A deep-learning model called Hetairos — led by Darui Jin with senior authors Moritz Gerstung (DKFZ) and Felix Sahm (Heidelberg University) — can classify 102 methylation-defined subtypes of central nervous system tumors from standard hematoxylin-and-eosin histology slides in approximately 12 minutes — compared with roughly 12 days for the gold-standard DNA methylation profiling — and outperformed a panel of five board-certified neuropathologists by a wide margin in a controlled comparison, according to a study published June 10, 2026, in Nature Cancer.
The model was developed by a team led by lead author Darui Jin, with senior co-corresponding authors Moritz Gerstung at the German Cancer Research Center (DKFZ) and Felix Sahm at Heidelberg University Medical Faculty and Heidelberg University Hospital. Validation drew on more than 11,000 digitized tissue sections from 9,606 patients across 11 medical centers on four continents, making it one of the largest multi-institutional assessments of an AI pathology tool for CNS tumors to date.
Head-to-head performance. In a histology-only comparison using 210 cases spanning the full spectrum of CNS tumor types, Hetairos achieved a top-1 accuracy of 0.68 — the proportion of cases in which the model’s single best prediction matched the molecular ground truth. The five neuropathologists, working from the same slides without molecular data, reached a mean top-1 accuracy of 0.30. At top-3 (whether the correct subtype appeared among the three most likely predictions), Hetairos scored 0.84 versus 0.50 for the specialists. For cases in which the model assigned its highest confidence scores, accuracy reached 0.87.
The study also included a prospective component: Hetairos ran in parallel with routine clinical diagnostics on 111 consecutive tumors without influencing treatment decisions, confirming the 12-minute turnaround.
What it does — and does not — do. The model predicts molecular subtype from morphology alone; it does not perform or replace molecular testing. Sahm stated in a release from DKFZ that the tool is “not intended to replace molecular analyses, but rather to specifically complement and accelerate them.” Rare subtypes underrepresented in training data remain a challenge, and the authors note that the validation centers were predominantly high-income academic institutions — a limitation with direct implications for generalizability.
Equity dimension. Because Hetairos runs on digitized versions of H&E sections — the same stain used in virtually every pathology laboratory worldwide — proponents argue it could extend sophisticated subtype classification to settings where DNA methylation arrays are unavailable or unaffordable. Gerstung described it as demonstrating the tool’s “enormous potential of AI-supported digital pathology to provide rapid and widely available diagnostic methods that were previously only possible with considerable technical effort.” Whether that potential translates to low-resource settings will depend on digital-pathology infrastructure — slide scanners, connectivity, and local implementation capacity — none of which is addressed in the current validation.
The model’s code is publicly available via the Gerstung lab’s GitHub repository. No confidence intervals for the accuracy figures were available from sources accessible to this desk; the primary paper is paywalled.