A Statistical Framework for Incorporating Multi-Omics Technologies for Precision Medicine

This thesis addresses important statistical challenges in precision medicine, the clinical practice to customise treatment plans for individual patients using genetic information. We propose methods, frameworks and procedures that tackle the discovery, translation and implementation of precision medicine through the use of omics data. Specifically, we examine three key aspects of precision medicine: (1) The use of targeted assays as opposed to whole-transcriptome technologies like microarray and bulk RNA-Sequencing. In particular, we focus on the impact of a gene-set selection bias on classical gene-set tests in targeted assays; (2) The challenges of stable feature selection methods when dealing with mediumthroughput targeted assays; and (3) The development of a prediction model on noisy and batched gene expression data from multiple sources that can be implemented under clinical constraints. In Chapter 1, we provide a brief overview of precision medicine with a special focus on targeted assays as a potential instrument for large-scale implementation. We present an overview of the statistical challenges that must be overcome in three phases of precision medicine in order to facilitate the final implementation. Chapter 2 of this thesis focuses on adapting the classical gene-set tests for targeted assays by proposing a new method, bcGST (bias-corrected Gene Set Test). We design synthetic simulations in eleven publicly available whole-genome expression data and show how bcGST improves over cases where the gene-set selection bias is ignored. Most of this chapter was published in Wang et al. (2019a). 3 In Chapter 3 we propose a novel variable selection method, APES (APproximated Exhaustive Search), for generalised linear models (GLMs). APES modernises the classical exhaustive variable selection problem for GLMs, enabling model selection with hundreds of variables. It is therefore capable of analysing data from commercially available targeted assays. The advantage of APES lies in its ability to approximate a genuine exhaustive search at a dramatically improved speed. We devise a comprehensive set of simulations to test APES’s performance and apply it to a real targeted assay. Most of this chapter was published in Wang et al. (2019b). Chapter 4 is motivated by irreproducible results in research due to inconsistencies in the scaling of omics data and the corresponding tuning of parameters in prediction models. We first define a notion of “transferability” to describe a model’s ability to maintain high predictive power across multiple omics datasets with no additional manipulations on the model. Then, we propose the Cross-Platform Omics Prediction (CPOP) procedure that constructs transferable models. These models have biologically relevant features that are statistically stable with respect to between-data noise. CPOP is specially designed for prediction across different omics platforms and prospective experiments that are commonly seen in clinical settings. We curate four melanoma datasets and a prospective targeted assay experiment to illustrate the novelty of CPOP. This thesis concludes with some final remarks in Chapter 5. In summary, this thesis contributes to precision medicine research by developing relevant, interpretable and implementable statistical methods.