Mobile Menu

Machine learning and Darwin combined to piece together tumour genetics

MOBSTER – model-based clustering in cancer, is a machine learning tool that incorporates theoretical population genetics from Darwin to examine cancers. Specifically, to identify important sub-clonal populations in tumours that confer either, a positive advantage, treatment resistance or potential metastasis.

Developed by researchers at The Institute of Cancer Research and Queen Mary University, and funded by Cancer Research UK,  this machine learning tool can more accurately analyse genetic data from cancers. In doing so, they discovered that tumours have a simpler genetic architecture than previously thought.

As cancers change over time through clonal evolutions, tumours are a mix of sub-populations arising from a smaller number of cells (intratumor heterogeneity). To assess these populations, bulk samples are taken and sequenced, with various clusters of reads being classed as different sub-populations.

Heterogeneous tumours are likely to be resistant to single-drug treatments as a drug would cause a bottleneck effect, killing all but the drug-resistant cells.

Previous methods were purely data-driven, which meant sub-populations were identified and used to study the evolution of cancer but did not consider evolutionary dynamics, such as neutral mutations. This meant systematic errors occurred leading to more sub-populations being identified and important mutations being overlooked.

Identifying subpopulations is an important step in combating drug-resistance in tumours. However, due to the lack of appropriate biomarkers, sometimes the level of heterogeneity in the tumour itself is used as a marker to indicate the likelihood of drug-resistance.

Therefore, correctly determining the number of subpopulations is important. Using MOBSTER and accounting for neutral mutations, the researchers found fewer groupings and discovered that tumours had a simpler genetic architecture than previously thought. They were also able to determine the age of each subclone.

MOBSTER was validated on the Pan-Cancer Analysis of Whole Genomes (PCAWG) international consortium, one of the largest available cohorts, of 2,606 cancers. The hope is that by improving the analysis of tumour heterogeneity and subclonal reconstruction, MOBSTER will lead to better and tailored treatment strategies for patients.

More on these topics

Cancer / Machine Learning