Cancer diagnosis has traditionally relied on microscopic examination of tissue samples and analysis of selected genetic mutations. However, researchers are increasingly focusing on epigenetic changes that regulate gene activity rather than alterations in DNA sequence alone.

This trend is also visible in clinical practice. In recent years, the World Health Organization has included molecular markers in the classification of tumours of the central nervous system, among them DNA methylation patterns, which to better differentiate between tumour types than classical histopathology.

The new classifier, developed by researchers at the Pomeranian Medical University in Szczecin, uses DNA methylation profiles and machine learning to identify cancers affecting multiple organs. The system was trained on data from nearly 17,000 patients and its results were published in the journal Genome Medicine.

According to the study's first author, Professor Tomasz Wojdacz, head of the Independent Laboratory of Clinical Epigenetics at the Pomeranian Medical University, the approach could significantly change cancer diagnostics by providing faster and more detailed tumour classification.

DNA methylation is one of the basic mechanisms of gene regulation. It involves adding methyl groups to selected fragments of the genome. This does not change the DNA sequence itself, but affects which genes are active and which are silenced. In healthy cells, this process is strictly controlled and responsible for the proper functioning of tissues. However, in cancer, it is dysregulated, leading to abnormal activation or deactivation of genes, the scientist explains.

"In other words: genetics determines how a gene is written; epigenetics, how a gene works," he adds.

Wojdacz says researchers once believed DNA mutations were the primary cause of cancer, but advances in sequencing have shown they do not fully explain how many tumours develop. Instead, cancers exhibit distinctive methylation patterns that can serve as molecular signatures.

"It turns out that these changes are so specific to tumours that they can be very precisely used to classify them," Wojdacz says.

To create a methylation profile, researchers analyse up to one million locations in the genome and identify about 1,500 characteristic changes, which are then used to train machine learning models.

Unlike earlier classifiers that often performed well only under research conditions, the new system was designed to be more resistant to methodological errors and variations between datasets. Researchers also expanded the analysis beyond a single organ to include cancers affecting multiple parts of the body.

Tests showed the classifier accurately identified almost 50 types of cancer, including brain, pancreatic and lung tumours, while maintaining strong performance on real clinical samples. Its accuracy was lower for some very rare cancers because of limited training data.

"If we have few samples, the model will be wrong more often. But the problem is not the method itself, but access to training data," he emphasises.

Although the classifier remains a research tool and has not yet been certified for clinical use, Wojdacz says it has considerable potential for future cancer diagnostics.

"They will be much faster, because classic histopathology requires the evaluation of preparations by a pathologist, while methylation analysis is largely automatic. From the moment the sample is collected, we can have a diagnosis within one day," he says.

Katarzyna Czechowicz (PAP)

kap/ zan/

tr. RL