LETTERS Functional dissection of protein complexes involved in yeast chromosome biology using a genetic interaction map Sean R. Collins1,2,3, Kyle M. Miller4, Nancy L. Maas4, Assen Roguev1,2, Jeffrey Fillingham5, Clement S. Chu1,2,3, Maya Schuldiner1,2,3, Marinella Gebbia5, Judith Recht6, Michael Shales5, Huiming Ding5, Hong Xu5, Junhong Han7, Kristin Ingvarsdottir8, Benjamin Cheng9, Brenda Andrews5, Charles Boone5, Shelley L. Berger8, Phil Hieter9, Zhiguo Zhang7, Grant W. Brown10, C. James Ingles5, Andrew Emili5, C. David Allis6, David P. Toczyski4, Jonathan S. Weissman1,2,3, Jack F. Greenblatt5 & Nevan J. Krogan1,2 Defining the functional relationships between proteins is critical for understanding virtually all aspects of cell biology. Large-scale identification of protein complexes has provided one important step towards this goal however, even knowledge of the stoichi- ometry, affinity and lifetime of every protein���protein interaction would not reveal the functional relationships between and within such complexes. Genetic interactions can provide functional information that is largely invisible to protein���protein interaction data sets. Here we present an epistatic miniarray profile (E-MAP)1 consisting of quantitative pairwise measurements of the genetic interactions between 743 Saccharomyces cerevisiae genes involved in various aspects of chromosome biology (including DNA rep- lication/repair, chromatid segregation and transcriptional regu- lation). This E-MAP reveals that physical interactions fall into two well-represented classes distinguished by whether or not the indi- vidual proteins act coherently to carry out a common function. Thus, genetic interaction data make it possible to dissect function- ally multi-protein complexes, including Mediator, and to organize distinct protein complexes into pathways. In one pathway defined here, we show that Rtt109 is the founding member of a novel class of histone acetyltransferases responsible for Asf1-dependent acet- ylation of histone H3 on lysine 56. This modification, in turn, enables a ubiquitin ligase complex containing the cullin Rtt101 to ensure genomic integrity during DNA replication. The synthetic genetic array (SGA)2 and diploid-based synthetic lethality analysis on microarray (dSLAM)3 approaches have enabled systematic identification of synthetic sickness/lethal (SSL) relation- ships in S. cerevisiae in which pairs of gene deletions are far more deleterious together than either of the individual deletions. Although individual SSL interactions can be difficult to interpret, the patterns of genetic interactions for gene mutations can be more informative because they provide high-resolution phenotypes that can be com- pared to identify functionally related genes1���5. Recently, we exploited the SGA strategy for generating double mutants to develop an approach, termed E-MAP1, that greatly facilitates such comparisons. An E-MAP comprises comprehensive and quantitative measure- ments of genetic interactions between pairs of mutations within a defined subset of genes linked to one or more specific biological processes1. E-MAPs are created by systematically generating yeast strains carrying each pair of mutations and measuring their growth rates. Genetic interactions are determined by comparing the observed fitness of the double mutants to an empirically determined typical fitness that would be expected on the basis of the growth defects associated with each mutation1,6. This technique allows for the identification of not only negative (aggravating) interactions, such as SSL pairs, but also positive (alleviating) interactions. Positive interactions include suppression, in which double mutants are healthier than the sicker of the two single mutants, as well as cases in which loss of one gene masks the effect of losing another, as is seen when two proteins act together in a common complex or pathway. We comprehensively evaluated pairwise genetic interactions for 754 alleles of 743 genes involved in various aspects of chromosome biology (Fig. 1a see also Supplementary Fig. 1 and Supplementary Data).The mutationsinclude deletionsof 663non-essentialgenes and constitutive hypomorphic alleles���constructed using the ���decreased abundance by messenger RNA perturbation��� (DAmP) strategy1���for 70 essential genes. Genes were selected based on published functional studies, protein���protein interaction data7,8, earlier genome-wide SSL studies2 and chemical sensitivity screens. The resulting E-MAP consists of a 754 by 754 matrix of genetic interaction scores, where each row corresponds to the pattern (or profile) of interactions for one mutant allele of a gene (Fig. 1a). Using hierarchical clustering, we reordered the matrix to sort genes accord- ing to the similarity of their genetic interaction profiles. The resulting map has a modular structure that distinguishes between major pro- cesses such as transcription and chromatin remodelling, DNA rep- lication and repair, and sister chromatid segregation. We illustrate the high-resolution functional information within these modules by focusing on a subcluster containing genes involved in DNA replica- tion and repair (Fig. 1b). The general DNA replication factors (for example, RPA (RFA1 and RFA2) and RFC processivity clamp loader subunits (RFC4 and RFC5)) cluster near each other, and are resolved from the DNA replication checkpoint complex, Mrc1���Csm3���Tof1. The E-MAP also distinguishes groups of genes involved in sensing and repairing DNA damage including the RAD52 epistasis group (RAD51, RAD52, RAD54, RAD55, RAD57), the MRX complex 1 Department of Cellular and Molecular Pharmacology, 2 The California Institute for Quantitative Biomedical Research, and 3 Howard Hughes Medical Institute, University of California, San Francisco, California 94158, USA. 4 Department of Biochemistry and Biophysics, Cancer Research Institute, University of California, San Francisco, California 94115, USA. 5 Banting and Best Department of Medical Research, University of Toronto, Toronto, Ontario M5S 3E1, Canada. 6Laboratory of Chromatin Biology, The Rockefeller University, New York, New York 10021, USA. 7Department of Biochemistry and Molecular Biology, Mayo Clinic, College of Medicine, Rochester, Minnesota 55905, USA. 8Gene Expression and Regulation Program, The Wistar Institute, Philadelphia, Pennsylvania 19104, USA. 9Michael Smith Laboratories, University of British Columbia, Vancouver, British Columbia V6T 1Z4, Canada. 10Department of Biochemistry, University of Toronto, Toronto, Ontario M5S 1A8, Canada. Vol 446|12 April 2007|doi:10.1038/nature05649 806 NaturePublishing ��2007 Group

(RAD50, MRE11, XRS2) and the 9-1-1 clamp (DDC1, MEC3, RAD17). The complete genetic interaction map, a useful resource for future functional studies, is available in Supplementary Data. Beyond allowing simple hierarchical clustering, patterns of genetic interactions provide an unbiased way to identify sets of genes that function together in a coherent manner1���3,5. If two proteins act to- gether to carry out a common function, one would expect deletions of the two encoding genes to have highly similar profiles of genetic interactions, as deletion of either gene would disrupt the same cel- lular process. Similarly, one would expect a positive genetic inter- action between the two deletions, because in the context of the first, the second deletion would incur no additional cost. This relationship can be formalized in a COP (complex or pathway) score (see Sup- plementary Methods)1,6, which synthesizes both expectations to cre- ate a single mathematical metric. Sets of genes connected by high COP scores are analogous to classically defined epistasis groups such as the well-studied RAD52 epistasis group (Fig. 1b and Supplemen- tary Fig. 2). Recently, we used large-scale affinity purification data7,8 to gen- erate a physical interaction map that quantitatively reports through a purification enrichment (PE) score on the relative likelihood of each protein���protein interaction (see http://interactome-cmp.ucsf.edu)9. The accuracy and completeness of this integrated physical interaction map and the present E-MAP now make it possible to explore broadly the relationship between physical complexes and genetically defined epistasis groups. To evaluate the predictive power of the COP score relative to the physically based PE score, we used the protein com- plexes in the Saccharomyces Genome Database (SGD)10 to define a trusted reference set of ���true positives��� and ���true negatives��� (see Supplementary Methods). We then used receiver operating char- acteristic (ROC) curves11, which measure the rate at which each approach identifies true positives versus true negatives, to compare the predictive power of the two approaches. Notably, the COP score identifies a distinct and large subset of protein���protein interactions with a specificity rivalling that of affin- ity purification (Fig. 2a see also Supplementary Fig. 2). A key value of the E-MAP, therefore, is that it divides physical interactions into two classes: one group in which the proteins function coherently and a second in which their patterns of genetic interactions indicate that the proteins carry out distinct or even opposing functions. In par- ticular, for pairs of physically interacting proteins, the histogram of either the genetic interaction scores or the correlation between gen- etic interaction patterns shows a roughly bimodal character (Fig. 2b, c). Thus, for a large fraction (somewhat greater than half) of physical Correlation between interaction partners Genetic interaction score c a b 1 ��� Specificity Frequency Frequency Sensitivity Figure 2 | Relationship between genetic epistasis groups and physical complexes. a, ROC curves comparing the power of the genetic interaction patterns���using the COP score (red) (see Supplementary Methods)���and large-scale affinity purification data (blue)���using a recent re-analysisof raw purification data9���to predict co-membership of pairs of proteins in the same physical complex. The slope of the initial portion of each curve serves as a measure of the score���s maximal accuracy. b, Distribution of direct genetic interaction scores for pairs of genes encoding physically interacting proteins (green) and non-interacting proteins (black) (see Supplementary Data). c, Distribution of the Pearson���s correlation coefficients between the interaction patterns for the same sets of gene pairs as in b. Nuclear pore DNA replication checkpoint SUMO complex Sgs1 complex MRX RAD52 epistasis group DNA damage checkpoint Mms4/Mus81 endonuclease DNA replication factors Replication fork processivity Repair by homologous recombination (HR) Acetylation of K56 and suppressing DNA damage during replication Maintenance of replication fork by single strand invasion and suppressing HR S G2/M NUP60 NUP84 NUP133 HEX3 SLX8 PRI1-DAmP RFC4-DAmP DPB11-DAmP HYS2-DAmP RFA1-DAmP RTT105 RFA2-DAmP RFC5-DAmP POL30-DAmP ELG1 POL32 RAD27 POL30-79 POL30-879 CDC9* TSA1 MRC1 CSM3 TOF1 SGV1 YBR094W MMS4 MUS81 HPR5 (SRS2) ESC2 TOP3 SGS1 RMI1 RTT107 RTT101 MMS1 MMS22 RTT109 ASF1 RAD57 RAD55 RAD51 RAD52 RAD54 XRS2 RAD50 MRE11 RAD9 MEC3 RAD17 RAD24 DDC1 RAD53-11 HST3 HOP2 Histone H3 K56 pathway ���3 ���2 ���1 0 1 2 3 (alleviating) (aggravating) Transcription and chromatin remodelling DNA replication and repair Chromatid segregation b a Figure 1 | Hierarchical clustering of genetic interaction patterns. a, Full ���clustergram��� of the patterns of interactions for all 754 mutations. Black horizontal bars indicate regions of the cluster corresponding to sets of genes implicated in the indicated functional processes. b, An enlargement of the ���DNA replication and repair��� subcluster from a. The dendrograms indicate the relative similarities of the full patterns of interactions for the indicated genes. Various subclusters are annotated according to their known functions. Blue and yellow represent negative and positive genetic interactions, respectively. Grey boxes correspond to missing data points. Here, and throughout, genes indicated with an asterisk correspond to deletions of spurious ORFs that overlap the indicated gene. Genes highlighted in red represent novel findings that are referred to in the text. NATURE|Vol 446|12 April 2007 LETTERS 807 NaturePublishing ��2007 Group