Machine learning uncovers patterns in DNA methylation

German researchers have developed a new method for analysing the epigenome — the genome’s control system that determines which genes are switched on and off. The machine-learning method identifies differentially methylated DNA regions without sample labels — a prerequisite for many existing algorithms — making it possible to identify previously hidden biological patterns as well as new subgroups of cells or diseases.

The activity of our genes is not determined by DNA sequence alone. The attachment of small chemical compounds — known as methyl groups — influences which genes are active, and which remain silenced. DNA methylation is thus a central component of the epigenome.

Changes to the epigenome play a crucial role in the development of our body, influence our aging process, and are relevant to numerous diseases such as cancer. To understand such changes, researchers specifically search for differentially methylated DNA regions (DMRs). However, existing methods usually require that the samples under investigation be assigned to known groups, such as healthy or diseased tissue. With complex clinical datasets, however, this information is often unknown.

With the software tool known as metilene³, researchers have now created a method that overcomes this limitation. The software can compare DNA methylation patterns between both predefined groups (supervised mode) and unlabelled samples (unsupervised mode).

In the latter mode, the software searches for DMRs without samples having to be pre-classified into groups such as ‘healthy’ or ‘diseased’. The software autonomously segments the genome based on methylation signals, grouping the samples fully automatically. This automatic classification makes it possible to visualise epigenetic similarities and developmental relationships between the samples. Previously unknown cell types or disease subgroups can thus be identified, as well as those regions in which the samples both resemble and differ from already known diseases or cell types. At the same time, the biological differences remain traceable, since every similarity and difference can be attributed to specific methylation patterns.

The researchers tested the method on various biological datasets. Using human blood cells, metilene³ reconstructed the known developmental pathways of various immune cell types based on their DNA methylation alone. In addition, the software identified regulatory DNA regions linked to transcription factors that control the identity of these cell types. The results demonstrate that the method not only distinguishes between cell groups but also identifies functionally relevant regulatory elements.

This approach also proved to be extremely powerful when applied to tumour data. In datasets on glioblastomas, metilene³ identified various molecular subgroups of the tumours and even detected individual samples with unusual biological properties.

In tissue samples from pancreatic cancer, the software was also able to trace the step-by-step progression from healthy tissue through precancerous lesions to the tumour. In the process, the researchers discovered DNA regions in which the binding sites of the transcription factors NF-κB and NFAT occur together particularly frequently. These regions could play an important role in the development of pancreatic cancer and serve as a key starting point for further investigations. This biological hypothesis must now be experimentally verified.

“The cancer-related changes in DNA methylation identified by the machine learning method allow us to draw direct conclusions about molecular disruptions in transcription factors,” said Professor Helene Kretzmer of the Hasso Plattner Institute. Especially for medical questions, she noted, the requirements for the interpretability of predictions are particularly high.

“Our method opens up new possibilities for discovering previously hidden biological relationships and therapeutic approaches,” added Professor Steve Hoffmann of the Leibniz Institute on Aging - Fritz Lipmann Institute.

With metilene³, researchers now have a tool that significantly expands the analysis of complex DNA methylation data. Particularly in the case of heterogeneous tissue samples or clinical datasets, where biological groups cannot always be defined in advance, the method can help to uncover previously hidden biological relationships. The study’s authors therefore see great potential in the software for researching aging processes, cancer and other diseases, as well as for identifying new biomarkers. In the future, the tool should also be used for other sequencing technologies and single-cell analyses to make the analysis of epigenetic data even more widely applicable.

“The better we understand epigenetic patterns, the more precisely we can decipher the biological processes behind them,” Kretzmer said. “The software provides an important basis for this and could help us to derive new hypotheses for biomedical research from large datasets.”

The method has been described in the journal Nature Communications, while the program’s source code is freely available at github.com/zzhu1372/metilene3.

Image credit: FLI/Kerstin Wagner, AI-generated with ChatGPT