The list of genetic mutations associated with human disease is long and continues to grow, and with it the challenges of deciphering what actually goes wrong in the cells, tissues and bodies of sick people. The problem is that the relationship between genotype and phenotype, in health or in disease, is typically very complex. It is the rare exception when a single mutation in a single gene causes disease.

One way to begin to probe this problem is to ask how disease mutations affect the way in which proteins interact. This is what Haiyuan Yu set out to do when he moved to Cornell University to start his laboratory. Yu and his colleagues began with a carefully curated map of directly interacting human proteins, using information both from the literature and from well-verified high-throughput yeast two-hybrid datasets. But just physical interaction data, Yu thought, was not enough. “When people talk about protein networks they have a sort of a mathematical view where proteins are dots, mathematical dots; there is no shape or structure,” he explains. “But we know that's not true; in the cell the structure of the protein is fundamentally important in determining its function.”

For those interacting human proteins for which crystal structures of the protein pair ('co-crystal' structures) have been solved, the researchers therefore added the interface structures to the network. But such information is not yet available for most putatively interacting human proteins. So Yu reached back to his bioinformatics-heavy past and used a homology modeling approach to predict the structures of interacting interfaces.

If for a particular human protein interaction there is a co-crystal structure of the orthologous yeast proteins, for example, Yu and colleagues used that information and the sequence homologies to predict the interface of the human proteins; notably, they first tested this strategy on existing co-crystal data to determine that it indeed makes reliable predictions. Combining both the experimental and predicted structural information with the binary interaction network produced the human structural interaction network (hSIN). Of its 4,222 binary interactions (between 2,816 proteins), about two-thirds of the interface structures are predicted by homology modeling. Then, with their scaffold built, Yu and colleagues combed through the Online Mendelian Inheritance in Man database and Human Gene Mutation Database and found more than 20,000 disease-associated mutations in about 600 genes that they could map to hSIN.

What did the scientists see in their bird's-eye view of disease? To mention just some of their observations, they note that missense mutations in disease-associated genes are enriched in protein interaction interfaces, they investigate the basis for pleiotropy and for locus heterogeneity effects, and they propose molecular hypotheses for the effects of disease-associated mutations. It should be noted that their observations apply to the subset of mutations that maps onto hSIN; the extent to which this analysis applies to other disease mutations remains to be seen. Finally, Yu and colleagues used hSIN to predict 292 new disease-associated genes, in 182 different diseases.

In the present work, the researchers validated experimentally only a few of their observations, using the yeast two-hybrid system, but Yu hopes to take this to an entirely different scale in the future. He envisions hSIN-derived predictions about the effects of disease mutations on protein interactions combined with rapid experimental testing of these predictions in high throughput.

There is a very long way to go before human disease can be understood mechanistically—for one thing, interaction networks such as hSIN do not take cell or tissue type into account—but at least the global effects of disease mutations on the generic protein interaction map are becoming a little clearer.