Randomness in Cancer Breakpoint Prediction.

Area of Science:

• Genomics
• Cancer Biology
• Bioinformatics

Background:

• Cancer genomes undergo frequent rearrangements like deletions, insertions, and translocations.
• Predicting structural variation breakpoints in cancer is challenging, with machine learning models often performing near random chance.

Purpose of the Study:

• To investigate the randomness of cancer breakpoint distributions using machine learning.
• To determine if machine learning models can predict breakpoint hot spots and understand determinants of structural mutagenesis.

Main Methods:

• Applied machine learning models to predict cancer breakpoint hot spots.
• Categorized cancer types based on breakpoint formation randomness.
• Evaluated different density thresholds and biases in hot spot definitions.
• Compared the prediction accuracy of hot spots versus individual breakpoints.
• Utilized positive-unlabeled learning to assess data set limitations.

Main Results:

• Hot spots are significantly more predictable than individual breakpoints.
• Some individual breakpoints demonstrate predictable power and should not be filtered.
• Positive-unlabeled learning highlighted potential data set insufficiencies.
• Cancer breakpoint distributions can be characterized by predictable dense regions and scattered individual breakpoints.

Conclusions:

• Cancer breakpoint landscapes are not entirely random; predictable regions and some individual breakpoints exist.
• Detecting mechanisms driving structural mutagenesis is possible.
• Careful analysis is needed to avoid discarding potentially informative individual breakpoints.