Not surprisingly, there are several different ideas on how to divide and conquer high-throughput functional genomics data. I will restrict my discussion here to gene expression data, although similar remarks could be made for sequence data. Conceptually there are two distinct problems. One is: given a module of disease-associated genes, how can we compute and/or experimentally predict which genes are good candidates for therapeutics? Before discussing this problem I will first give an overview of different approaches to the other problem: identifying a module of genes.

A module is a group of genes that are related in some way to each other and therefore a module is effectively a measure of similarity. Grouping genes into modules depends on an exact mathematical definition of similarity. For example, if similarity is defined as the distance in a network, then a graph theoretical calculation will be used. However, if gene functional associations are used, then gene similarity will be measured in terms of gene ontology (GO) or correlation in gene expression values. Therefore, different algorithms are used for dividing the genes into modules, a fact that could be confusing for the clinical researcher.

The need to reduce the complexity of the original high-throughput gene expression data was realized early on in its analysis 2. Applying established engineering concepts, such as principal component analysis (PCA) and singular value decomposition (SVD), reduced the dimensionality of the data. Instead of analyzing scattered points (the samples) in a high-dimensional space equaling the number of genes, the data could thereby be projected into a two- to four-dimensional space. However, it turned out to be difficult to make a biological interpretation of the resulting linear combinations of large numbers of genes. This problem forced the development of different strategies in which the available knowledge on a limited number of genes could be used to predict the functions of as-yet uncharacterized genes.

The use of hierarchical clustering in the classic compendium study on yeast data by Rosetta Inpharmatics 3 grouped genes (shown as rows) by their similarity of expression across several experimental conditions (columns). Novel gene function was then predicted by inspecting genes in the same cluster as genes with known functions. Subsequent work by Eran Segal and colleagues 4 developed more statistically sound procedures for identifying robust modules using a Bayesian formalism applied to microarray data generated from cancer samples.

It became clear, however, that a similarity measure based only on correlations was insufficient, because the clusters (modules) or Bayesian modules did not have an internal network structure that could be used for a more refined analysis. As a consequence, a large number of studies addressing this problem appeared in the literature at the beginning of 2004. The idea was that if we could identify the wiring within cellular networks, various different algorithms could be applied to find 'connected groups' in such networks. Such an analysis would then provide more biological insights into the mechanisms of disease.

Now, how can such networks be found using only a small number of experimental samples with a large number of genes? This is an impossible problem from the point of view of engineering system identification, because the number of possible networks consistent with the data is prohibitively large 5. The key simplifying insight came from Ideker and Lauffenburger 6 and was later developed by Nicolas Luscombe and colleagues in a pioneering paper 7. Here, the edges (or connections) in the network were simply defined by transcription factor binding experiments, and gene expression data were used to select the subsets of edges that were active under different conditions.

This idea of defining edges in a network using a static scaffold has since been reused using various data types (protein-protein interaction data, pathways from a database, text mining and DNA variants). The network of interest is then defined by combining the gene expression data with the scaffold, leaving only the active edges. By searching through such an active network using graph algorithms it is then possible to define 'more' connected parts in a well defined manner, thereby providing modules with an intrinsic network structure.

All the above approaches basically begin with a large, complex dataset, which is then simplified by dividing the data into smaller modules. Interestingly, Vidal and colleagues 8 demonstrated that this process can be reversed. They instead began with four well characterized breast cancer genes and, by using these ideas, constructed a module in which the genes were 'close' as defined by expression and proteomic data in several species.

Benson and colleagues 1 have now contributed to a disease-oriented modular analysis by combining several of the above ideas in a novel manner, as summarized in the flow chart in Figure 1. First, because allergic disease involves multiple cells in different tissues and because no prior characterization of key genes was available, they turned to several different sets of gene expression microarray data in order to find a reference disease-associated gene around which they could construct a module. Using the idea that disease-associated genes tend to interact, they could search for other disease-associated genes that were 'close'. For this purpose, the authors used a graph algorithm that identified a connected clique of 103 disease-associated genes from the microarray data.

The T-cell receptor signaling pathway turned out to be a pathway shared by these 103 genes, as detected by the Ingenuity Pathway Analysis tool, which identifies physical, transcriptional and enzymatic interactions from the literature 1. Experimental analysis of this pathway in patient-derived cells revealed strong activation of the ITK gene, which is also known to be located in the genomic susceptibility region for allergy. Combining a promoter analysis of the ITK gene with expression data revealed that the transcription factor GATA3 regulated ITK.

Finally, using available databases, 47 genes were identified as interacting with GATA3. The expression data were used to filter out 10 inactive genes, thus leaving a final module of 37 disease-associated genes around the GATA3 transcription factor 1. The construction of this module was accompanied by several experimental tests at various stages, providing confidence to the analysis.

The problem of selecting therapeutic targets within a module has not received much attention in studies that have used a modular approach for reducing complexity. There are various ideas from graph theory on how to compute mathematically defined properties, such as clustering and connectivity in large networks, which then could suggest which nodes are essential. However, essentiality is not necessarily equivalent to disease association. Experimental investigators have instead performed target selection using the full dataset in combination with extensive experimental testing. This is, by most measures, an inefficient and expensive procedure.

The analysis by Benson and colleagues 1 is important because it highlights the difficulty of selecting a disease-associated target from a module of 37 genes despite the elegant prior reduction of complexity. They resorted to using a connectivity criterion, selecting the IL7R gene because it had the largest number of connections, and they were also able to demonstrate that perturbing the IL7R gene affected other genes and the T-cell phenotype. There are probably several other disease-associated genes in their module that warrant further experimental investigation.

Benson and colleagues 1 have introduced a useful procedure for defining a module of disease-associated genes. As with most complex diseases, the study of allergy is complicated by the fact that the disease affects several cell types and tissues. The process of identifying such modules therefore requires the kind of stringent experimental validation as was performed by the Benson team 1. Despite their careful analysis, because there are other transcription factors for the ITK gene that are active in the expression datasets there is a significant risk that several disease-associated genes remain that were not captured in their module.

The second step of selecting a gene for therapeutics from a module is even more problematic because we are currently lacking systematic tools for this selection problem. Furthermore, it is not unlikely that an efficient therapy could require targeting of several disease-associated genes simultaneously. However, the number of combinations of three genes that can be chosen from a small ten-gene module, for example, quickly exceeds what is experimentally feasible to study.

In conclusion, Benson and colleagues 1 have devised an interesting method for finding disease-associated genes, but it needs to be evaluated on other complex diseases. Their study also makes clear that the problem of prioritizing disease-associated genes within a module for therapeutic studies in the clinic is still unsolved.