A novel assessment of whole-mount Gleason grading in prostate cancer to identify candidates for radical prostatectomy: a machine learning-based multiomics study

Abstract

Purpose: This study aims to assess whole-mount Gleason grading (GG) in prostate cancer (PCa) accurately using a multiomics machine learning (ML) model and to compare its performance with biopsy-proven GG (bxGG) assessment. Materials and Methods: A total of 146 patients with PCa recruited in a pilot study of a prospective clinical trial (NCT02659527) were retrospectively included in the side study, all of whom underwent ⁶⁸Ga-PSMA-11 integrated positron emission tomography (PET) / magnetic resonance (MR) before radical prostatectomy (RP) between May 2014 and April 2020. To establish a multiomics ML model, we quantified PET radiomics features, pathway-level genomics features from whole exome sequencing, and pathomics features derived from immunohistochemical staining of 11 biomarkers. Based on the multiomics dataset, five ML models were established and validated using 100-fold Monte Carlo cross-validation. Results: Among five ML models, the random forest (RF) model performed best in terms of the area under the curve (AUC). Compared to bxGG assessment alone, the RF model was superior in terms of AUC (0.87 vs 0.75), specificity (0.72 vs 0.61), positive predictive value (0.79 vs 0.75), and accuracy (0.78 vs 0.77) and showed slightly decreased sensitivity (0.83 vs 0.89) and negative predictive value (0.80 vs 0.81). Among the feature categories, bxGG was identified as the most important feature, followed by pathomics, clinical, radiomics and genomics features. The three important individual features were bxGG, PSA staining and one intensity-related radiomics feature. Conclusion: The findings demonstrate a superior assessment of the developed multiomics-based ML model in whole-mount GG compared to the current clinical baseline of bxGG. This enables personalized patient management by identifying high-risk PCa patients for RP.

Keywords: Gleason grading; PSMA; machine learning; multiomics; prostate cancer.

Figure 1

Flowchart of the study cohort. PCa: prostate cancer; TMA: tumor microarray; FFPE: formalin-fixed paraffin-embedded; SUV: standardized uptake value.

Figure 2

Genomics profile indicates the heterogeneity of the 51 investigated biological pathways in 65 PCa patients. The top bar shows TMB and CNV burden distribution. The top panel shows the correlation of genes/pathways mutation profile with ISUP groups. The top dendrogram shows the clustering patterns of genes/pathways based on their mutation profiles. TMB: tumor mutational burden; CNV: copy number variant; ISUP: International Society of Urological Pathology; PCa: prostate cancer.

Figure 3

Representative images of H&E staining and PSA staining on TMA slides revealing less PSA expression when PCa tissue is more aggressive. A. Representative images of H&E staining for each ISUP grade: a. Patient 1, GS 6 (3+3); b. Patient 2, GS 7 (3+4); c. Patient 3, GS 7 (4+3); d. Patient 4, GS 8 (4+4); e. Patient 5, GS 9 (4+5); according to ISUP consensus 2019. B. Representative images of PSA expression in each ISUP grade core. a. Patient 1, high PSA expression; b. Patient 2, relatively high PSA expression; c. Patient 3, moderate PSA expression; d. Patient 4, relatively low PSA expression; d. Patient 5, negative PSA expression. The corresponding H&E core and PSA core are from the same cylinder of the same patient. The scale bars of the overview core and enlarged details are 400 μm and 100 μm respectively. C. The maximum H-score of PSA is significantly different between ISUP high and low groups (p < 0.0001). TMA: tumor microarray; PSA: prostate-specific antigen; ISUP: International Society of Urological Pathology; GS: Gleason score.

Figure 4

ML Performance of the ISUP prediction in PCa. A. Performance comparison of the five ML algorithms (KNN, RF, SVM, LGR, XGB). Ranked by AUC, the RF model had the best performance. B. Overall comparison of different performance metrics between the RF model and ISUP derived from needle biopsy. C. Comparison of the mean permutation importance between different types of features. D. Detailed comparison of different performance metrics between the RF model and ISUP derived from needle biopsy. E. The top 10 performing features in ISUP prediction based on permutation importance over all cross-validation folds. F. SHAP importance of the eight features included in the final RF model trained on the entire dataset. Each dot represents a single patient and higher feature values are labeled as red while lower values are blue. The increasing positive SHAP values are indicative of the model's tendency to predict high ISUP while decreasing SHAP values indicate the tendency of the model to predict low ISUP. KNN: k-nearest neighbors; RF: random forest; XGB: extreme gradient boosting; SVM: support vector machine; LGR: logistic regression; AUC: area under the curve; ACC: accuracy; SNS: sensitivity, SPC: specificity; PPV: positive predictive value; NPV: negative predictive value; ML: machine learning; SUVmean: mean standardized uptake value; PSAmax: maximum H-score of PSA expression on three cores of TMA slides; PSAavg: average H-score of PSA expression on three cores of TMA slides.

Figure 5

Proposed diagnostic flowchart for prostate cancer (PCa) management. PSAmax: maximum H-score of PSA expression on three cores of TMA slides; PSA: prostate-specific antigen.