AI in PCa Detection
Studies demonstrate that machine learning (ML) excels at identifying novel PCa biomarkers from public datasets like TCGA-PRAD and GSE series. RankProd and genetic algorithm-optimized artificial neural networks (ANNs) applied to microarray data from GSE6956, GSE8218, and others yielded a 15-gene signature with AUC 0.953 for distinguishing tumor from normal tissue; genes like FABP5, C1QTNF3, and LPHN3 were validated in PCa samples. Analysis of 22,014 genes from GSE60329 and GSE71016 selected 30 via Random Forest for an ANN model, achieving 88.9% accuracy on GSE46602. A deep learning framework pinpointed noncoding somatic mutations altering chromatin accessibility, linking them to androgen receptor (AR) pathways with 5-fold cross-validation, AUC 0.86.
Convolutional neural networks (CNNs) applied to 449 TCGA-PRAD histopathology slides predicted expression of 5,419 transcripts and the cell cycle progression (CCP) score for aggressiveness assessment. A Random Committee classifier on microarray data reached 95.1% accuracy in tissue classification. These findings underscore AI's diagnostic precision across neural networks, CNNs, and ensembles.
Prognostic Applications
Nine studies linked AI-derived signatures to metastasis, biochemical recurrence (bRFS), and survival. Integration of 10 ML algorithms (e.g., CoxBoost, survival-SVM) across 10 datasets created the Epithelial Cell Marker Gene Prognostic Signature (ECMGPS), stratifying patients with 1-, 3-, and 5-year bRFS AUCs up to 0.797; low-risk groups showed superior bRFS (p<0.001 in TCGA/MSKCC). A deep neural network (P-NET) differentiated primary PCa from castration-resistant PCa (CRPC) in 1,013 samples with AUC of 0.93, identifying MDM4 and FGFR1 as advanced disease markers.
ML on cfDNA from 341 metastatic patients pinpointed 16 castration resistance genes (AR most predictive), with AUC 0.74, PPV 94%, tied to RTK/MAPK/PI3K pathways. SPAG1 and PLEKHF2 amplifications were flagged for lymph node risk via ML clustering. A 15-gene prognostic index outperformed PSA/Gleason for 5-year OS (AUC 0.808).
Treatment Response Prediction
AI predicts therapy outcomes effectively. Preconditioned Random Forest on 606,563 SNPs from 324 patients forecasted post-radiation genitourinary toxicity (AUC 0.70). A deep neural network modeled AR mutation responses to antiandrogens (AUC 0.89), excelling in resistant phenotypes. A replication stress signature (RSS) via seven ML algorithms identified immunotherapy-responsive high-RSS tumors and targets like TOP2A. ECMGPS predicted anti-PD-L1 benefits in IMvigor210.
Clinical Readiness and Limitations
While AI promises personalized PCa care—superior to conventional stats in many metrics—implementation lags due to absent clinical validation, heterogeneity, and clinician-bioinformatician silos. No economic analyses exist, and signatures like ECMGPS need prospective trials. The review, mirroring sample formats on UroToday, advocates interdisciplinary efforts for theranostics and targets like irinotecan for high-RSS cases. Future integration could refine screening, prognosis, and therapies beyond PSA/Gleason limitations.
Written by: Andrey Bazarkin,1 Mark Taratkin,1 Stanislav Vovdenko,1 Aleksandr Androsov,2 Maria Balashova,3 Andrey Morozov,1 Alina Itskevich,4 Ekaterina Laukhtina,1,5 Evgenii Bezrukov,1 Nirmish Singla,6 Leonid Rapoport,1 Evgenii Shpot,1 Petr Glybochko1
- Institute for Urology and Reproductive Health, Sechenov University, Moscow, Russia
- Department of Pediatric Surgery, Division of Pediatric Urology and Andrology, Sechenov University, Moscow, Russia
- Department of Medical Genetics, Sechenov University, Moscow, Russia
- Institute for Clinical Medicine named after N.V. Sklifosovskiy
- Department of Urology, Comprehensive Cancer Center, Medical University of Vienna, Vienna, Austria
- Department of Urology, James Buchanan Brady Urological Institute, Johns Hopkins University School of Medicine, Baltimore, United States