Phenotypic traits are now known to stem from the interplay between genetic variables across many if not every level of biology. The field of gene regulatory network (GRN) inference is concerned with understanding the regulatory interactions between genes in a cell, in order to build a model that captures the behaviour of the system. Perturbation biology, whereby genes or RNAs are targeted and their activity altered, is of great value for the GRN field. By first systematically perturbing the system and then reading the system's reaction as a whole, we can feed this data into various methods to reverse engineer the key agents of change.

The initial study sets the groundwork for the rest, and deals with finding common ground among the sundry methods in order to compare and rank performance in an unbiased setting. The GeneSPIDER (GS) MATLAB package is an inference benchmarking platform whereby methods can be added via a wrapper for testing in competition with one another. Synthetic datasets and networks spanning a wide range of conditions can be created for this purpose. The evaluation of methods across various conditions in the benchmark therein demonstrates which properties influence the accuracy of which methods, and thus which are more suitable for use under given characterized condition.

The second study introduces a novel framework NestBoot for increasing inference accuracy within the GS environment by independent, nested bootstraps, \ie repeated inference trials. Under low to medium noise levels, this allows support to be gathered for links occurring most often while spurious links are discarded through comparison to an estimated null distribution of shuffled-links. While noise continues to plague every method, nested bootstrapping in this way is shown to increase the accuracy of several different methods.

The third study applies NestBoot on real data to infer a reliable GRN from an small interfering RNA (siRNA) perturbation dataset covering 40 genes known or suspected to have a role in human cancers. Methods were developed to benchmark the accuracy of an inferred GRN in the absence of a true known GRN, by assessing how well it fits the data compared to a null model of shuffled topologies. A network of high confidence was recovered containing many regulatory links known in the literature, as well as a slew of novel links.

The fourth study seeks to infer reliable networks on large scale, utilizing the high dimensional biological datasets of the LINCS L1000 project.  This dataset has too much noise for accurate GRN inference as a whole, hence we developed a method to select a  subset that is sufficiently informative to accurately infer GRNs. This is a first step in the direction of identifying probable submodules within a greater genome-scale GRN yet to be uncovered.