Abstract

Systems Biology is a field that models complex biological systems in order to better understand the working of cells and organisms. One of the systems modeled is the gene regulatory network that plays the critical role of controlling an organism's response to changes in its environment. Ideally, we would like a model of the complete gene regulatory network. In recent years, several advances in technology have permitted the collection of an unprecedented amount and variety of data such as genomes, gene expression data, time-series data, and perturbation data. This has stimulated research into computational methods that reconstruct, or infer, models of the gene regulatory network from the data. Many solutions have been proposed, yet there remain open challenges in utilising the range of available data as it is inherently noisy, and must be integrated by the inference techniques. The thesis seeks to contribute to this discourse by investigating challenges of performance, scale, and data integration.

We propose a new algorithm BENIN that views network inference as feature selection to address issues of scale, that uses elastic net regression for improved performance, and adapts elastic net to integrate different types of biological data. The BENIN algorithm is benchmarked on a synthetic dataset from the DREAM4 challenge, and on real expression data for the human HeLa cell cycle. On the DREAM4 dataset BENIN out-performed all DREAM4 competitors on the size 100 subchallenge, and is also competitive with more recent state-of-the-art methods. Moreover, on the HeLa cell cycle data, BENIN could infer known regulatory interactions and propose new interactions that warrant further experimental investigation.


Keys words: gene regulatory network, network inference, feature selection, elastic net regression.