Network-based identification of driver pathways in clonal systems

Highly ethanol-tolerant bacteria for the production of biofuels, bacterial pathogenes which are resistant to antibiotics and cancer cells are examples of phenotypes that are of importance to society and are currently being studied. In order to better understand these phenotypes and their underlying genotype-phenotype relationships it is now commonplace to investigate DNA and expression profiles using next generation sequencing (NGS) and microarray techniques. These techniques generate large amounts of omics data which result in lists of genes that have mutations or expression profiles which potentially contribute to the phenotype. These lists often include a multitude of genes and are troublesome to verify manually as performing literature studies and wet-lab experiments for a large number of genes is very time and resources consuming. Therefore, (computational) methods are required which can narrow these gene lists down by removing generally abundant false positives from these lists and can ideally provide additional information on the relationships between the selected genes.

Other high-throughput techniques such as yeast two-hybrid (Y2H), ChIP-Seq and Chip-Chip but also a myriad of small-scale experiments and predictive computational methods have generated a treasure of interactomics data over the last decade, most of which is now publicly available. By combining this data into a biological interaction network, which contains all molecular pathways that an organisms can utilize and thus is the equivalent of the blueprint of an organisms, it is possible to integrate the omics data obtained from experiments with these biological interaction networks. Biological interaction networks are key to the computational methods presented in this thesis as they enables methods to account for important relations between genes (and gene products). Doing so it is possible to not only identify interesting genes but also to uncover molecular processes important to the phenotype.

As the best way to analyze omics data from an interesting phenotype varies widely based on the experimental setup and the available data, multiple methods were developed and applied in the context of this thesis:

In a first approach, an existing method (PheNetic) was applied to a consortium of three bacterial species that together are able to efficiently degrade a herbicide but none of the species are able to efficiently degrade the herbicide on their own. For each of the species expression data (RNA-seq) was generated for the consortium and the species in isolation. PheNetic identified molecular pathways which were differentially expressed and likely contribute to a cross-feeding mechanism between the species in the consortium.

Having obtained proof-of-concept, PheNetic was adapted to cope with experimental evolution datasets in which, in addition to expression data, genomics data was also available. Two publicly available datasets were analyzed: Amikacin resistance in E. coli and coexisting ecotypes in E.coli. The results allowed to elicit well-known and newly found molecular pathways involved in these phenotypes.

Experimental evolution sometimes generates datasets consisting of mutator phenotypes which have high mutation rates. These datasets are hard to analyze due to the large amount of noise (most mutations have no effect on the phenotype). To this end IAMBEE was developed. IAMBEE is able to analyze genomic datasets from evolution experiments even if they contain mutator phenotypes. IAMBEE was tested using an E. coli evolution experiment in which cells were exposed to increasing concentrations of ethanol. The results were validated in the wet-lab.

In addition to methods for analysis of causal mutations and mechanisms in bacteria, a method for the identification of causal molecular pathways in cancer was developed. As bacteria and cancerous cells are both clonal, they can be treated similar in this context. The big differences are the amount of data available (many more samples are available in cancer) and the fact that cancer is a complex and heterogenic phenotype. Therefore we developed SSA-ME, which makes use of the concept that a causal molecular pathway has at most one mutation in a cancerous cell (mutual exclusivity). However, enforcing this criterion is computationally hard. SSA-ME is designed to cope with this problem and search for mutual exclusive patterns in relatively large datasets. SSA-ME was tested on cancer data from the TCGA PAN-cancer dataset. From the results we could, in addition to already known molecular pathways and mutated genes, predict the involvement of few rarely mutated genes.