Machine learning-based refinement of genomic structural variations and genomic analysis of ALK-rearranged non-small cell lung cancer

Abstract

Genomic structural variations have been reported to be involved in tumorigenesis by disrupting tumor suppressor genes and activating oncogenes. With the advent of whole-genome sequencing, structural variations can be detected at a single base-pair resolution. However, currently available tools for the detection of structural variations report many false-positive calls. Thus, a substantial downstream effort has been required to refine the call set. On the other hand, genomic analyses on ALK-rearranged non-small cell lung cancer have been limited to protein-coding regions. This dissertation consists of mainly two parts. Part 1 describes our study on the development of a machine learning-based system for the refinement of structural variations. Our machine learning-based method demonstrated higher precision and recall than previous methods. In addition, our method rescued pathogenic structural variations ignored in previous studies. Our approach can serve as a next-generation pipeline for facilitating accurate and scalable genome analyses. Part 2 demonstrates a genomic analysis of ALK-rearranged non-small cell lung cancer. True-positive list of structural variations was easily obtained by using the method developed in Part 1. We identified that ALK fusion genes were mostly formed by complex genomic rearrangements. Compared to other non-small cell lung cancer, the gain of chromosome 5q was more recurrently observed in ALK-rearranged cases. The ALK rearrangements and 5q gain events were estimated to have occurred more than a decade earlier than diagnoses in most cases. ALK-rearranged lung cancer usually had low mutation burdens and few smoking-induced mutations regardless of smoking history. In addition, less biallelic inactivation of TP53 and more frequent TERT amplification were observed in ALK-rearranged lung cancer.