Development and Application of Methods for Cancer Genome Analysis

Abstract

Cancer is a disease of the genome. The transition from healthy to malignant cell is driven by the gradual accumulation of genetic mutations that drive tumorigenesis. Identification of these cancer driver events and characterization of the cancer genome, using next generation sequencing, has driven a revolution in our understanding and the treatment of cancer. Central to these advances are oncologists who leverage the tumors’ unique genomic mutations to personalize treatment, and cancer genome analysts who sequence tens of thousands of tumors to identify novel therapeutic opportunities. Clinical care grounded in genomic profiling requires sensitive and specific somatic mutation detection to ensure well-reasoned decision-making. A key step in somatic mutation detection is comparison of the tumor sample to a matched germline control. Sensitivity to detect somatic variants is greatly reduced when the matched normal sample is contaminated with tumor cells. Tumor in normal contamination can happen frequently in clinical settings where tissue is scarce. To overcome this limitation, we developed deTiN, a method that estimates tumor-in-normal contamination, and improves detection sensitivity when using a contaminated normal. Through simulation, we show that deTiN accurately estimates TiN levels and improves mutation detection sensitivity. We leverage deTiN to analyze a cohort of clinical chemo-resistant testicular germ-cell tumors. Using deTiN and other methods for interrogation of the cancer genome, we show that the primary somatic feature of germ-cell tumors is highly recurrent amplifications of one chromatid paired with deletions of its sister. These are significantly enriched in germ-cell tumors when compared to 6,509 tumors across 19 cancer types from The Cancer Genome Atlas (TCGA). This copy number alteration signature is active in germ-cell tumors throughout their development and is associated with loss of pluripotency markers in chemoresistant recurrent tumors. Also critical in improving genomic analysis of tumors, is the ability to quickly analyze large cohorts to identify new therapeutic targets. We demonstrate that recently developed machine-learning libraries facilitate implementation of methods for GPUs and accelerate computations. Our implementations of methods for cancer genome analysis, using these libraries, ran >200 times faster than current versions. Transition to GPU-based implementations for genomics methods will enable investigation of previously unanswerable hypotheses involving more complex models, larger datasets, and more accurate empirical measurements, all requiring significantly more computation.