Metastatic prostate cancer comes in many varieties, and accurately understanding cancer phenotype is essential for informing treatment decisions. Neuroendocrine prostate cancer is an especially aggressive subtype than emerges from epithelial cells over time because of high cell-fate plasticity. It is resistant to androgen therapies and is often associated with poor prognosis. However, its clinical and molecular presentation is diverse, and diagnosing it is challenging even for skilled physicians.
Dr. Michael Haffner, a Fred Hutch physician-scientist specialized in prostate cancer, explained, “Recognizing this phenotype remains difficult due to its diverse histologic appearances and the limitations of conventional diagnostic methods. There is a critical need for objective, quantitative approaches that can detect neuroendocrine features in routinely collected tissue samples and improve the consistency of tumor classification. To address this need, we collaborated with Dr. Harmon, an expert in computer vision, and a multidisciplinary team of colleagues in pathology and medical oncology, including Drs. Cora Sternberg, Ana Aparicio, Colm Morrissey and Peter Nelson.”
In this multi-institutional study, involving Fred Hutch, the NCI, Weill Cornell and MD Anderson, the team developed NEURAL-PC, an artificial intelligence deep learning-based computational tool that evaluates histologic images of standard tumor biopsies. It extracts cellular features like the shape, texture, and distribution of nuclei to predict the neuroendocrine nature of a sample. It provides each sample with an overall score ranging from 0 to 1, with higher values indicating more neuroendocrine-like features, as well as a spatial probability map to show neuroendocrine phenotypic variation. Validation of the tool, described in a recent JCI Insight publication, showed strong evidence of its efficacy.
In samples from a cohort of people receiving care for metastatic prostate cancer at the University of Washington, NEURAL-PC scores strongly correlated with the expression of two well-established, clinically used neuroendrocrine markers, synaptophysin (SYP) and insulinoma-associated protein (INSM1). When applied to a retrospective cohort of clinical metastatic prostate biopsy samples, NEURAL-PC differentiated the neuroendocrine phenotype from adenocarcinoma. Neuroendocrine prostate cancer samples received a median neuroendocrine feature score of 0.877 while adenocarcinoma samples received a median score of 0.243.
To understand whether NEURAL-PC could detect emerging neuroendocrine features in transitional tumors, the research team tested the tool on samples from a mouse model engrafted with patient-derived tumor samples that reproducibly develop a neuroendocrine phenotype after castration. At the baseline timepoint, NEURAL-PC uniformly called tumors neuroendocrine-negative. At 24 weeks after castration, tumors were heterogeneous, with spatial probably maps showing focal points of neuroendocrine features, and at 46 weeks, tumors uniformly exhibited a neuroendocrine phenotype. This suggests that NEURAL-PC can detect neuroendrocine features as they begin to emerge, an encouraging result for improving early diagnosis and treatment of neuroendocrine prostate cancer.
NEURAL-PC may also be a powerful prognostic tool. Based on tumor histology samples, it can stratify patients into either low or high risk of death within two years. In three separate cohorts, the predicted low-risk group had a longer median survival than the high-risk group. In one cohort, the median survival of the low-risk group was greater than 24 months while that of the high-risk group was just 14.3 months.
Dr. Haffner is encouraged by the performance of NEURAL-PC. “Its predictions were associated with molecular characteristics and clinical outcomes, demonstrating that routine histologic images contain valuable biological and prognostic information that can be extracted using artificial intelligence.” This information is vital to tailoring treatments for patients with metastatic prostate cancer. Accurate diagnosis ensures that people who need intensive therapy get it, and that people who don’t are spared unnecessary treatment.
Dr. Haffner and colleagues hope to further expand computational tools to gather more pertinent information about disease from standard histologic samples. He shared, “We plan to explore whether these image-based signatures can identify emerging treatment-resistant phenotypes, inform therapeutic selection, and provide clinically actionable biomarkers beyond conventional histopathologic assessment.”
Fred Hutch/University of Washington/Seattle Children’s Cancer Consortium Members Drs. Colm Morrisey, Michael Schweizer, Ruben Raychaudhuri, Evan Yu, Heather Cheng, Ruth Etzioni, Petere Nelson, and Michael Haffner contributed to this research.
The spotlighted research was funded by the National Institutes of Health, the US Department of Defense Prostate Cancer Research Program, the Doris Duke Charitable Foundation, the V Foundation, PCF Felix Fang PC-SYNERGY award, the Safeway Foundation, the Brotman Baty Institute for Precision Medicine, and the UW/FHCC Institute for Prostate Cancer Research.
Chen Z, Sayar E, Guevara D, Richards H, Zhang H, Patel RA, Gawne AC, Liu LJ, Coleman I, Dumpit R, Morrissey C, Schweizer MT, Raychaudhuri R, Graham LS, Yu EY, Cheng HH, Ding C-KC, Wang Y, Turkbey B, Chanel-Vos C, Fedorov C, Otilano JR, Kane T, Manohar J, Sigouros M, Nauseef JT, Molina A, Nanus D, Tagawa ST, Mosquera JM, Beltran HP, Etzioni R, Nelson PS, Soundararajan R, Aparicio AM, Sternberg CN, Haffner MC, Harmon SA. 2026. Deep learning-based histologic classifiers enable molecular subtyping of metastatic prostate cancer. JCI Insight. doi: 10.1172/jci.insight.201872.
Science Spotlight writer Ashley Person is a PhD candidate in the Cohn lab in the Vaccine and Infectious Disease Division at Fred Hutch. She studies how HIV-infected cells persist over time in people living with HIV on long term treatment.