Some prostate cancers change character after hormone treatment. They become neuroendocrine prostate cancer, a subtype that no longer responds to drugs targeting the androgen receptor and usually carries a poor outlook. Spotting it is hard even for experienced pathologists, because it looks different from patient to patient and confirming it normally needs extra stains.
A team led by Dr Michael Haffner at Fred Hutch Cancer Center and Stephanie Harmon of the US National Institutes of Health has built an AI tool that looks for it in ordinary tissue slides. The tool, called NEURAL-PC, is described in JCI Insight and in a Fred Hutch Science Spotlight published on 28 October. The work involved Fred Hutch, the US National Cancer Institute, Weill Cornell and MD Anderson.
In brief
- NEURAL-PC, an AI tool from a Fred Hutch-led team, scores routine biopsy slides for signs of neuroendocrine prostate cancer, an aggressive form of the disease.
- In 63 past biopsies it separated that subtype from the common form with an AUROC of 0.921, the paper says.
- It is a research tool tested on past samples. It is not an approved diagnostic, and the authors call for prospective studies.
How NEURAL-PC reads a slide
NEURAL-PC works on H&E slides, the standard pink-and-purple stained tissue sections made for almost every biopsy. According to Fred Hutch, it “extracts cellular features like the shape, texture, and distribution of nuclei” and gives each sample a score from 0 to 1, where higher means more neuroendocrine-like. It also draws a map showing where in the tissue those features appear.
The model was trained on tissue from the University of Washington’s rapid autopsy programme: 722 small tissue cores from 204 metastatic sites in 52 patients, the paper says. The training was guided by measurable nuclear features, which the authors say makes the model easier to interpret.
The results on past biopsies
The main test used 63 clinical biopsies collected in the past at the University of Washington: 24 neuroendocrine, 38 adenocarcinoma, the common form, and one mixed case. NEURAL-PC reached an AUROC of 0.921, a measure where 1 is perfect separation and 0.5 is chance. Neuroendocrine samples had a median score of 0.877 and adenocarcinoma samples 0.243. Scores also tracked two markers pathologists already use, synaptophysin and INSM1. On a separate set of 118 slides with gene-expression data, it scored an AUROC of 0.810 against pathologists’ assessment.
The team also used the scores to sort patients into lower and higher risk of death within two years. In the University of Washington cohort, the high-risk group had a median survival of 14.3 months against more than 24 months for the low-risk group. In 82 patients from the C-COLA phase II trial, the split was 12.3 months against 20.4 months, a statistically significant difference. In a Weill Cornell cohort of 66 patients, the gap was 8.62 against 19.6 months, but the paper says it did not reach statistical significance.
“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,” Haffner said in the spotlight.
Catching the change as it happens
The researchers also tested tumours from a mouse model with patient-derived tissue that turns neuroendocrine after castration. Before treatment, the tool scored a sample at 0.04. At week 46, when the tumours were fully neuroendocrine, it scored 0.851. The in-between stage is less clear. Fred Hutch’s write-up says the maps showed focal neuroendocrine features at week 24, but the week-24 example in the paper was scored 0.154 and classed as negative, even though staining showed scattered marker-positive cells.
Why routine slides are the prize
This is TSN’s analysis. The appeal of NEURAL-PC is that it asks for nothing new. Hospitals already make H&E slides for nearly every biopsy, so a tool that can flag a dangerous subtype from them could help decide which samples need extra stains or genetic tests, and which patients might need different treatment. It follows other attempts to pull hidden signals from standard slides, such as King’s College London’s HRDPath for ovarian cancer and Sophont’s Holon-S1 pathology model. The commercial interest is real too, as Agilent’s planned purchase of Proscia shows.
Small cohorts, past samples
The evidence is still early. Every test used past samples, the training set came from 52 patients, and only a handful of mixed cases were available. The paper itself says “future prospective work and validation at additional clinical sites are warranted”. The survival splits are encouraging, but one of the three did not reach significance, and the mouse results show the tool can miss an early, patchy change. Haffner says the team next wants to see whether these image signatures can spot emerging treatment resistance and help choose therapy. Like UCLA’s MAESTRO drug-pairing AI, NEURAL-PC produces strong leads. It will need prospective trials before it can guide a single patient’s care.
Sources
- Fred Hutch Cancer Center, Science Spotlight, “Artificial intelligence takes on neurodendocrine prostate cancer diagnosis”, 28 October 2026
- Chen et al., “Deep learning–based histologic classifiers enable molecular subtyping of metastatic prostate cancer”, JCI Insight, 8 October 2026, DOI 10.1172/jci.insight.201872
Related stories
- KCL AI Spots DNA-Repair Defect in Ovarian Cancer From Routine Slides
- Sophont’s Holon-S1 Aims to Predict Tumour Gene Activity From Routine Slides
- UCLA’s MAESTRO AI Ranked 3,592 Compounds to Pair With a Cancer Drug
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
