Scientists evaluate a tool that identifies tumors with more than 97% accuracy

An international team led by experts from Charité–Universitätsmedizin Berlin developed an artificial intelligence model to classify more than 170 tumor types with an accuracy rate exceeding 97%.

The breakthrough, published in the journal Nature Cancer, aims to transform cancer diagnosis by enabling accurate analysis even with fragmentary data and without the need for invasive biopsies.

“The model is not only accurate but also interpretable, giving specialists a clear understanding of how predictions are made, which is critical in the clinical setting,” explained Dr. Philipp Euskirchen, co-senior author of the study.

Diagnosis based on DNA methylation

According to the authors, the technique is based on the analysis of DNA methylation patterns, an epigenetic process that regulates gene activity and whose alteration is characteristic of tumor cells. This “epigenetic fingerprint” allows for precise differentiation of cancer subtypes, even those with similar histological features.

“Hundreds of thousands of epigenetic modifications act as on/off switches for individual sections of genes. Their patterns form a unique and unmistakable fingerprint,” said Euskirchen, a researcher at the Berlin headquarters of the German Cancer Consortium and the Charité Institute of Neuropathology.

This model was designed to work with data obtained through various technological platforms, such as microarrays or nanopore sequencing.

“Our goal was to develop a model that would accurately classify tumors, even if they were based only on parts of the entire tumor epigenome or if profiles were collected using different techniques and with varying degrees of accuracy,” said bioinformatician Sören Lukassen, head of the Medical Omics group at the Berlin Institute of Health at Charité.

Accuracy in more than 5,000 clinical cases

The model was validated with more than 5,000 tumor samples, achieving an accuracy of 99.1% in brain tumors and 97.8% in general cancers. Furthermore, it maintained high performance even when incomplete or low-resolution genomic data were used.

“Our model enables highly accurate diagnosis of brain tumors in 99.1% of cases and is more accurate than existing AI solutions to date,” Euskirchen stated.

One of the most significant contributions of this model is its application in scenarios where performing a biopsy is risky. In these cases, cerebrospinal fluid analysis may be sufficient to generate a reliable diagnosis, eliminating the need for surgical intervention, according to these specialists.

“We examined the cerebrospinal fluid using nanopore sequencing, a novel, very rapid, and efficient form of genetic analysis. The classification performed by our models revealed a central nervous system lymphoma, which allowed us to quickly initiate appropriate chemotherapy,” Euskirchen recounted a recent clinical case.

Furthermore, the model can make predictions in a matter of seconds, making it particularly useful in contexts where rapid responses are required.

The model could not only improve diagnostic accuracy but also advance the identification of genes associated with specific tumor subtypes, potentially leading to new biomarkers and individualized therapeutic strategies.

“We detected how methylation in specific genes, such as MUM1, is associated with rare tumor subtypes, which will help us better understand their biological mechanisms,” Euskirchen added.

While the results are promising, the authors caution that some rare tumor types are underrepresented in the datasets, limiting the current scope of the model in certain cases. They also encountered difficulties in differentiating between similar subtypes, such as papillary and clear cell renal cell carcinomas.

In collaboration with the German Cancer Consortium (DKTK), clinical trials are being planned at eight cancer centers in Germany. One goal is to evaluate the use of this tool during surgery and in routine clinical practice.

“Although the architecture of our AI model is much simpler than previous approaches and therefore still explainable, it offers more accurate predictions and, consequently, greater diagnostic certainty,” Lukassen concluded.