Cellular organization is believed to be modular, meaning cellular functions are carried out by modules composed of clusters of genes, proteins and metabolites that are interconnected, co-regulated or physically interacting. In turn, these modules interact together and thereby form complex networks that taken together is considered to be the interactome. 

Modern high-throughput biological techniques have made high-scale accurate quantification of these biological molecules possible, the so called omics. The simultaneous measurement of these molecules enables a picture of the state of a cell at a resolution that was never before possible. Mapping these measurements aids greatly to elucidate a network structure of interactions. The ever growing size of public repositories for omics data has ushered in the advent of biology as a (big) data science and opens the door for data hungry machine learning approaches in biology. 

Complex diseases are multi-factorial and arise from a combination of genetic, environmental and lifestyle factors. Additionally, diagnosis and treatment is complicated by the fact that these genetic, environmental and lifestyle factors can vary between patients and may or may not give rise to different disease phenotypes that still classify as the same disease. Genetically, there is substantial heterogeneity among patients and therefore the emergence of a disease phenotype cannot be attributed to a single genetic mutation but rather to a combination of various mutations that may vary from patient to patient. As complex diseases can have different root causes but give rise to a similar disease phenotype, the implication is that different root causes perturb similar components in the interactome. Most of the work in this thesis is aimed at developing methods and computational pipelines to identify, analyze and evaluate these perturbed disease specific sub-networks in the interactome, so called disease modules. 

We started by collecting popular disease module inference methods and combined them in a unified framework, an R package called MODifieR (Paper I). The package uses standardized inputs and outputs, allowing for a more user-friendly way of running multiple disease module inference methods and the combining of modules. Next, we benchmarked the MODifieR methods on a compendium of transcriptomic and methylomic datasets and combined transcriptomic and methylomic disease modules for Multiple Sclerosis (MS) to a highly disease-relevant module greatly enriched with known risk factors for MS (Paper II). Subsequently, we extended the functionality of MODifieR with software for transcription factor hub detection in gene regulatory networks in a new framework with a graphical user interface, MODalyseR. We used MODalyseR to find upstream regulators and identified IKZF1 as an important upstream regulator for MS (Paper III). Lastly, we used the growing large-scale repositories of gene expression data to train a Variational Auto Encoder (VAE) to compress and decompress gene expression profiles with the aim of extracting disease modules from the latent space. Utilizing the continues nature of the latent space in VAE’s, we derived the differences in latent space representations between a compendium of complex disease gene expression profiles and matched healthy controls. We then derived disease modules from the decompressed latent space representation of this difference and found the modules highly enriched with disease-associated genes, generally outperforming the gold standard of transcriptomic analysis of diseases, top differentially expressed genes (Paper IV). 

To conclude, the main scientific contribution of this thesis lies in the development of software and methods for improving disease module inference, the evaluation of existing inference methods, the creation of new analysis workflows for multi-omics modules, and the introduction of a deep learning-based approach to the disease module inference toolkit.