The recent expansion of gene-expression and protein-interaction data sets means that we can predict increasingly complex models of gene-regulatory networks. But how does one recognize the correct model? Trey Ideker and colleagues have devised an automated method to minimize the effort of testing and refining gene-network predictions.

Combining gene-expression, promoter-binding and protein-interaction data, the authors constructed models of numerous gene-regulatory pathways in yeast. In some cases interactions between components were ambiguous — that is, more than one model was possible. One way to validate a particular prediction is to make genetic deletions of network components and determine whether the resulting changes in gene expression are consistent with the model. The authors developed an automated method that prioritized the deletion experiments that would provide the most information about their ambiguous models.

Three of the most informative experiments related to the same model — that for the regulatory pathway downstream of the yeast SW14 and SOK2 genes. Analysis of the deletions that were indicated by the automated system confirmed two predicted regulatory pathways within this model, whereas a third was rejected. Furthermore, when the results of these deletion experiments were combined with data from a previous analysis of 273 single-gene knockouts that was used to create the original model, many previously ambiguous interactions were resolved.

How does this method compare with other experimental validation strategies? Alternative methods might be to prioritize the deletion of hubs — genes that participate in many interactions — or to delete genes at random. The authors' automated method is more effective than either of these approaches at removing ambiguity. The method can also estimate the number of future experiments needed to fully refine a model.

The method described here has its limitations — it would be more useful if it could deal with experiments that involve the deletion of more than one gene. Nonetheless, this approach provides a useful tool for validating networks and could be made more powerful by extending its scope to multiple deletions in the future.