The construction of protein-protein interactomes is an important step toward understanding cell developmental and regulatory programs. Protein-protein interactions (PPIs) have been examined by using high-throughput experimental techniques, such as yeast two-hybrid assays and affinity purification, and by computation-based approaches, including gene ontology and phylogenetic profiles. Now, Zhang et al. have developed an approach called PrePPI (predicting protein-protein interactions) that combines structural- and nonstructural-interaction data to predict PPIs on a genome-wide scale. PrePPI performs structural modeling on a pair of query proteins, making predictions based on close structural neighbors as well as homology models and remote structural relationships, greatly expanding the number of interactions that can be modeled. The resulting PPI model is scored and, by using a Bayesian network, a likelihood ratio is calculated that incorporates nonstructural evidence, such as coexpression levels and functional similarities. The authors used PrePPI to score all possible pairwise interactions in the yeast and human proteomes, predicting about 31,400 high-confidence interactions for yeast and nearly 318,000 for humans. Nineteen predictions were experimentally validated by using coimmunoprecipitation assays. One set of predictions involved potential PPIs formed by suppressor of cytokine signaling 3 (SOCS3). SOCS3's inhibitory function had already been established in the JAK-STAT pathway, but PrePPI predicted interactions with GRB2 and RAG3, components of the RAS-MAPK pathway, which were experimentally confirmed. PrePPI performs better than methods that rely on structural or nonstructural evidence alone. The authors attribute its relative success to the use of both close and remote structural neighbors, an efficient scoring strategy for their interaction models that allows them to examine a very large number of models, and the use of a Bayesian network to evaluate the predicted interaction models. Because PrePPI relies on structural representatives of sequence families, it falls short in mapping membrane protein interactions. It also presently does not take post-translational modifications into account. However, PrePPI offers a feasible alternative to high-throughput methodologies while also providing a rough model for what a predicted PPI may look like. (Nature doi:10.1038/nature11503, published online 30 September 2012)