Recognition of Regions Containing Breakpoints in Cancer Genomes with Deep Learning Methods

In AD 2000, the results of a comprehensive study of many classes of structural variants (SVs) — which are a type of cancer mutations — were presented by the Pan-Cancer Analysis of Whole Genome (PCAWG) Consortium. Among many classes of SVs, chromothripsis — in which many breakpoints, and, accordinly, SVs, occur in a single event — is of particular interest. The aim of this paper is to determine, by means of classical and deep machine learning methods, whether there be a link between the primary DNA structure and the formation of cancer breakpoints in chromothripsis. Five models have been constructed — random forest, gradient boosting, CNN, RNN and Transformer — of which random forest proved the most accurate but even it could not achieve a ROC-AUC score higher than 0.6. The obtained results suggest that there is no significant relationship between the primary structure of DNA segments and the emergence of chromothirpsis breakpoints therein.