Summary: An international team of researchers has developed Hetairos, a deep-learning artificial intelligence that predicts the molecular classification of brain and spinal cord tumors from routine stained tissue sections in minutes. Trained on a large, diverse global dataset, Hetairos delivers rapid molecular subtyping that closely matches DNA methylation diagnostics, offering a faster, more accessible diagnostic option for central nervous system (CNS) tumors.
Hetairos was trained on more than 11,000 digitized histological slides from 9,606 patients across eleven medical centers on four continents. Using those slides—each linked to molecular DNA methylation-based diagnoses—the model learned to recognize microscopic patterns that correlate with molecular tumor subtypes. The result is an AI tool that reduces diagnostic turnaround time from roughly twelve days for full molecular testing to about twelve minutes after slide digitization, while providing classification across nearly the entire World Health Organization (WHO) spectrum of CNS tumor entities.
Key Facts
- Addressing the DNA methylation bottleneck: Accurate classification of many CNS tumors currently depends on DNA methylation profiling, which delivers high diagnostic resolution but requires specialized labs, substantial tissue, and considerable cost and time—often up to two weeks and unavailable in many low-resource settings.
- Large, diverse training set: Hetairos was built and validated on a global library of over 11,000 digitized H&E slides from 9,606 patients, with ground-truth labels established by molecular methylation diagnostics across multiple centers and continents.
- Comprehensive molecular coverage: The AI distinguishes 102 distinct methylation-defined tumor subtypes, covering nearly the full diagnostic range of current WHO CNS tumor classification.
- Superior performance against experts: In a direct comparison using only histology slides, Hetairos achieved 68% definitive accuracy on 210 challenging cases, compared with an average of 30% for five senior neuropathologists. Considering the top three candidate diagnoses, the AI reached 84% versus 50% for human experts.
- Rapid results: In prospective clinical evaluation, full molecular diagnostics took on average 12 days, whereas Hetairos produced molecular subtype predictions in approximately 12 minutes on standard hardware once slides were digitized. Including slide preparation, results were often available within 24 to 48 hours.
- Confidence-guided triage: The model reports a confidence score for each prediction. For roughly 50–70% of cases it flags predictions as “high-certainty,” where accuracy rises to about 87–88%. Even when confidence is lower, Hetairos reliably narrows the differential diagnosis from over 100 possibilities to a small set of likely candidates.
- Explainable image mapping: Hetairos highlights the specific microscopic regions that most influenced its decision, helping pathologists review and interpret AI findings, choose targeted molecular tests, and prioritize tissue sampling for further analysis.
Source: DKFZ
Brain and spinal cord tumors are highly heterogeneous. In recent years, clinicians have come to rely on molecular profiling—especially DNA methylation analysis—to resolve diagnostic uncertainty that cannot be resolved by morphology alone. While methylation profiling provides exceptional diagnostic precision, it is resource-intensive and can be slow or unavailable outside specialized centers.
How Hetairos was developed
Hetairos was developed by researchers led by Moritz Gerstung and Felix Sahm with collaborators across multiple institutions. The project aimed to predict methylation-defined molecular subgroups using only routinely prepared hematoxylin and eosin (H&E) slides. The training and validation set included over 11,000 slides from 9,606 patients. For each slide, the reference molecular diagnosis was established via DNA methylation profiling. The dataset spans a broad range of patient demographics and clinical settings to improve generalizability.
In routine use, Hetairos not only predicts a molecular subtype but also provides a confidence metric and a heatmap identifying regions that contributed most to the prediction. This helps clinicians verify AI reasoning, select regions for targeted genomic testing, and accelerate patient management.
Hetairos versus human experts
Five experienced neuropathologists assessed the same set of 210 difficult cases using only stained tissue sections. Hetairos achieved a single-best accuracy of 68%, compared with an average of 30% for the specialists. When evaluating the top three suggested diagnoses, Hetairos reached 84% accuracy versus 50% for humans. Developers note that the system currently struggles most with very rare tumor types, where human expertise remains essential, but performance should improve with even larger and more diverse datasets.
Clinical and practical implications
Hetairos can be particularly helpful when tissue quantity limits molecular testing, when molecular results are ambiguous, or where access to advanced molecular labs is restricted. By narrowing possible diagnoses and highlighting diagnostically important regions on the slide, the AI can streamline subsequent testing and shorten time to treatment. The tool is intended to complement—not replace—molecular diagnostics and expert review, serving as a fast decision-support system that increases diagnostic reach and efficiency.
Key Questions Answered
A: Hetairos was trained on thousands of H&E slides linked to known DNA methylation results. It learned to recognize subtle, reproducible microscopic patterns that correlate with specific molecular subtypes—patterns often too subtle for routine human observation. By mapping those visual signatures to molecular labels, the AI infers likely methylation-based subtypes directly from histology.
A: No. Hetairos is designed as an advanced diagnostic assistant. It speeds up and narrows differential diagnoses, but human neuropathologists remain essential for interpreting complex or rare cases, integrating clinical context, and making final treatment decisions. The AI functions as a digital copilot to increase accuracy and efficiency.
A: The primary benefit is time saved. For aggressive CNS tumors, faster molecular insights can enable earlier, more targeted treatment planning. Hetairos can provide highly informative subtype predictions within minutes of slide digitization, often enabling actionable guidance within 24–48 hours—much sooner than conventional molecular workflows.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full by the editorial team.
- Additional context was added by the staff to clarify clinical implications.
About this brain tumor and AI research news
Author: Sibylle Kohlstädt
Source: DKFZ (German Cancer Research Center)
Contact: Sibylle Kohlstädt – DKFZ
Image credit: Neuroscience News
Original Research: Open access. “Hetairos is a histology-based artificial intelligence model for predicting central nervous system tumor methylation subtypes” by Darui Jin et al., published in Nature Cancer. DOI: 10.1038/s43018-026-01186-3
Abstract
Hetairos is a histology-based AI model that predicts 102 methylation-defined CNS tumor subtypes from digital H&E slides. Built and validated on data from 9,606 patients and over 11,000 slides across four continents, Hetairos identifies 50–70% of cases with high confidence, achieving approximately 0.87 accuracy for those high-confidence predictions. In a direct histology-only comparison, Hetairos outperformed five board-certified neuropathologists (0.68 versus 0.30). Prospective evaluation in routine diagnostics replicated the performance and reduced diagnostic turnaround time from roughly 12 days (molecular testing) to about 12 minutes (AI inference). Hetairos can support decision-making across pediatric and adult CNS tumors by narrowing differentials and guiding efficient molecular testing.