Reconstructing gene regulatory networks from micro-array data is important for understanding the underlying mechanism behind cellular processes. Numerous studies have been devoted to addressing this problem. One of the more popular methods is to view the gene regulatory network as a Bayesian inference network, and to apply structure learning methods to determine the topology of the gene regulatory network. There are, however, several major defects with the Bayesian structure learning approach for reconstructing gene regulatory networks such as the computational cost, non-unique solutions, does not take into account the special topology of gene regulatory network, and inefficiency in exploiting prior knowledge.

In this paper, we present a knowledge driven matrix factorization framework for reconstructing modular gene regulatory network. Unlike the Bayesian network approach, this approach views the reconstruction of gene regulatory network as a matrix factorization problem. The key advantage of the proposed framework in comparison to a Bayesian approach is its computational efficiency and power in exploiting the domain knowledge. Gene expression data is initially used to estimate the correlation matrix W. A matrix factorization algorithm is developed to decompose W into W=MCM' with a priori knowledge of gene ontology integrated to regularize M and C matrices. Here, matrix M provides the information about the gene regulatory modules. Each element of M indicates the confidence of assigning a gene to a module. Matrix C reveals the regulatory interaction between modules. Each element of C indicates the likelihood of two regulatory modules to interact with each other.

The matrix factorization approach was initially applied to yeast cell cycle expression data and the resulting modules were compared to that of the Bayesian approach, e.g. using the gene module network in Genomica [1]. Upon comparing the precision and recall, our approach performed better than the Bayesian method. We then applied the matrix factorization approach to gene expression from HepG2 cells exposed to different free fatty acids and identified biologically meaningful modules [2]. For example, all the genes in the TCA cycle were found in one module. Modules containing genes relevant to electron transport chain complex I, electron transport chain complex IV, ATP synthesis, were similarly identified. Analyzing the C matrix identified the connection or interaction between the modules. We found that many relevant modules predicted by this approach were indeed biologically related, for example the TCA module was found to have a strong interaction with the complex I, complex IV, and ATP modules. Indeed, it is well-known that these four processes are connected and involved in energy production. In summary, we have developed a knowledge driven matrix factorization approach that can simultaneously identify gene modules and how these modules are connected in a network.

1. Segal E. et al. (2004) A module map showing conditional activity of expression modules in caner, Nature Genetics, 36 (10):1090-1098. 2. Srivastava, S.; Chan, C. (2007) “Hydrogen peroxide and hydroxyl radical mediate palmitate-induced cytotoxicity to hepatoma cells: relation to MPT", Free Radical Research, Free Radical Research, 41(1),38-49.