Interactive Inference and Experimental Design

This section is based on a paper [5] by T. Iddeker, V. Thorsson and R. Karp. Our goal is to infer the underlying genetic networks from a series of steady-state gene expression profile for a set of perturbations. We assume the Boolean genetic network model for the gene network. Moreover, we shall restrict ourselves only to acyclic networks. In the case of acyclic networks there is no need for assumptions about the time delays of the components. Moreover, even if most networks do have feedback loops, there are generally few of them, and the main pathway is often acyclic. So this model may be able to reconstruct informative networks with few feedback loops. The analysis of cyclic networks is complicated by the possibility of oscillatory behavior. For cyclic networks, one may adopt either a synchronous model in which each component has a fixed, known delay, or an asynchronous model in which the delays are unknown and even nondeterministic. The proposed strategy is based on repeated and interactive application of two analytical methods: the predictor and the chooser. According to this strategy, the underlying network of interest is exposed to an initial series of genetic and/or biological perturbations and a steady-state gene expression profile is generated for each. Next, a method called the predictor is used to infer one or more hypothetical Boolean networks consistent with these profiles. When several networks are inferred, the predictor returns only the most parsimonious, as measured by those networks having the fewest number of interactions. Depending on the complexity of the genetic network and the number of initial perturbations, numerous hypothetical networks may exist. Accordingly, a second method called the chooser is used to propose an additional perturbation experiment to discriminate among the set of hypothetical networks determined by the predictor. The two methods may be used iteratively and interactively to refine the genetic network: at each iteration, the perturbation selected by the chooser is experimentally performed to generate a new gene expression profile, and the predictor is used to derive a refined set of hypothetical gene networks using the cumulative expression data.