNA and Protein Assay (NanoString Technologies, Seattle, WA). The cytotoxic T-lymphocyte–associated antigen-4 (CTLA4), CD74, and transmembrane protein 173 (TMEM173), known to influence tumor TME and consequently, immune response; six known PCa biomarkers such as androgen receptor (AR), programmed death ligand-1 (PD-L1), prostate-specific membrane antigen (PSMA), and PSA, and six immune cell markers (CD4, CD8, CD20, CD68, CD163, and CD57) for IHC analysis in the ranked ROIs in 20 RP cases. We also included cleaved caspase-3 (CC3), a marker for programmed cell death, and Ki67, a marker for cell proliferation, for analysis (Data Supplement). PSMA was also used as a prostate epithelial mask for tissue segmentation in the image analysis. Eight to ten representative areas of 0.08 mm² (each) from high-ranked ROIs (defined by MS 0.8-1, red in the heatmap, PCa-r) and relatively lower ranked ROIs (defined by MS < 0.8, nonred in the heatmap, PCa-nr) from the PCa area, respectively, were selected for analysis. Biomarkers were also analyzed in total PCa (PCa-t), PCa periphery (PCa-p, defined as an area ±0.5 mm around the PCa border), and benign prostatic tissue (BPT) regions (Fig 3). The PCa biomarkers were stained and analyzed in the selected ROIs, and the results were compared.

FIG 3.

A representative whole slide image (WSI) included for biomarker analysis. (A) H&E slide. (B) Heatmap. (C) Slide annotations for biomarker and immune cell analysis: larger irregular red circled areas denote total PCa (PCa-t), the area between the outer red and inner dotted line denote the PCa periphery (PCa-p), small red circles within PCa denote high-ranked PCa ROIs (red, PCa-r), small yellow/blue circles within PCa denote lower-ranked PCa ROIs (nonred, PCa-nr), large irregular green circled area denotes benign prostatic tissue (BPT), and small yellow/blue round circles within BPT denote representative BPT included for analysis. (D) Tissue segmentation: red highlights epithelium; green, stroma; and yellow, nonepithelium and nonstroma. H&E, hematoxylin and eosin; PCa, prostate cancer; ROI, region of interest.

Immunohistochemistry.

A representative block containing both PCa and non-PCa areas was selected from each case to produce unstained serial sections for biomarker analysis by the pathologist (W.H.). The antibody information is provided in the Data Supplement. Automated multiplex immunohistochemistry was performed in the Pathology TRIP Laboratory at UW. Automated multiplex immunohistochemistry was performed on the Ventana Discovery Ultra Biomarker Platform (Roche Diagnostics, Indianapolis, IN) following the standard protocol after optimization of each antibody. The combinations of multiplexed immunostaining are listed in the Data Supplement.

Biomarker analysis.

The stained slides were scanned using a Leica Biosystems Aperio AT2 scanner in the Pathology TRIP Laboratory at UW. Halo software modules (HALO v3.1.1076.405, Tissue Classifier, Multiplex IHC, and Spatial Analysis, Indica Labs, Albuquerque, NM) were used for biomarker analysis. Tissue was segmented into epithelial and stromal compartments (Fig 3D). The positive thresholds for biomarkers were carefully determined by assessing the background and true signals in the context of tissue morphology for each biomarker using the software. The percentage of positive cells with biomarker expression and immune cell densities (cell count/mm²) in the ROIs were compared (PCa-r v PCa-nr, PCa-t v PCa-p, and PCa-t v BPT).

Statistical analysis.

Comparisons between groups were conducted using a two-sample t-test or nonparametric Wilcoxon rank-sum test. The Bonferroni adjustment was used for experiment-wise error control when comparing multiple primary markers. The Benjamini-Hochberg False Discovery Rate method was used for controlling the false discovery rate when evaluating the full panel of markers.

RESULTS

Novel ROIs Predictive of Early Recurrence

Using the validation set, we measured the predictive power of the AI-MS in predicting 3-year BCR using the OR—at the 50th percentile split between low and high scores, we see an OR of 0.138 (0.071 to 0.272; P < .0001) for the high AI-MS score in having a recurrence-free survival of > 3 years in univariate analysis. When we split the score into low, moderate, and high and compared the 20th percentile at the low and high end, we observed an OR of 0.056 (0.017 to 0.187; P < .0001). Similar OR was observed in multivariate analysis with respect to the pT stage, GGG, and margin status (Table 1).

TABLE 1.

Univariate and Multivariate Analyses of the AI-Morphometric Score for Predicting the 3-Year Biochemical Recurrence Rate

We then evaluated the AI-MS using the AUC metric. We found that the AI-MS was superior to each of the pathologic parameters (GGG, pT stage, and margin status) in predicting 3-year BRC, with an AUC of 0.78 versus 0.62, 0.64, and 0.51 for each parameter, respectively. When AI-MS was combined with the pathologic parameters, the AUC was improved to 0.84. Importantly, the AI-MS was also able to stratify high-risk patients within the various Gleason Grade Groups. The AUCs were 0.84 and 0.81 for GGG2 and GGG3, respectively (Table 2).

TABLE 2.

Univariate and Multivariate Analyses of Model Features for Predicting 3-Year Biochemical Recurrence

ROI Validation and Biomarker Discovery

Significant differences in biomarker expression and immune cell distribution were found between the high and low AI-MS–ranked ROIs (labeled as PCa-r and PCa-nr, respectively). Specifically, Ki67 and PD-L1 protein expressions were found to be significantly higher and CC3 protein expression is significantly lower in PCa-r than in PCa-nr (P < .05). Examining the epithelial and stromal compartments of the ROIs separately, we found that the Ki67 index was significantly higher in the epithelial compartment of PCa-r than in the epithelial compartment of PCa-nr, whereas CD163 density was significantly higher and CTLA-4 density was significantly lower in the stromal compartment of PCa-r than in the stromal compartment of PCa-nr. TMEM173 expression was significantly higher in the epithelial compartment of both the PCa-r and total PCa-r than in the epithelial compartment of the PCa-nr and total PCa-nr, respectively (Data Supplement).

Comparing PCa-t (total PCa) with PCa-p (PCa periphery) or BPT regions, we observed that Ki67 index; AR, CD74, PSA, PSMA, and TMEM173 expressions; and densities of CD8-, CD20-, and CD163-positive cells were also significantly higher, whereas CC3 protein was significantly lower in the PCa-t than in the PCa-p and BPT (P < .001). Examining the epithelial and stromal compartments of PCa-t, PCa-p, or BPT separately, we found that there were significantly more CD8 and fewer CD57-, CD68-, and CTLA-4–positive cells in the epithelial compartment of the PCa-t than in that of the PCa-p or BPT and significantly more CD8, CD20, and CD163 and fewer CD57-, CD68-, and CTLA-4–positive cells in the stromal compartment of PCa-t than in that of PCa-p or BPT (P < .05; Data Supplement).

DISCUSSION

The major strength of our platform is its ability to identify top adverse patterns (ROIs) predictive of early recurrence, even in PCa with low or intermediate GGG, with a high degree of accuracy (78%, with ROI/AI-MS alone; 84% by adding pathologic parameters [Table 2]), by extracting both visual and subvisual features including morphologic features of PCa and TME from H&E-stained WSI. To the best of our knowledge, this is the first study using weighted ROIs mapped to actual patient outcome data to predict PCa outcomes. More importantly, the high-ranked ROIs allow us to pursue focused biomarker exploration, which has never been attempted in any cancer type. Interestingly, similar to the only such analysis in mesothelioma that has ever been published, we also found that the stromal compartment of PCa has a major impact on regulating disease aggressiveness. We observed that the stromal regions within the PCa-r had significantly higher CD163 and lower CTLA-4 cell densities compared with PCa-nr ROIs. There are significant differences in CD8, CD20, CD57, CD68, CD163, and CTLA-4 distribution in PCa-t, PCa-p, and BPT regions.

The results of biomarker analysis in the ranked ROIs and other tissue regions demonstrated the difference in molecular and immunophenotypic characteristics of the high-ranked ROIs compared with the low-ranked ROIs. Specifically, we found that four of the 15 biomarkers studied (CC3, Ki67, PD-L1, and TMEM173) were significantly different between the high- and low-ranked ROIs (PCa-r v PCa-nr) and between PCa-t and PCa-p and BPT regions. The validity of the ROIs was further confirmed by the significant differences in the expressions of AR, PSA, and PSMA in addition to CC3, Ki67, and PD-L1 between PCa-t and BPT, which were in concordance with the previously published studies.^28-32 We recognize that these observations were made in a small sample set. However, achieving statistical significance in the small sample set gives us confidence that a larger sample set will further strengthen the validation. Once validated in a larger sample set, this method will make a significant impact on PCa risk stratification. A more accurate risk stratification of PCa beyond GGS will have a far-reaching impact on management of localized PCa, with options ranging from active surveillance, localized treatment with radiation or surgery, and use of androgen deprivation therapy and/or chemotherapy in an adjuvant setting. In addition, this should have a high impact on potential enrollment of high-risk patients into clinical trials of novel and existing agents at an earlier stage of the disease to potentially improve disease outcomes. Patient enrichment strategies using enrollment of high-risk patients as identified by our AI platform into trials could result in shorter trials with a relatively smaller cohort of patients with potentially higher treatment effects of promising agents.

Our analysis also revealed a potentially new STING pathway–related PCa biomarker—TMEM173,³³ which correlated with PCa development and progression that has never been reported before. It has been well established that cellular oxidative stress plays a key role in PCa progression.^34-37 Karin and his coworkers have demonstrated that nuclear factor kappa B activation in PCa cells modifies PCa and its microenvironment and increases B-cell (CD20) densities in therapy-resistant PCa tissues.^38,39 Our AI-enabled, focused analysis identified significantly increased TMEM173 expression in PCa-r regions compared with the PCa-nr region—denoting the role of TMEM173 in PCa progression. This, combined with the high CD20 cell density observed in the PCa-r region, suggests a possible new pathway of PCa progression present in primary prostate tissues in some patients, who have relatively short PFS.

In summary, this study presents an AI-based innovative approach for identifying PCa driver regions for PCa risk stratification, outcome prediction, and novel biomarker/target detection, which might have a significant impact in clinical trial design and personalized patient management.

DATA SHARING STATEMENT

The source code used in this study will be available upon request after our patent application is approved by the US Patent and Trademark Office.

AUTHOR CONTRIBUTIONS

Conception and design: Wei Huang, Ramandeep Randhawa, Parag Jain, Shivaani Kummar, George Wilding, Hirak Basu, Rajat Roy

Financial support: Rajat Roy

Administrative support: Hirak Basu, Rajat Roy

Collection and assembly of data: Wei Huang, Ramandeep Randhawa, Parag Jain, Samuel Hubbard, George Wilding

Data analysis and interpretation: Wei Huang, Ramandeep Randhawa, Parag Jain, Jens Eickhoff, George Wilding, Hirak Basu, Rajat Roy

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/cci/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).