Cancer genomes are susceptible to multiple rearrangements by deleting, inserting, and translocating genomic regions. Recently, the problem of finding determinants of breakpoint formations was approached with machine learning methods; however, unlike cancer point mutations, breakpoint prediction appeared to be a more difficult task, and various machine learning models did not achieve high prediction power often slightly exceeding the threshold of random guessing. This raised the question of whether the breakpoints are random noise in cancer mutagenesis or there exist determinants in structural mutagenesis. In the present study, we investigated randomness in cancer breakpoint genome distributions through the power of machine learning models to predict breakpoint hot spots. We divided all cancer types into three groups by degree of randomness in their breakpoint formation. We tested different density thresholds and explored the bias in hot spot definition. We also compared prediction of hot spots versus individual breakpoints. We found that hot spots are considerably better predicted than individual breakpoints; however, some individual breakpoints can also be predicted with a satisfactory power, and thus, it is not proper to filter them from analyses. We demonstrated that positive-unlabeled learning can provide insights into insufficiency of cancer data sets, which are not always reflected by data set sizes. Overall, the present results support the view that cancer breakpoint landscape can be represented by predictable dense breakpoint regions and scattered individual breakpoints, which are not all random noise, but some are generated by detectable mechanism.
Keywords: cancer breakpoint hot spots; cancer breakpoints; cancer genome rearrangements; machine learning; random forest.