Marc Vidal and colleagues respond:

The distorting effect of limited sampling on the overall topology of interactome models, relative to the actual network they are supposed to represent, is a subject of great interest. Our paper showed that out of four different topological models examined, scale-free networks were most likely to produce scale-free samples, although we could not formally exclude any of the other models by the criteria tested. Friedel and Zimmer suggest that limited sampling not only affects the degree distribution of a network, but can also distort its average local clustering coefficient (CC)1.

Their study concurs with our conclusions in two respects: first, partial networks can have different topological properties relative to the complete network they are derived from; and second, the power-law model is the most likely of the tested models. By demonstrating that the exact topology of the interactome could not be determined unambiguously from currently available data sets, our initial report is a reminder that the quality and coverage of interactome network maps needs to be improved. To question the generally accepted view that actual interactome networks are indeed scale-free, given the topology of the current maps, only one alternative model is needed and we proposed three of them. On the contrary, to conclude that the interactome can only be scale-free, as suggested by Friedel and Zimmer, would require proof by negation. Out of the large number of potential interactome topologies, how many can be excluded because they are inconsistent with the observations drawn from available maps, and what are the common properties of the remaining ones? A high CC could be one of these characteristics, in which case some scale-free models would remain good candidates. However, many other possibilities have yet to be excluded before the issue is resolved.

We suggest that before concluding that “the underlying network very likely also is scale-free and highly clustered,” it would be appropriate to test whether other topological models with high CC could give rise to the topology observed in currently available interactome maps. For instance, is it sound to limit the rewiring of a scale-free model to generate a highly clustered scale-free model network so as to avoid “a deviation from the power-law behavior for nodes with low degrees”? Could the sampling of such networks result in sampled networks having topological properties consistent with those of the protein-protein interaction maps currently available?