Networking proteins in yeast.

T advent of genome sequencing projects—culminating in the recent reports of the human sequence (1, 2)—has resulted in both the identification of novel genes and proteins as well as the proliferation of the ‘‘omes’’ that come from their analyses: the proteome (the complement of proteins), transcriptome (the complement of mRNA transcripts), metabolome (the complement of metabolites), and so on. These end products of global assays are needed to interpret the large fraction (typically close to half) of predicted proteins for which no proteins of similar structure exist or have been functionally characterized. The report by Ito et al. (3) is the largest contribution to date in the effort to generate the protein interactome, or map of protein–protein interactions, for the yeast Saccharomyces cerevisiae. Yeast has been the major proving ground for functional genomics methods from the time its genome was sequenced in 1996 (4). Most such approaches use the underlying principle of ‘‘guilt by association’’ as the means of elucidating function. For example, genes that are coexpressed or proteins that are found in the same complex or in the same location are likely to be involved in the same or related cellular process. Theoretical methods to deduce function include bioinformatic analyses based on protein homology, phylogenetic relationships, and protein domain fusions (5). Empirical methods elucidate gene function by diverse approaches that include expression profiling, screens for biochemical activity, identification of proteins in macromolecular complexes by mass spectrometry, systematic gene disruptions, and determinations of protein interactions. The most popular means to carry out this last method on a genomewide basis is the yeast two-hybrid system (6), a genetic assay based on the properties of site-specific transcriptional activators. Hybrid proteins are generated in yeast composed of a DNA-binding domain fused with a protein X and a transcriptional activation domain fused with a protein Y; the interaction of X and Y leads to the expression of a reporter gene whose product is easily assayed, generally by growth of the yeast on a defined media. Ito et al. (3) have followed up their earlier pilot study (7), by using a global two-hybrid approach in which they constructed a DNA-binding domain hybrid and an activation domain hybrid for each of the '6,000 predicted yeast proteins. They generated 62 pools of each type of yeast transformant, containing up to 96 independent hybrids each, followed by a systematic mating of the 62 3 62 pools to yield 3,844 sets of diploids. Subsequent recovery and sequencing of DNA from diploids positive for four different twohybrid reporters identified the genes encoding the pairs of interacting proteins. This approach resulted in 4,549 twohybrid positives among 3,278 proteins. An independent yeast two-hybrid project (8) used two other strategies: an individual DNA-binding domain hybrid tested against a library of all activation domain hybrids, and an individual DNA-binding domain hybrid tested against an array of '6,000 separate activation domain transformants. This independent study resulted in the identification of 957 putative interactions involving 1,004 proteins. Surprisingly, the overlap in the data among all three approaches is small, and neither of the two studies recapitulates more than '13% of the published interactions detected by the community of yeast biologists using conventional single protein analyses. The high fraction of false negatives may be explained by several factors such as the use of fulllength proteins vs. the protein domains used in other studies, the differing levels of hybrid protein expression, the different reporter genes, and other variables in the two-hybrid assay. This lack of overlap between datasets indicates that the screens to date are far from saturating and suggests that the yeast interactome may be larger than estimates based on earlier studies. These studies beg the question of what does it mean when a two-hybrid interaction has been detected in a genomewide approach. Ito et al. (3) focus on a core dataset of 806 interactions among 797 proteins instead of their complete dataset of 4,549 interactions among 3,278 proteins. The core dataset included cases in which the interactions were detected more than three times and excluded redundant interactions detected in both orientations of the two-hybrid assay. This focus is reasonable, given that roughly 3,000 of the interactions were identified only once or twice, and that a mere 15 proteins account for 1,504 (or '33%) of the interactions. Thus, this large-scale study likely contains a fraction of false positives, as is the case with the other recent two-hybrid efforts using proteins of S. cerevisiae (8), Helicobacter pylori (9), and Caenorhabditis elegans (10). Some of these are artifactual pairs in which a transcriptional signal occurs even though the two proteins do not interact with each other, and some are real two-hybrid interactions that do not correspond to interactions that occur in vivo. Such false positives also arise, of course, when individual researchers carry out two-hybrid searches with their favorite proteins, but they are far more likely to be discarded (or at least not reported) in the absence of any confirmatory data. Not only are these data downsizing a luxury that the genomic researcher cannot take advantage of, but individual researchers may upsize their data by additional experimentation. Thus, the hints to function in the datasets from these large-scale approaches may be best validated through conventional single protein analyses. Despite complications from redundancy and false positives, the useful information from these protein interaction projects falls into at least four categories. First, interactions of an uncharacterized protein with proteins of defined function can lead to a tentative assignment of function for the novel protein. For example, Ito et al. (3) suggest from the interaction data that the protein Ydr016c is involved in the