The technology is now available to measure the expression levels of a set of genes in a cell/tissue. Measurements taken from different "states" of the tissue (e.g. different developmental stages, environmental conditions, disease states) provide clues about the underlying regulatory pathways.

A great challenge for computational biology is to develop methods that are able to (partially) elucidate regulatory pathways given gene expression data.

Recently, a variety of methods have been proposed for addressing this problem. They differ mainly in how they represent (i.e. model) genetic networks. The method described by Weaver et al. is interesting because it can account for the "graded" nature of gene expression (unlike Boolean networks), yet it requires less data than other models (such as differential equations) that are able to represent this gradation.

• clusters
• Boolean networks
• differential equations
• Bayesian networks
• weight matrices

• transcription modeled as discrete state transitions
• assumes that regulatory interactions are independent (e.g. can't account for transcriptional regulators that have different activities depending on their protein partners)
• very similar to a single-layer neural network

"Learning" a model that explains a set of expression data involves finding values of the matrix Z such that the resulting weight matrix model closely accounts for the data.

The goal is to find a model that is very close to the true, underlying regulatory network.

M. Craven, June 10, 1999