Cancer Breakpoint Hotspots Versus Individual Breakpoints Prediction by Machine Learning Models

Abstract

Genome rearrangement is a hallmark of all cancers. Cancer breakpoint prediction appeared to be a difficult task, and various machine learning models did not achieve high prediction power. We investigated the power of machine learning models to predict breakpoint hotspots selected with different density thresholds and also compared prediction of hotspots versus individual breakpoints. We found that hotspots are considerably better predicted than individual breakpoints. While choosing a selection criterion, the test ROC AUC only is not enough to choose the best model, the lift of recall and lift of precision should be taken into consideration. Investigation of the lift of recall and lift of precision showed that it is impossible to select one criterion of hotspot selection for all cancer types but there are three to four distinct groups of cancer with similar properties. Overall the presented results point to the necessity to choose different hotspots selection criteria for different types of cancer.