Statistical methods for the detection, analyses and integration of biomarkers in the human genome and transcriptome

Abstract: Most human diseases have been shown to have a genetic basis that is linked to regulation of gene expression at the transcriptional or post-transcriptional level. In the central dogma of biology, deoxyribonucleic acid (DNA) is transcribed to messenger ribonucleic acid (mRNA), and then translated into proteins; dysfunction in any of these processes may contribute to the development of disease. Sources of such potential irregularities include, but not limited to, the following: point mutations in DNA sequences, copy number alterations (CNAs) and abnormal mRNA and microRNAs (miRNAs) expression. MiRNAs are a type of non-coding RNA that inhibit the transcription and/or translation of specific target mRNAs. Current technologies allow the identification of biomarkers and study of the complex interplay between DNA, mRNA, miRNA and phenotypic variation. This thesis aims to tackle the statistical challenges that have arisen with the application of these technologies to investigate various genomic and transcriptomic alterations. In study I, modified least-variant set normalization for miRNA microarray, a new algorithm and software were developed for microRNA array data normalization. The algorithm selects miRNAs with the least array-to-array variation as the reference set for normalization. The selection process was refined by accounting for the considerable differences in variances between probes. Data are provided to show that this algorithm results in better operating characteristics than other methods. In study II, joint estimation of isoform expression and isoform-specific read distribution using multi-sample RNA-Seq data, a joint model and software were developed to estimate isoform-specific read distribution and gene isoform expression, using RNA-sequencing data from multiple samples. Observation of similarities in the shape of the read distributions solves the problem that the non-uniform read intensity pattern is not identifiable from the data provided by one sample. In study III, integrated molecular portrait of non-small cell lung cancers, molecular markers at the DNA, mRNA and miRNA level that can distinguish between different histopathological subtypes of non-small cell lung cancer were identified. Additionally, using integrated genomic data including CNAs and mRNA and miRNA expression data, three potential driver genes were identified in non-small cell lung cancer, namely MRPS22, NDRG1 and RNF7. Furthermore, a potential driver miRNA, hsa-miR-944, was identified. In study IV, integration of somatic mutation, expression and functional data reveals potential driver genes predictive of breast cancer survival. An analytic pipeline to process large-scale whole-genome and transcriptome sequencing data was created, and an integrative approach based on network enrichment analyses to combine information across different types of omics data was proposed to identify putative cancer driver genes. Analysis of 60 patients with breast cancer provided evidence that patients carrying more mutated potential driver genes had poorer survival.