Although the genome is mostly invariant in each cell of an organism, genes can have different expression patterns related to environmental conditions or developmental programs. For a given genome, different transcriptome states can be observed, depending on complex networks regulating adaptation and homeostasis. Fundamental questions about the topology of networks as protein interactome 1, metabolome 2 or transcriptional regulation networks 34 have been previously addressed. Interfaces like KEGG 5, Biocyc 6 or GenMapp 7 help biologists to put these results into a cellular context by mapping expression states onto metabolic network representations. Moreover, to visualize and edit networks, different tools like Cytoscape 8 or Osprey 9 are available. To understand the biology of the studied systems better, the trend is clearly towards the aggregation of multiple sources of biological information.

This article focuses on the analysis of transcriptome states. Numerous microarray gene expression datasets are now available, giving the opportunity to get a picture of the transcriptome state in various cellular conditions. Various methods to mine compendia of transcriptome states and to try to understand their biological meaning already exist. One of the most successful is based on the assumption that genes having similar expression profiles across a set of conditions are likely to be involved in the same biological process 10. Many approaches, including multivariate analysis, hierarchical clustering and SOM have been used to find patterns of expression and to create biologically relevant clusters (see 11 for review). By analysing whole microarray datasets, these "global" clustering approaches have already led to the formulation of interesting testable predictions 12. However, it has been previously emphasized that classical clustering methods could be inefficient on large number of biologically unrelated datasets 13. Indeed, in response to environmental changes, gene expression is modified only in a fraction of the transcriptome and the signal is hence diluted over the whole dataset. Few methods have been proposed to find these regulatory sub-signatures 1314, and in fine, the most reliable way to go deeper into the data to capture interesting trends is to be an expert in the field. That is why tools allowing the biologist to mine microarray results, in order to find such expression modifications in a sub-set of genes related to his area of expertise, are highly desirable. Such "gene-centric" clustering analysis that distinguishes differentially-expressed genes in specific parts of the transcriptome (centred around a "seed gene") should overcome some drawbacks of global analysis approaches.

The ideal tool would be gene-centric, and would provide understandable outputs mapping biological knowledge onto results. We present here MiCoViTo, a user-friendly tool for identifying and visualizing groups of genes having similar expression in two sets of microarray experiments representing two distinct transcriptome states.

As an illustration, we used MiCoViTo to compare genes co-expressed during the cell cycle (Cho el al. dataset 23) for two pairs of cyclin: CLN1/CLN2 and CLN1/CLB2 (Figure 2, B1). Results depend mainly on the proximity of seed expressions. Indeed, two genes close one to another like CLN1 and CLN2 (both implicated in the G1 phase), lead to the identification of numerous correlated genes also implicated in the G1 phase (Figure 2, B2) like PCL1 24, MCD1 25 or RNR1 26. But if CLN1 is compared to a gene implicated in another phase of the cycle like CLB2 (G2/M transition), no correlation at all is observed (Figure 2, B3). Moreover, the upper right and the lower left groups show genes implicated in the G2/M transition and the G1 phase like CLB2 and CLN1 respectively.

MiCoViTo aims to help biologists achieve an intuitive understanding of their own data using an approach based on the comparison of sets of microarray experiments. The user-friendly web interface is designed to be accessible to those with no particular technical skill.

The major drawback of visualization tools based on biologists' knowledge and intuition is that it does not give a computer readable output and cannot be applied systematically to find interesting expression patterns. Particularly, seed choice is a crucial point of our approach because it precisely defines which part of the transcriptome to focus on. The gene-centric approach presented in this paper empowers biologists to focus on a particular sub-part of the transcriptome. In order to assist seed selection process, online lists of candidates for each set of microarray experiments are proposed. Those candidates are ranked according to their neighbourhood density (density is the number of genes co-expressed with seed according to a given expression distance threshold). Genes with a large neighbourhood density are those whose nearest neighbourhoods contain a lot of genes, which means that their expression profiles are similar to the expression of many other genes. Such information may be used as a starting point when no prior information about the transcriptome state topology is available.

As initiated in a recent paper 27, one possible future direction is to use this approach to compare transcriptomes from different organisms captured in the same state. This will be possible if we can find functional annotation in the same format for two organisms and if we are able to define pairs of seeds having inter-organism correspondence. An annotation project like GO 28 addresses the first question. To define relevant inter-organism seeds, two ways appear realistic. The first one is to assume that sequence-based conservation implies function conservation. This is not always the case, but one can expect MiCoViTo results to be incoherent if seeds are not functionally related. Alternatively, one can start from pairs of genes previously described as orthologous. S. pombe could provide a good benchmark system to test this approach, since GO annotations are available and microarray datasets describing states previously studied in S. cerevisiae are publicly available. Moreover well-characterised S. cerevisiae/S. Pombe orthologous genes have been listed by the Sanger Center (Valerie Wood, personnal communication).

MiCoViTo allows users to compare and visualize different transcriptome states in a gene-centric way. The generated clusters of genes are mapped onto existing biological knowledge to gain a higher level view of the ongoing transcriptional changes. Upload of personal data onto the site is possible but not mandatory since a compendium of more than fifty yeast microarray datasets is available.

At present, this tool is restricted to S. cerevisiae, but a natural future direction will be to incorporate data originating from different species as well as orthology tables to allow comparison of seeds from different organisms.

MiCoViTo is available at http://www.transcriptome.ens.fr/micovito/. The source code and database scheme are freely distributed to academic users upon request to the authors.

A more detailed description of MiCoViTo including a step by step tutorial can be found online.

MiCoViTo: Microarray Comparison Visualization Tool; DTT: dithitreitol; MIPS: Munich Information center for Protein Sequences; RSA tools: Regulatory Sequence Analysis Tools; SGD: Saccharomyces Genome Database; TF: transcription factor; yMGV: yeast Microarray Global Viewer.

GL conceived, implemented MiCoViTo and drafted the manuscript, PMsuggested the first version of the visualization interface, contributed todiscussions and drafted the manuscript, PV and CJ are GL's Ph.D. advisors, they both contributed to discussions, SV provided help with algorithms and coordinated the project. All authors read and approved the final manuscript.