Building a genomic variant based prediction model for lung cancer toxicity

Abstract: Since the completion of the the Human genome project in 2003, the evident complexity of our genome and its regulation has only grown. The idea that having sequenced the human genome would solve this mystery was quickly discarded. With the decreasing costs of DNA sequencing, a plethora of new methods have evolved to further understand the role of non-coding regions of our genome, which makes up 98% its length. Genetic variations in these regions are therefore abundant in the human population, but their e ects are hard to characterize. Many non-coding variants have been linked to complex diseases such as cancer predisposition. This thesis aims to investigate the potential e ects of non-coding variants on drug toxicity, that is, how severe the adverse e ects of a drug are to the treated patients. More specifically it will study the effects of two cancer drugs, Gemcitabine and Carboplatin, on a set of 96 patients with lung cancer. To do this we use spatial data acquired by the promoter-targeting method HiCap as well as expression data obtained from blood cell lines. Using the variants obtained through whole genome sequencing of the patients, a supervised learning approach was attempted to predict the final toxicity experienced by the patients. The large number of variants present among the comparably few patients resulted in poor accuracy. The conclusion was drawn that the resolution of HiCap is too low compared to the density of variants in the non-coding regions. Additional data, such as transcription factor Chip-Seq data, and transcription factor motifs are needed to locate potentially contributing variants within the interactions.