NATURE BIOTECHNOLOGY VOL 18 DECEMBER 2000 http://biotech.nature.com 1257Functional genomics strategies are of increasing importance in char-acterizing proteins newly discovered by genome sequencing pro-jects. In one approach, large sets of mutants created systematically1or randomly2are analyzed by physiological tests. In a secondapproach, DNA arrays are interrogated in order to cluster genes thetranscripts of which are co-regulated with genes of known func-tions3–5. A third approach uses computational strategies to assignfunctions based on the coevolution of sets of proteins6or on theexistence of a fusion protein in one organism corresponding to twoseparate proteins in another organism7,8. As a fourth approach, weuse protein–protein interaction data (for example, ref. 9) to classifyproteins based on the properties of their interacting partners.The identification of protein interactions has escalated in scalefrom the analysis of small numbers of proteins10–12to more compre-hensive analyses covering more than 1,000 proteins13. In particular,the ∼6,000 predicted proteins of the yeast S. cerevisiae have beenused extensively in two-hybrid searches to detect interacting part-ners14,15. When the data from these studies are combined with previ-ously published interactions from the entire community of yeastresearchers, they total over 2,709 putative interactions encompass-ing 2,039 different proteins16,17.Here we analyze all of these S. cerevisiae interactions in an effortto diagram the set of links within a large protein network. This net-work and the accompanying computational approaches can be usedto view interactions among proteins within any defined functionalcategory or cellular localization. The interaction data can be used topredict function for those uncharacterized yeast proteins that havepartners of known function. Such approaches will find additionalutility when applied to other organisms with increasing numbers ofinteractions, and as uncharacterized yeast proteins placed on theinteraction map are used to predict possible functions for theirorthologs in other species.Results1,548 proteins can be linked by protein interactions. We analyzed2,709 published interactions involving 2,039 yeast proteins, availablefrom public databases16,17and from two recent large-scale studies14,15(see Experimental Protocol). Our analysis includes only direct inter-actions identified by biochemical experiments or two-hybrid stud-ies, but not protein complexes for which specific protein contacts areunknown. We sought to determine whether comprehensive interac-tion maps could be assembled from these data, and if so, whethersuch maps afforded insight into functional relationships amongcharacterized and uncharacterized proteins. To visualize interac-tions, we developed a software program based on the graph-drawinglibrary “AGD” (http://www.mpi-sb.mpg.de/AGD). Surprisingly,only a single large network of protein interactions was obtained,containing 2,358 links among 1,548 individual proteins (Fig. 1A).The next largest network contained only 19 proteins; 9 networkscontained between 5 and 11 proteins; and the remaining 193 net-works contained 4 or fewer proteins (data not shown), for a total of204 independent networks (each interaction appears in only a singlenetwork).Proteins have been assigned to 42 “cellular role” categories in theYeast Protein Database (YPD)16. The term “function” is used here asin YPD to mean the cellular role a protein is engaged in and not itsprecise biochemical activity. The YPD categories are broad, and 39%of the 1,485 characterized proteins in the networks are assignedmore than one cellular role. Most proteins within one of these func-tional classes cluster in a specific region of the large network, if a“cluster” includes any three or more proteins of the same functionseparated by no more than two other proteins. For example, inFigure 1, proteins involved in chromatin structure are highlighted ingray (89% of all such annotated proteins within clusters); in red,cytokinesis proteins (75% in clusters); in blue, membrane fusionproteins (94% in clusters); in green, cell structure proteins (90% inclusters); and in yellow, lipid metabolism proteins (58% in clusters).A major concern in the delineation of these 204 networks is thequality of both the data and the database annotations. For example,interactions derived from genomic two-hybrid approaches are gen-erally uncorroborated by additional experiments, and false positivesare commonplace with this technique. Annotations of “cellular role”sometimes do not match experimentally determined properties, andother annotation information often has been transferred fromA network of protein–protein interactions in yeastBenno Schwikowski1,2*, Peter Uetz3, and Stanley Fields3,41The Institute for Systems Biology, 4225 Roosevelt Way NE, Suite 200, Seattle, WA 98105. 2Department of Computer Science and Engineering,University of Washington, Box 352350, Seattle, WA 98195. 3Departments of Genetics and Medicine, 4Howard Hughes Medical Institute, University of Washington,Box 357360, Seattle, WA 98195. *Corresponding authors ([email protected], [email protected], [email protected]).Received 27 June 2000; accepted 13 October 2000A global analysis of 2,709 published interactions between proteins of the yeast Saccharomyces cere-visiae has been performed, enabling the establishment of a single large network of 2,358 interactionsamong 1,548 proteins. Proteins of known function and cellular location tend to cluster together, with 63%of the interactions occurring between proteins with a common functional assignment and 76% occurringbetween proteins found in the same subcellular compartment. Possible functions can be assigned to aprotein based on the known functions of its interacting partners. This approach correctly predicts a func-tional category for 72% of the 1,393 characterized proteins with at least one partner of known function,and has been applied to predict functions for 364 previously uncharacterized proteins.Keywords: Protein interactions, Saccharomyces cerevisiae, functional genomics, proteomics, bioinformaticsRESEARCH ARTICLES© 2000 Nature America Inc. • http://biotech.nature.com© 2000 Nature America Inc. • http://biotech.nature.comhomologous proteins without experimental confirmation. To assessthe reliability of these networks, we determined how well they couldbe used to predict function for