Latent-variable models for drug response prediction and genetic testing

High-throughput DNA sequencing and related biotechnologies revolutionized our understanding of human genomics and diseases with genetic component, particularly of cancer -- one of the leading causes of death world-wide. Despite the progress in cancer research and availability of over 150 FDA-approved anti-cancer drugs, the cancer treatment is often unsuccessful. Identifying the best cancer treatment using computational models to personalize drug response prediction holds great promise to improve patient’s chances of successful recovery. Unfortunately, the computational task of predicting drug response remains very challenging. In this thesis I develop a deep latent-variable machine learning model with amortized variational inference that improves accuracy of drug response prediction over the currently used models. Besides increased expressiveness of this model thanks to parameterization by neural networks, the achieved improvement stems from integration of drug-induced perturbation profiles, a resource not fully utilized before. Clinical trial datasets of cancer treatments which also include genomic characterization of the tumours are small and scarce, therefore for the vast majority of drugs only responses in pre-clinical biological models are available. To this end, I assess applicability of popular domain adaptation approach, based on domain-invariant representation learning, to the drug response prediction task. I conclude that necessary conditions of successful domain adaptation are often not satisfied in the available datasets and as such many current methods are misguided. Last but not least, in this thesis I also contribute to the area of non-invasive prenatal testing. Using a hidden Markov model I propose a method for analysis of cell-free DNA fragments isolated from maternal plasma that also contain admixture of DNA fragments derived from the fetal genome. Here presented method is a first proof-of-concept for non-invasive sub-chromosomal CNV detection.Ph.D