Learning to better understand: Novel bioinformatics algorithms for cancer research

Abstract

In this thesis, we focused on developing new bioinformatics algorithms with the ultimate aim of creating a better understanding of cancer. Our research targeted 3 unsolved problems.

First, a main difficulty in treating cancer is that it is often unclear which mutations are present. As a result, cells with undetected mutations may escape therapy and re-grow. Our solution involves sampling from multiple regions of the tumor to capture a larger number of mutations, and constructing a tree that represents the order in which these mutations originated.

Second, many cancers present with structural variants in non-coding regions, yet their function is unknown. We found that these structural variants often disrupt the folding of the genome, thereby enabling novel interactions between genes and regulatory elements, such as enhancers. We introduce an algorithm that uses machine learning to detect from gene expression data which structural variants may be causal for cancer.

Third, genes often lose function through non-genetic mechanisms, such as methylation. However, obtaining such data types is costly, and therefore often only genetic disruptions, i.e. mutations, can be detected. We found that genes with loss of function often leave behind specific signatures of mutations and structural variants in the genome. Using this information, we developed machine learning algorithms that can detect aberrant gene function even if loss cannot be measured directly.