Metastatic castration-resistant prostate cancer (mCRPC) is defined by disease progression despite androgen deprivation therapy (ADT), with serum testosterone suppressed to castration levels (below 50 ng/dl or 1.7 nmol/l). It is one of the most aggressive forms of prostate cancer and carries elevated morbidity and mortality. Patients with metastatic prostate cancer on ADT typically progress to castration resistance after an average of 18 to 36 months. Frontline therapies, including oral hormonal agents such as enzalutamide and abiraterone, and taxane-based chemotherapies, frequently encounter resistance, making durable disease control exceptionally difficult.
This narrative review from the Aga Khan University and Western University examines how artificial intelligence (AI), machine learning (ML), and deep learning (DL) can be applied to overcome therapeutic resistance in mCRPC. The authors synthesize current literature on AI-driven approaches that leverage integrated datasets spanning genomics, proteomics, and clinical parameters. The review is structured around pathomics, radiomics, genomics, recurrence prediction, biomarker discovery, drug development, and registry management.
The paper provides important definitions distinguishing AI, ML, and DL. ML is a subset of AI where algorithms learn from data to make predictions without explicit programming. Specific ML techniques discussed include decision trees, k-nearest neighbors, logistic regression, Naive Bayes, principal component analysis, Random Forests, and eXtreme Gradient Boost (XGBoost). DL is a specialized form of ML employing multilayered artificial neural networks to extract complex features from vast datasets automatically, making it powerful for image analysis and molecular profiling.
The authors note that AI/ML/DL have already demonstrated significant contributions in breast, gastric, thyroid, and colorectal cancer research, improving diagnostic accuracy and predicting therapeutic responses. They argue that this track record in other cancer types provides a strong rationale for deploying these technologies against mCRPC, where treatment resistance remains the central clinical problem.
The review details multiple molecular mechanisms driving therapeutic resistance in mCRPC. AR pathway alterations: Mutations in the ligand-binding domain (LBD) of the androgen receptor (AR) can convert receptor antagonists into agonists. Up to 64% of mCRPC patients exhibit AR overexpression or amplification. AR splice variants such as AR-V7, which lack the LBD, remain constitutively active and directly fuel resistance. AR overexpression may also arise from coregulator overexpression, increased receptor stability, or gene amplification leading to enhanced phosphorylation or acetylation of histones at AR enhancers. Glucocorticoid receptor (GR) upregulation has been implicated in enzalutamide resistance, since AR and GR share a chromosomal binding site.
Efflux proteins and apoptosis evasion: Overexpression of P-glycoprotein and multidrug resistance proteins expels chemotherapy drugs from tumor cells, reducing drug efficacy. Tumor cells also evade destruction by inhibiting apoptosis and repairing DNA damage, processes regulated by non-coding RNAs. Epigenetic changes, including covalent modifications of histone proteins, further contribute to disease progression.
Neuroendocrine differentiation and lineage plasticity: Tumor cells can undergo neuroendocrine differentiation, gaining increased adaptability through lineage plasticity, which is proposed as a mechanism for both therapeutic resistance and disease progression. PI3K/AKT activation: This pathway provides alternative survival signaling for tumor cells, promoting proliferation even when the AR axis is blocked.
Tumor microenvironment: The mCRPC microenvironment is highly immunosuppressive, driven by inflammatory signals and hypoxia. Regulatory T cells, M2 macrophages, and myeloid-derived suppressor cells dominate, impeding anti-tumor activities of dendritic cells, NK cells, B cells, and cytotoxic T cells. PARP inhibitor resistance: Aberrations in homologous recombination repair (HRR) genes such as BRCA1, BRCA2, and PALB2 along with ABC efflux transporters reduce PARP inhibitor efficacy. Lu-PSMA resistance: Low linear energy transfer causes primarily single-strand DNA breaks, and tumors with low or heterogeneous PSMA expression may not receive adequate radiation doses.
The review includes an extensive Table 1 summarizing landmark clinical trials in mCRPC. Abiraterone: The COU-AA-301 trial (n = 1,195) showed abiraterone plus prednisone improved median overall survival versus placebo (14.8 vs. 10.9 months; HR = 0.65; p < 0.001). COU-AA-302 (n = 1,088) demonstrated longer radiographic progression-free survival (16.5 vs. 8.3 months; HR = 0.53; p < 0.001) and a 25% decrease in risk of death (HR = 0.75; p = 0.01).
Enzalutamide: The AFFIRM trial (n = 1,199) showed improved median overall survival (18.4 vs. 13.6 months; HR = 0.63; p < 0.001). PREVAIL (n = 1,717) demonstrated an 81% reduction in risk of radiographic progression (HR = 0.19; p < 0.001) and 29% reduction in death risk (HR = 0.71; p < 0.001). The STRIVE trial showed enzalutamide improved progression-free survival versus bicalutamide (19.4 vs. 5.7 months; HR = 0.24; p < 0.001), and TERRAIN confirmed similar benefit (15.7 vs. 5.8 months; HR = 0.44; p < 0.0001).
Taxanes: TAX 327 (n = 1,006) established docetaxel every 3 weeks as standard, with median survival of 19.2 months versus 16.3 months for mitoxantrone (HR = 0.79; p < 0.004). The TROPIC trial (n = 775) showed cabazitaxel improved survival over mitoxantrone (15.1 vs. 12.7 months; HR = 0.70; p < 0.0001). PARP inhibitors: PROfound showed olaparib extended imaging-based PFS in BRCA1/2 or ATM-mutated patients (7.4 vs. 3.6 months; HR = 0.34; p < 0.001). TRITON3 demonstrated rucaparib prolonged imaging-based PFS in the BRCA subgroup (11.2 vs. 6.4 months; HR = 0.50; p < 0.001). PROpel showed abiraterone plus olaparib improved imaging-based PFS (24.8 vs. 16.6 months; HR = 0.66; p < 0.001).
Radioligand therapy: The VISION trial (n = 831) demonstrated that 177Lu-PSMA-617 plus standard care improved both imaging-based PFS (8.7 vs. 3.4 months; HR = 0.40; p < 0.001) and overall survival (15.3 vs. 11.3 months; HR = 0.62; p < 0.001). ALSYMPCA showed radium-223 improved median OS (14.0 vs. 11.2 months; HR = 0.70; p = 0.002). Despite these advances, the authors emphasize that overall survival rates remain low, underscoring the need for AI-driven strategies to improve treatment selection and overcome resistance.
The review examines AI applications in computational pathology (pathomics). While prostate biopsy and Gleason scoring historically serve as the diagnostic cornerstone for localized prostate cancer, their role extends into mCRPC through molecular profiling of metastatic biopsies. The authors note that 58% of patients in one study harbored theoretically actionable mutations, yet only a small subset received matched therapies, highlighting a gap between molecular findings and clinical application that AI could help bridge.
Morozov et al meta-analysis: AI-assisted histological assessment across 8,000 prostate biopsies and 458 prostatectomy cases achieved diagnostic accuracies ranging from 83.7% to 98.3%. Sensitivity for diagnosing prostate cancer exceeded 90% (range 87% to 100%), while specificity varied between 68% and 99%. PANDA challenge (Bulten et al): Using over 10,000 digitized biopsies, AI algorithms achieved agreements with expert uropathologists of kappa = 0.862 (95% CI, 0.840-0.884) on the US validation set and kappa = 0.868 (95% CI, 0.835-0.900) on the European validation set, demonstrating cross-continental reproducibility.
DeepDx Prostate: Jung et al validated this AI diagnostic tool using 593 whole-slide images (130 normal, 463 adenocarcinomas) against Gleason scores assessed by three expert uropathologists. DeepDx demonstrated cancer detection accuracy comparable to original pathology reports but with higher concordance with reference grade groups and Gleason scores. Paige Prostate: This clinical-grade AI tool was evaluated on 105 prostate core needle biopsies. Four pathologists initially achieved 95.0% diagnostic accuracy unaided. With AI assistance, there was a reduction in atypical small acinar proliferation reports, fewer immunohistochemistry studies, fewer second opinions needed, and less time required for reading and reporting slides.
The authors emphasize that while these advances have primarily focused on localized disease, their integration into molecular pathology pipelines for evaluating metastatic lesions represents a promising avenue for precision oncology in mCRPC. Early detection through AI-assisted pathology increases the likelihood of successful localized treatment, potentially preventing or delaying progression to mCRPC.
Radiomics: Wang et al investigated texture features from multiparametric MRI (mp-MRI) to predict bone metastases in prostate cancer. They analyzed 976 features from T2-weighted and dynamic contrast-enhanced T1-weighted MRI scans of 176 patients. Combining information from both MRI sequences showed improved prognostic performance compared to using either sequence alone (p < 0.01 for association with bone metastasis). TRUS AI model: Li et al developed an AI model using 1,696 two-dimensional transrectal ultrasonography (TRUS) images from 142 patients. They trained a ResNet50 network with three classification models: original image (Whole), biopsy needle tract (Needle), and combined image using Feature Pyramid Networks (FPN). The FPN model outperformed others and senior radiologists, achieving AUC of 0.667 and detecting 82.9% of prostate cancer cases versus 55.8% by radiologists.
Lymph node involvement (LNI): Faiella et al evaluated 16 studies on AI models for LNI detection and prediction. MRI-based AI models showed comparable LNI prediction accuracy to standard nomograms, while CT and PET-CT based models demonstrated strong diagnostic and prognostic capabilities. The authors note that radiomics' quantitative approach complements traditional imaging modalities, potentially enhancing both sensitivity and specificity.
Commercial genomic classifiers: Three ML-powered genomic tests are commercially available. Decipher analyzes gene expression to predict disease recurrence post-treatment. Oncotype DX Genomic Prostate Score assesses tumor aggressiveness to guide treatment decisions, especially post-surgery. Prolaris measures cell proliferation genes to predict disease aggressiveness. All three demonstrated rigorous quality criteria and clinical utility in prognostication of localized prostate cancer, providing additional prognostic information beyond standard clinicopathologic variables.
TMPRSS2-ERG detection: Dadhania et al used a DL model based on MobileNetV2 architecture to identify TMPRSS2-ERG rearrangement from digitized H&E slides of 392 cases. AUC of the ROC curves ranged between 0.82 and 0.85, with sensitivity of 75.0% and specificity of 83.1% at 20x magnification. Gene expression classifier: Mena et al developed an ML classifier using 47 genes from 550 samples (The Cancer Genome Atlas), validated on four external datasets (463 samples total). The Random Forest algorithm with majority class downsampling achieved an average sensitivity of 0.90, specificity of 0.80, and AUC of 0.84. Key genes identified included DLX1, MYL9, FGFR, CAV2, and MYLK.
Huang et al developed an AI-powered method for predicting 3-year biochemical recurrence (BCR) of prostate cancer post-prostatectomy. Using deep convolutional neural networks trained on whole slide images and patient data, the model extracted both visual and microscopic morphological features to identify predictive regions of interest (ROIs). The AI-derived morphometric scores achieved an AUC of 0.78 for predicting 3-year BCR, significantly outperforming the traditional Gleason Grade Group, which had an AUC of only 0.62. This represents a meaningful improvement in prognostic discrimination.
Liu et al conducted a systematic review evaluating AI effectiveness in predicting BCR after prostatectomy. AI algorithms incorporating radiological features demonstrated higher accuracy compared to those based solely on pathological or clinicopathological data. In some cases, AI outperformed traditional prediction methods entirely. However, the review identified significant variability in AI performance across studies due to differences in study designs, patient inclusion criteria, and follow-up data, indicating the need for standardized approaches.
Eminaga et al developed an AI-based system using histology images to predict prostate cancer recurrence, validated with multi-institutional datasets of 2,647 patients followed over 10 years. The system demonstrated superior predictive accuracy compared to existing grading systems and successfully categorized prostate cancer into four distinct risk groups, with high consensus observed among pathology experts.
Biomarker discovery: Through focused analysis of high-scored ROIs, Huang et al also identified TMEM173 as a potential new prostate cancer biomarker. TMEM173 is associated with the STING (stimulator of interferon genes) pathway, a signaling cascade involved in innate immune activation. This discovery exemplifies how AI-driven morphological analysis of whole slide image data can uncover novel biomarkers that might be missed by conventional pathological examination, opening new avenues for both prognostication and therapeutic targeting.
CancerOmicsNet: This AI system uses deep graph learning to predict how well multitargeted kinase inhibitors will work against various tumor types. It achieved an AUC of 0.83 in predicting therapeutic effects, outperforming other computational approaches. By modeling the complex interactions between drug molecules and cancer cell signaling networks, CancerOmicsNet accelerates the identification of promising drug candidates without requiring exhaustive laboratory screening.
AndroPred: Gagare et al developed ML and DL algorithms to predict androgen receptor (AR) inhibitors using a dataset of 2,242 compounds. The DL-based prediction model outperformed all other approaches, achieving accuracies of 92.18% on the training dataset and 93.05% on the test dataset. Given that AR signaling is central to mCRPC progression and resistance, tools like AndroPred could expedite discovery of novel AR-targeting agents. The authors note that further experimental assay validation is needed to establish full reliability.
TraRe (Blatti et al): This Bayesian ML tool used RNA sequencing to determine the response to abiraterone in mCRPC across different transcriptional networks. TraRe identified a transcriptional module linked to immune response, enriched with rewired regulatory networks in treatment responders versus non-responders. Key transcription factors included CEBPE, GATA1, KLF1, and MYB (involved in granulocyte differentiation, erythroid development, and hematopoietic pathways), along with TAL1 and NFE2 as pivotal regulators. TraRe also identified factors like SREBF2, SMAD7, SOX8, and SNAI2 driving transcriptional modules implicated in cancer progression. Three transcription factors (ELK3, MYB, and MXD1) functioned differently in non-responders, underscoring the complexity of transcriptional rewiring in mCRPC.
Registry management (CAPRI-3): In the Netherlands, NLP-based text-mining software is widely used across hospitals to semiautomatically gather data from electronic health records (EHRs). The automated data extraction achieved completeness and accuracy of 92.3% or higher, except for date fields and inaccessible data. Dosimetry prediction: Xue et al used ML to predict organ-level dosimetry for mCRPC patients receiving radioligand therapy, using PET imaging data and clinical information as inputs and Hermes software for dosimetry calculations. This approach could personalize radiation plans to maximize tumor kill while minimizing damage to healthy organs.
Treatment discontinuation prediction: Deng et al built ML models integrating data from three clinical trial cohorts to predict treatment discontinuation due to adverse effects in CRPC patients. Various models were tested, including linear regression, Cox regression, logistic regression, and nonlinear tree-based methods. Tree-based models performed best overall, with Random Forest (RF) showing the highest performance in two of three cohorts. The RF model identified patterns distinguishing patients who discontinue treatment, with laboratory test results and medical history having the greatest influence on predictions. This approach could help personalize treatment plans and reduce unwanted side effects.
Genomic mutation identification: Lin et al explored next-generation sequencing of circulating cell-free DNA (cfDNA) and applied ML algorithms to distinguish patients with castration-resistant from castration-sensitive prostate cancer (CSPC). The study revealed genomic alterations specific to mCRPC in the PI3K, RTK, G1/S, and MAPK signaling pathways, highlighting potential therapeutic targets for drug development.
Ongoing clinical trials (Table 3): Several active trials incorporate AI into prostate cancer management. These include: (1) a prospective validation of an AI-based diagnostic model analyzing whole-slide images for lymph node metastasis detection after radical prostatectomy; (2) AI-based measurements of tumor burden in PSMA PET-CT at Skane University Hospital (Sweden) to evaluate how total tumor burden in cm3 predicts overall survival; (3) IP6-CHAIROS, evaluating the Galen Prostate AI system for triaging pathology slides within the NHS; (4) PI-CAI Challenge, validating AI algorithm performance against radiologists for detecting clinically significant prostate cancer (csPCa) in MRI for patients with PSA levels of 3 ng/ml or higher; and (5) CeleScan-R, testing quickDWI (an accelerated whole-body diffusion-weighted MRI technique with DL denoising filters) that reduces acquisition times by up to 50%.
These trials represent the translational pipeline from AI research to clinical deployment. The diversity of approaches, spanning pathology, PET imaging, MRI, and histology, reflects the broad applicability of AI across the mCRPC diagnostic and treatment workflow. The outcomes of these trials will be critical for determining whether AI tools can be reliably integrated into routine clinical practice.
Data accessibility and infrastructure: ML and DL algorithms require vast training datasets, but hospital data is considered confidential and is rarely shared between institutions. Establishing AI-supporting platforms for storage, processing, and analysis demands expensive hardware, software, and skilled personnel, costs often too high for individual research teams. This is compounded by the need for continuous model improvement through training on expanding datasets.
Disease heterogeneity and annotation complexity: mCRPC involves a variety of genetic subtypes, including BRCA1/2-mutated, AR-V7 positive, and neuroendocrine-transformed tumors. AI models trained on localized prostate cancer datasets may perform poorly when applied to mCRPC, where resistance mechanisms differ dramatically from hormone-sensitive disease. Data labeling is particularly challenging in histopathology: while AI has shown promise in Gleason grading, detecting subtle features like neuroendocrine differentiation or intratumoral heterogeneity requires expert manual annotation and large sample diversity. Variations in reporting standards across biopsy pathology, MRI protocols, and molecular testing platforms further hinder cross-center generalizability.
Black box problem and accountability: DL algorithms often fail to provide justifications for their predictions, creating the "black box problem." This lack of transparency poses legal difficulties, as systems cannot justify potentially erroneous recommendations. In mCRPC specifically, treatment decisions hinge on nuanced variables such as AR splice variants, genomic alterations, and immunohistochemical profiles. Without clear justifications for AI outputs, clinicians may hesitate to trust these models. The review cites a cautionary example: an AI application for post-pneumonia risk assessment malfunctioned and incorrectly recommended discharging asthmatic patients, posing a direct health risk.
Data privacy and replication gaps: The 2018 DeepMind Health/Google acquisition exemplifies privacy concerns. The NHS shared data from over 1 million patients with DeepMind servers without obtaining patient consent. Current ML research also lacks consistency in reporting and fails to address the "last mile" of clinical evaluations through randomized trials. Many studies focus on technical performance with historical data rather than real-world clinical validation, and some replication studies have failed to reproduce original results. Low- and middle-income countries (LMICs) face additional barriers including limited internet access, poor infrastructure, financial constraints, and populations in rural areas with inadequate healthcare access, making AI deployment particularly challenging.
The authors conclude that AI algorithms have demonstrated clear potential in prostate cancer across multiple disciplines. In diagnostics, AI systems have achieved high accuracy in Gleason grading (83.7%-98.3%), cancer detection (sensitivity above 90%), and cross-continental reproducibility (kappa 0.86+). In drug discovery, tools like CancerOmicsNet (AUC = 0.83) and AndroPred (92-93% accuracy) show how computational approaches can accelerate identification of therapeutic candidates. In resistance profiling, TraRe and cfDNA-based ML classifiers are revealing the transcriptional and genomic underpinnings of treatment response.
The review highlights that AI should be regarded as a valuable tool to support and strengthen clinical judgment, not as a substitute for human decision-making. Medicine is a dynamic field with unpredictable circumstances that often require intuition and innate abilities. The fear that AI will replace the human workforce is, the authors argue, a flawed perspective. Instead, AI can help oncologists tailor personalized treatment plans for mCRPC, integrating genomic data with clinical parameters for optimized therapy selection.
Several areas require further development. Standardized multi-institutional datasets with diverse genetic subtypes need to be established. Explainability methods must be integrated into clinical AI tools so that predictions for AR splice variant status, HRR gene alterations, and treatment response can be transparently justified. Regulatory frameworks governing AI in healthcare must mature, particularly around patient consent and data governance. The ongoing clinical trials (IP6-CHAIROS, PI-CAI, CeleScan-R, and others) will provide critical evidence for whether AI tools can transition from research to routine clinical deployment.
For mCRPC specifically, the convergence of pathomics, radiomics, and genomics through AI platforms could enable truly integrated precision oncology. The ability to simultaneously analyze histological patterns, imaging features, and molecular profiles could help clinicians predict which patients will develop resistance to abiraterone, enzalutamide, PARP inhibitors, or Lu-PSMA therapy before that resistance clinically manifests. This proactive approach, rather than reactive switching of therapies after progression, represents the transformative potential of AI in mCRPC management.