Researchers have developed an AI model that predicts DNA methylation–based central nervous system (CNS) tumor subtypes directly from routine hematoxylin and eosin (H&E)–stained histology slides, potentially reducing diagnostic turnaround times from weeks to minutes while helping to guide molecular testing. The study, led by Darui Jin, PhD, a postdoctoral fellow in the Division of AI in Oncology at the German Cancer Research Center Deutsches Krebsforschungszentrum (DKFZ), Heidelberg, Germany, was published in Nature Cancer.

“Hetairos demonstrates the enormous potential of AI-supported digital pathology to provide rapid and widely available diagnostic methods that were previously only possible with considerable technical effort," stated co-corresponding author Moritz Gerstung, PhD, the Division Head for Artificial Intelligence and Cancer Evolution at DKFZ and Professor at the University of Heidelberg. 

Study and Model Details

Molecular profiling—particularly DNA methylation analysis—has become a cornerstone of modern CNS tumor classification and is incorporated into the 2021 World Health Organization classification system. However, methylation testing requires specialized laboratories, adequate tissue samples, and often takes approximately 2 weeks to complete, limiting its accessibility worldwide. The investigators sought to determine whether AI could extract and aggregate comparable molecular information directly from routinely available H&E slides, thereby accelerating diagnosis and expanding access to advanced tumor classification.

The resulting model, named Hetairos (Greek for "companion"), was trained to classify 102 methylation-defined CNS tumor subtypes using digital pathology images. The investigators trained and validated the algorithm using more than 11,000 H&E slides from 9,606 patients collected at 11 institutions across four continents, making it one of the largest multicenter studies of AI in neuropathology to date.

With Hetairos, each whole-slide image was preprocessed to identify tissue-containing areas, then features were extracted using a pretrained histopathology foundation model, and finally features were aggregated using embedding-based multiple-instance learning. For predicting subtypes, a deep ensemble was created from the combined outputs of multiple independently trained models for enhanced accuracy. The model incorporated the whole-slide images together with patient age and tumor location and was benchmarked against methylation-based diagnoses as the ground truth. 

The model's classifications were compared with that of five board-certified neuropathologists based on H&E slide images alone. 

Key Results

Across the internal validation cohort, Hetairos correctly identified the top diagnosis in 75% of all cases and achieved a top-3 accuracy of 87%, or the identification of the three most likely classes. The model also estimated its own probability, or confidence. Approximately 70% of internal cases received high-confidence predictions, and within this subset, top-1 accuracy increased to 88%. In half of the incorrect cases, the true class belonged to the same superfamily as the predicted class. Even among lower-confidence cases, the correct diagnosis was included within the model's three most likely predictions in 71% of cases, substantially narrowing the differential diagnosis and potentially reducing the need for broad molecular testing.

Generalizability was evaluated using 10 external datasets comprising 4,645 tumors and 5,498 slides from institutions in Europe, North and South America, and Asia. Overall top-1 accuracy decreased modestly to 68%, largely because a greater proportion of external cases were assigned low confidence. However, performance remained consistent in the high-confidence cases (55%), with 87% top-1 accuracy, suggesting that the model appropriately calibrated its confidence across different practice settings.

When the model was compared side-by-side with neuropathologists, using 210 H&E slides without access to molecular data or ancillary testing, Hetairos achieved a 68% top-1 accuracy, compared with an average of 30% for the specialists. For the three most likely diagnoses, the AI achieved 84% accuracy, whereas the neuropathologists averaged approximately 50%. The advantage was greatest for more common tumor subtypes and higher-confidence cases, although human experts performed similarly for several of the rarest entities.

Dr. Jin noted, "The study shows that artificial intelligence is capable of deriving molecular information directly from routine tissue sections and thus fundamentally changing cancer diagnostics."

Beyond classification accuracy, the investigators evaluated Hetairos prospectively in routine clinical practice. Between August 2024 and June 2025, the system analyzed 210 consecutive tumor samples at the University Hospital Heidelberg, in parallel with standard diagnostics without influencing patient management. Whereas conventional integrated diagnosis took about 12 days, or about 16 days with the addition of molecular testing, Hetairos generated predictions in approximately 12 minutes once digitized slides were available. Top-1 predictions agreed with the final integrated diagnosis in 90.2% of high-confidence cases and reached 94.3% accuracy among cases with high-confidence methylation results. Accuracy was 45.5% for the model among low-confidence cases. 

Another notable feature is the model's interpretability. Hetairos generates heat maps identifying the tissue regions that most strongly influenced each prediction, allowing pathologists to review the AI's reasoning and identify optimal regions for additional molecular testing when needed. According to the authors, this capability may improve diagnostic efficiency while providing educational value for recognizing subtle histologic patterns associated with specific molecular subtypes.

As senior author Felix Sahm, MD, PhD, Medical Director of the Department of Neuropathology, University Hospital Heidelberg, explained, "We developed Hetairos primarily as a tool to support diagnostics. It is not intended to replace molecular analyses, but rather to specifically complement and accelerate them."

DISCLOSURES: This study was supported by CRC/SFB 1389 UNITE Glioblastoma. Open access funding was provided by DKFZ. For study author disclosures, as well as data and code availability, visit nature.com. 

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.