AI model detects more than 170 types of cancer with remarkable precision
Researchers at Charité – Universitätsmedizin Berlin have developed crossNN, an artificial intelligence model that can classify more than 170 cancer types using DNA methylation patterns. The system achieves 99.1% precision for brain tumours and 97.8% for pan-cancer classification, working across multiple sequencing platforms including nanopore technology.
Researchers have developed a groundbreaking artificial intelligence model capable of diagnosing more than 170 different types of cancer with exceptional accuracy, potentially revolutionising tumour classification in clinical practice.
Cross-platform compatibility addresses diagnostic challenges
The new system addresses a critical limitation in current cancer diagnostics. Today, far more types of tumours are known than the organs from which they arise, with each tumour having distinct characteristics, growth rates, and metabolic peculiarities. The treatment of individual diseases depends decisively on accurate tumour type identification, particularly as new targeted therapies address specific cellular structures or block particular signalling pathways.
“Against the backdrop of increasingly personalised, rapidly developing cancer medicine, precise diagnosis at a certified tumour centre is the way forward for successful treatment,” stated Prof. Martin E. Kreis, Chief Medical Officer at Charité.
Traditional diagnostic approaches face significant obstacles when tissue sampling is impossible or carries high risk for patients. The MRI image of a brain tumour in an inauspicious location, requiring a high-risk biopsy for a patient presenting with double vision, exemplifies such challenging clinical scenarios that prompted the research team to seek innovative diagnostic solutions.
Epigenetic fingerprints enable precise tumour classification
The crossNN model exploits epigenetic modifications in tumour cells – chemical changes that determine which parts of genetic information are read and when. “Hundreds of thousands of epigenetic modifications act as on and off switches for individual gene sections. Their patterns form a unique, unmistakable fingerprint,” explained Dr Philipp Euskirchen, scientist at the Berlin site of the German Cancer Consortium and lead author of the study published in Nature Cancer on 6 June 2025.
“In tumour cells, the epigenetic information is altered in a characteristic way. Based on their profiles, we can differentiate between tumours and classify them,” Dr Euskirchen said. Remarkably, for brain tumours, even cerebrospinal fluid samples can suffice for analysis, potentially eliminating the need for surgical intervention entirely.
The research demonstrates the model’s effectiveness across multiple sequencing platforms, including nanopore sequencing, targeted bisulfite sequencing, and various microarray platforms. This cross-platform compatibility represents a crucial advancement, as different genomic coverage and sequencing depths have traditionally required separate, platform-specific analytical approaches.
Simple architecture delivers superior performance
Unlike complex deep learning models, crossNN employs a remarkably simple neural network architecture consisting of a single-layer perceptron. This streamlined design paradoxically enhances both performance and interpretability whilst reducing computational requirements.
“Although the architecture of our AI model is far more simple than previous approaches and therefore remains explainable, it delivers more precise predictions and therefore greater diagnostic certainty,” said Dr Sören Lukassen, head of the Medical Omics working group at the Berlin Institute of Health at Charité.
The model was trained using reference datasets containing thousands of tumour samples. For brain tumour classification, the system achieved 99.1% precision, whilst the pan-cancer model demonstrated 97.8% precision across all organ sites. Validation studies encompassed more than 5,000 tumour samples generated on different platforms, confirming the model’s robustness and scalability.
Clinical implementation shows immediate benefits
The model’s clinical utility has already been demonstrated. A patient presenting with double vision benefited directly from this technology. “We examined the cerebrospinal fluid using nanopore sequencing, a novel, very fast and efficient form of genetic analysis. The classification by our models revealed that it was a lymphoma of the central nervous system, enabling us to promptly initiate appropriate chemotherapy,” Dr Euskirchen explained.
The authors noted in their discussion that the model “enables instantaneous predictions from methylation profiles generated by multiple platforms, including WGBS, targeted methyl-seq, low-coverage nanopore whole-genome sequencing and several microarray platforms.”
Future clinical trials and regulatory approval
The research team, collaborating with the German Cancer Consortium (DKTK), is planning clinical trials with crossNN at all eight DKTK locations across Germany. Intraoperative applications are also under investigation, with the ultimate goal of transferring this precise and comparatively inexpensive tumour determination to routine clinical care.
The model’s fully explainable architecture represents a crucial factor for future regulatory approvals in clinical application. Unlike black-box artificial intelligence systems, crossNN’s decision-making process can be understood and validated, meeting essential requirements for medical device certification.
Wider implications for cancer diagnosis
The accuracy of the methodology surprised even the researchers themselves. The study’s authors concluded that their work “offers a machine learning framework for cross-platform DNA-methylation-based classification of cancer, enabling the development of rapid, resilient, interpretable and accurate diagnostic tests.”
Looking ahead, the authors suggested that “these methods hold promise to become valuable diagnostic tools for all types of cancer well beyond neuro-oncology,” potentially transforming cancer diagnostics across multiple medical specialties.
• The crossNN platform is accessible through an intuitive webbased interface at https://crossnn.charite.de, allowing immediate tumour classification from uploaded methylation data.
Reference:
Yuan, D., Jugas, R., Pokorna, P., et. al. (2025). crossNN is an explainable framework for cross-platform DNA methylation-based classification of tumors. Nature Cancer.
https://doi.org/10.1038/s43018-025-00976-5
