Machine learning method uncovers hidden patterns in DNA methylation

Researchers from Berlin, Potsdam, and Jena have developed a new machine-learning method called metilene3, which analyzes DNA methylation patterns without the need for sample labels. This breakthrough allows for the identification of previously hidden biological patterns and new subgroups of cells or diseases. DNA methylation, a central component of the epigenome, influences gene activity and is linked to various biological processes, including development, aging, and disease.

Existing methods typically require samples to be assigned to known groups, but metilene3 can compare methylation patterns within unlabeled samples. The software autonomously segments the genome based on methylation signals, grouping samples automatically. This classification enables the visualization of epigenetic similarities and developmental relationships between samples, revealing previously unknown cell types or disease subgroups.

The method was tested on various biological datasets, including human blood cells and tumors. In blood cell datasets, metilene3 reconstructed known developmental pathways and identified regulatory DNA regions. In tumor data, the software identified molecular subgroups and detected unusual biological properties in individual samples.

Additionally, it traced the progression from healthy tissue to cancer in pancreatic cancer samples, identifying key DNA regions associated with transcription factors that may play a role in cancer development. The researchers emphasize the interpretability of their method, which allows for direct conclusions about molecular disruptions in transcription factors.

This tool significantly expands the analysis of complex DNA methylation data, particularly for heterogeneous tissue samples or clinical datasets where biological groups may not be pre-defined. The study's authors see great potential for using metilene3 in researching aging processes, cancer, and other diseases, as well as identifying new biomarkers.