Stanford University School of Medicine, California, USA

A systems biologist encourages modelling by the millions.

In a typical modelling study, we write down equations, solve them, and see whether they account for known data. If they do, we claim to understand some bit of biology. One huge caveat is that many other models might have matched the data just as well.

Researchers from Peking University in Beijing and the University of California, San Francisco, have devised a satisfying way of dealing with this problem (W. Ma et al. Mol. Syst. Biol. 2, 70; 2006).

Their starting point was epithelial patterning in the fruitfly Drosophila. During embryogenesis, a system known as the 'segment polarity network' generates repeating stripes of gene expression. The stripes are initially fuzzy and later become sharp. Ma et al. set out to see what simple gene circuits were best suited to this sharpening process.

They formulated differential-equation models for about 14 million ways of connecting two or three segmentation genes, then randomly chose 100 sets of parameters that defined the strength of the interactions for each gene. They then carried out computations for each combination to determine which of them converted fuzzy stripes into sharp ones.

Many topologies worked for at least one parameter set. But only a fraction worked for more than one or two. Interestingly, the most robust topologies were all variations on the same design — each had three sub-circuits, one 'stripe generator' motif and two bistable 'response sharpeners'. These findings give hope that complex networks may be decomposed into modular sub-circuits with understandable functions.

Comprehensively examining millions of models is a lot of work, but is not impossible. And, as Ma et al. show, it can yield important insight that could not have been derived from studies of one or two.