Goutsias and Lee have a paper in Curr. Pharma Design (2007) 13:1415-1436 which is an interesting mix of a review with some new results. I’ll focus here on their review of computational and experimental approaches, as I think they do a pretty good job of categorizing approaches to the network modeling problem. Their “gets us grants” application is to try to tackle colon cancer as the system of interest … but I refer you to their paper for those details.

Networks are ubiquitous in biology (genes, proteins, post-translational, metabolic, cellular, organisimal, social, etc). Knowing both the connectivity and dynamics of a network are important to understanding the functions and constraints of the system. Attempts to tackle regulatory networks have come from experimental techniques (expression profiling, motif discovery, ChIP-whatever, and RNAi) and computational techniques. Computational approaches concentrate on a particular level of focus … be it the high level connectivity (gene networks) or the more detailed stochastic and thermodynamic molecular models. Following Goutsias’s terminology:

- Gene networks are directed graphs where the edges do not necessarily represent direct interactions. The goal is to capture interactions among genes, the connectivity. The canonical examples include the Lee et. al. and Harbison papers in S. cerevisiae.

- At a more detailed level are transcriptional regulatory systems. These utilize mathematical models of a dynamic system, typically first order rate equations. They require having some idea of the functional form for regulation. They capture both qualitative (connectivity) and quantitative (kinetic) aspects of the system … but require significantly more detail (parameters).

- A simplified type of transcriptional regulatory system is the boolean network. Here the parameters are reduced to binary valued functions and the interactions are modeled as logical functions. This is effectively a discrete time model approach that grossly simplifies the system. Therefore, only qualitative descriptions arise for these models.

- The most realistic models are those that not only consider the kinetic aspects of the system but also the inherent stochastic variability. For example, dynamic bayesian networks use conditional probability density functions to quantify uncertainty. These models are untractable unless some amount of the underlying system is assumed, typically the connectivity.

In all cases, these models tend to assume acyclic network topologies and hence linear systems of equations can model the transcriptional dynamics well. This is a limitation, however, for most complex biological systems. Gene perturbations are key to dissecting system behavior.

Goutsias, J., Lee, N.H. (2007). Computational and Experimental Approaches for Modeling Gene Regulatory Networks. Current Pharmaceutical Design, 13(14), 1415-1436. DOI: 10.2174/138161207780765945

## Leave a Reply