Dynamic modularity in protein interaction networks predicts breast cancer outcome Ian W Taylor 1,2 , Rune Linding 1,3 , David Warde-Farley 4,5 , Yongmei Liu 1 , Catia Pesquita 6 , Daniel Faria 6 , Shelley Bull 1,7 , Tony Pawson 1,2 , Quaid Morris 4,5 & Jeffrey L Wrana 1,2 Changes in the biochemical wiring of oncogenic cells drives phenotypic transformations that directly affect disease outcome. Here we examine the dynamic structure of the human protein interaction network (interactome) to determine whether changes in the organization of the interactome can be used to predict patient outcome. An analysis of hub proteins identified inter- modular hub proteins that are co-expressed with their interacting partners in a tissue-restricted manner and intramodular hub proteins that are co-expressed with their interacting partners in all or most tissues. Substantial differences in biochemical structure were observed between the two types of hubs. Signaling domains were found more often in intermodular hub proteins, which were also more frequently associated with oncogenesis. Analysis of two breast cancer patient cohorts revealed that altered modularity of the human interactome may be useful as an indicator of breast cancer prognosis. Transcriptome analyses have been extensively applied as molecular diagnostic and prognostic tools in breast cancer. Recently, the prog- nostic predictive performance of gene expression signatures has been improved by incorporating interactome data 1 , suggesting that altered gene expression in breast cancer might disturb the higher-level organization of the interactome and affect disease outcome. To investigate this possibility, we first identified proteins that have many interacting partners (so called ‘hubs’) in a network of protein-protein interactions curated from the literature and high- throughput sources 2 (Supplementary Fig. 1a online). Next, we obtained genome-wide expression data measured in 79 human tissues 3 , and quantified the extent to which a hub and its interacting partners were co-expressed in the same tissues (Supplementary Methods online). We used the average Pearson correlation coefficient (PCC) of co-expression of a hub protein and its partners to identify whether interactions are context specific (that is, interacting proteins are not always co-expressed) or constitutive (that is, interacting proteins are always co-expressed). This revealed a multi-modal dis- tribution that appeared to be the superposition of distinct populations of hubs centered over increasing average PCC values (Fig. 1a, red asterisks). Randomly reassigning the expression data to different gene products in the same network resulted in an approximately normal distribution of PCC values (Fig. 1a, black dashed line). The shoulder (marked with a black asterisk) is largely due to strongly correlated gene products that have a high probability of reforming interactions with their true interactors when randomized (data not shown). We observed a similar multi-modal distribution using a literature-curated source alone 4 (Supplementary Fig. 1b) or a different high-confidence human PPI database 5 (Supplementary Fig. 1c). The human interactome thus has two classes of hubs. One class displays low correlation of co-expression with its partners. We call these hubs intermodular hubs, as first proposed for the yeast inter- actome 6,7 . A second class, termed intramodular hubs, displays more highly correlated patterns of co-expression (Fig. 1a). These features reflect a modular architecture. Restricting the analysis to interactions conserved between yeast and humans revealed a single peak at high average PCC, suggestive of largely intramodular hubs (Fig. 1b). Previous analyses showed that the assembly of intramodular hubs into macromolecular complexes constrains intramodular hub evolu- tion 6 . This is visualized as a cluster of highly correlated interactions interconnecting intramodular hubs in the human interactome (Sup- plementary Fig. 1a; green edges between blue nodes). Modular structure can confer higher-order function to inter- actomes, such that intermodular hubs provide temporally and spatially restricted linkages to intramodular hubs that in turn fulfill specific functions, often as multi-subunit macromolecular machines 8,9 . For example, most components of the 26S proteasome show highly correlated expression and function together to mediate protein degradation (Supplementary Fig. 2a online). However, three hub components (PSMB1, PSMB2 and PSMD9) are intermodular, reflecting tissue-specific modulation of the proteasome 10,11 . Using the Gene Ontology (GO) molecular function database 12 , we found that intramodular hubs shared more functional similarity with their partners than did intermodular hubs (Student’s t-test, P o 0.02, Supplementary Fig. 2b). Received 25 September 2008; accepted 18 December 2008; published online 1 February 2009; doi:10.1038/nbt.1522 1 Samuel Lunenfeld Research Institute, Mount Sinai Hospital, 600 University Ave., Toronto, Ontario M5G 1X5, Canada. 2 Department of Molecular Genetics, University of Toronto, 1 King’s College Circle, Room 4396, Toronto, Ontario M5S 1A8, Canada. 3 Cellular & Molecular Logic Team, Institute of Cancer Research (ICR), Section of Cell, Molecular & Systems Section, 237 Fulham Road, London, SW3 6JB, UK. 4 The Terrence Donnelly Centre for Cellular and Biomolecular Research, 160 College St., Toronto, Ontario M5S 3E1, Canada. 5 Department of Computer Science, University of Toronto, 10 King’s College Road, Room 3303, Toronto, Ontario M5S 3G4, Canada. 6 Faculty of Sciences, University of Lisbon, Campo Grande, 1749-016, Lisbon, Portugal. 7 Dalla Lana, School of Public Health, University of Toronto, 155 College St., Toronto, ON M5T 3M7, Canada. Correspondence should be addressed to J.W. (wrana@lunenfeld.ca). NATURE BIOTECHNOLOGY VOLUME 27 NUMBER 2 FEBRUARY 2009 199 LETTERS © 2009 Nature America, Inc. All rights reserved.