From Statistical to Biological Interactions via Omics Integration
Year: 2016
Type: dissertation
Abstract: The XXI century opened a new ‘Big Data’ era in which, thanks to rapid technological advancements and appearance of high throughput technologies, vast amounts of unstructured omics data (e.g., transcriptomic, genomic, etc.) are generated every day. This thesis mainly focuses on solving the problems related diverse omics data integration and interaction identification tasks. Particular attention is given to useful knowledge extraction in the context of complex diseases including pathological mechanisms with the development of software tools and pipelines. The diseases covered included glioblastoma multiforme, asthma, and ankylosing spondylitis. Interactions detection in genomic data requires standardization of the protocols. In Chapter 3, we tested the impact of different settings in a genome-wide association interaction study (GWAIS). Some of the settings included marker selection strategy, the LD pruning, lower order effects adjustment, analytical tool. We were able to show that even small changes in each setting can have drastic impacts requiring careful assessment of proper settings and results comparisons across several analysis protocols. The greatest impact was attributed to the input dataset composition highlighting the importance of the marker selection strategy and use of prior knowledge. Expression of genes can be affected by nearby (‘cis’) or distant (‘trans’) genotypes. Thus, we developed methodology to identify complex trans/cis regulatory mechanisms between expression and genotype data in the context of asthma (CAMP data). Significant overlap between ‘trans’ and ‘cis’ gene regulatory components related to immune and signaling pathways was clearly identified matching asthma disease pathology. The semi-parametric Model-Based Multifactor Dimensionality Reduction (MB-MDR) method was applied for the first time in the context eQTL study achieving low false discovery and family-wise error rates (FDR and FWER). Identification of a meaningful data structure from omics data is a pressing topic nowadays. Gene regulatory networks (GRN) conveniently summarize large amounts of data allowing for useful knowledge generation. GRN inference is especially attractive for deciphering of complex diseases mechanisms allowing biologists to formulate a better hypothesis. We were able to generate GRNs from a single source (e.g., microarray expression data) using conditional inference forest (CIF) with more attractive features compared to classical Random-Forest (RF) including unbiased node variable selection even in the context of highly correlated variables particularly relevant in transcriptomics. The CIF methods provided attractive features and performance characteristics coupled to valuable pathological insights into type 1 diabetes.
Source: ORBi (University of Liège)
Author: Kyrylo Bessonov
Institution: University of Liège
Language: English
Subfield: Genetics
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W2464617931