0.87 accuracy on high-confidence predictions, 50-70% of cases covered, and about 12 minutes per slide: Hetairos walks into CNS tumor diagnostics carrying numbers that make you raise one eyebrow and immediately ask where the trapdoor is.

The context is brain and spinal cord tumors, where the name on the diagnosis is not just medical stationery. It can change prognosis, treatment planning, clinical trial eligibility, and whether everyone in the room gets to sleep tonight. Modern neuropathology increasingly relies on DNA methylation profiling, which reads chemical tags on DNA that help identify tumor subtypes. Think of methylation as the tumor’s browser history: not the whole genome, but enough behavioral evidence to make the case awkwardly specific.

The catch? Methylation testing needs specialized equipment, enough tissue, money, and time. In the Hetairos study, molecular testing took about 12 days on average. Twelve days is fine for sourdough. Less charming when a patient is waiting on a brain tumor diagnosis.

The Slide Whisperer, Allegedly

Hetairos is an AI model that predicts 102 methylation-based central nervous system tumor subtypes from ordinary H&E-stained pathology slides. H&E is the classic pink-and-purple stain pathologists use every day. It is basically the default Instagram filter for tissue, except the stakes are wildly higher and nobody is adding sparkle effects.

The researchers trained and validated Hetairos on 9,606 patients and more than 11,000 slides from 11 centers across four continents. The model chops giant whole-slide images into small tiles, extracts visual features using a pretrained vision transformer, then combines those tile-level clues with patient age and tumor location. From there, it ranks possible tumor subtypes.

That last part matters. Hetairos is not just blurting one answer like a chatbot that skimmed the manual. It outputs probabilities. When confidence was high, its top prediction matched the methylation-based label with about 0.87 accuracy in external validation. When confidence was low, top-1 accuracy fell hard, around 0.45 to 0.46, but the correct answer was often still in the top three.

Translation: when Hetairos shrugs, you should notice the shrug.

Beating Humans, With Asterisks

In a direct histology-only comparison, Hetairos outperformed five board-certified neuropathologists: 0.68 accuracy versus 0.30. That sounds brutal until you read the fine print, which is where the interesting stuff usually hides wearing a tiny fake mustache.

The human experts had to choose among a very granular list of methylation-defined classes using H&E images alone. That is not standard practice for many diagnoses, because pathologists usually combine morphology with immunohistochemistry, molecular tests, clinical context, and the sacred diagnostic ritual of “please send more tissue.” So the comparison shows something real, but not “AI replaces neuropathologists.” It shows that subtle visual patterns in routine slides may contain more molecular signal than humans can reliably unpack unaided.

The authors themselves frame Hetairos as a support tool. That is the sane interpretation. It can narrow the differential, flag likely subtypes, guide which molecular tests to order first, and maybe spare some labs from firing the entire expensive testing cannon at every case.

The Part Where We Do Not Clap Too Fast

The limitations are not decorative. Rare tumor subtypes performed worse, especially when the training set had fewer than 20 examples. External cohorts also pushed confidence down, which is what you want from a cautious model but still means deployment will need local validation. A model that knows when it is confused is better than one that confidently drives into a lake, but it is still confused.

There is also a deeper issue: methylation profiling remains the reference standard here. Hetairos predicts methylation-defined classes from morphology. That is clever, but it does not magically measure methylation. If the model and molecular test disagree, the answer is not automatically “trust the faster one because it has a cool Greek-ish name.”

Still, the clinical angle is strong. In prospective routine diagnostics, high-confidence Hetairos calls agreed with final integrated diagnoses in 120 of 133 cases, about 90.2%. Running on consumer-grade hardware, it generated reports in about 12 minutes after slide scanning. If reproduced widely, that could help hospitals triage cases faster, especially where methylation profiling is scarce or slow.

Why This Is More Than A Party Trick

Hetairos fits a larger trend: AI pathology models are getting better at linking what tissue looks like to molecular biology. DEPLOY previously predicted 10 major CNS tumor categories from histology and methylation-related signals. MethyLYZR attacked the speed problem from another direction, using sparse nanopore methylation data for near-real-time classification. Explainable methylation-classifier work is also probing which genomic regions drive these decisions, because “the model vibes with enhancer regions” is not quite the trust framework clinicians deserve.

The exciting version of this future is not a robot pathologist in a lab coat. It is a workflow where routine slides give fast, calibrated suggestions, molecular testing confirms the tricky stuff, and scarce resources go where they actually help. The skeptical version asks: will this hold up across scanners, staining protocols, rare subtypes, and messy real-world cases? Good. Keep asking that. Medicine improves faster when the hype has to show receipts.

References

1. Jin, D. et al. “Hetairos is a histology-based artificial intelligence model for predicting central nervous system tumor methylation subtypes.” Nature Cancer (2026). DOI: 10.1038/s43018-026-01186-3. PubMed: PMID 42270902

2. Hoang, D.-T. et al. “Prediction of DNA methylation-based tumor types from histopathology in central nervous system tumors with deep learning.” Nature Medicine 30, 1952-1961 (2024). DOI: 10.1038/s41591-024-02995-8

3. Brändl, B. et al. “Rapid brain tumor classification from sparse epigenomic data.” Nature Medicine (2025). DOI: 10.1038/s41591-024-03435-3

4. Hovestadt, V. et al. “Explainable artificial intelligence of DNA methylation-based brain tumor diagnostics.” Nature Communications 16, 1787 (2025). DOI: 10.1038/s41467-025-57078-0

5. “DNA methylation.” Wikipedia. https://en.wikipedia.org/wiki/DNA_methylation

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.