Research ArticleNetwork Modeling

Integration of Protein Abundance and Structure Data Reveals Competition in the ErbB Signaling Network

See allHide authors and affiliations

Science Signaling  17 Dec 2013:
Vol. 6, Issue 306, pp. ra109
DOI: 10.1126/scisignal.2004560


The mechanisms of context-specific differences in signal transduction, such as those that occur among different cell types, are not fully understood. One possibility is that differences in the abundance of proteins change signaling outputs because these proteins compete for binding to hub proteins at critical network branch points. Focusing on the ErbB signaling, we created a protein interaction network that included information about protein domains and analyzed the role of competing protein interactions. By leveraging three-dimensional protein structures to infer steric interactions among binding partners for a common binding domain or linear motif (node) and including information about protein abundance and interaction affinities, we identified a large number of competitive, mutually exclusive (XOR) protein interactions. Modeling changes in protein abundance with different patterns of partner proteins and XOR nodes (XOR motifs) revealed that each motif conferred a different response. We experimentally investigated the XOR motif containing the hub protein Ras and its binding partners RIN1 (Ras and Rab interactor 1) and CRAF (v-raf–leukemia viral oncogene 1). Consistent with the computational prediction, overexpression of RIN1 in cultured cells decreased the phosphorylation of CRAF and its downstream targets. Thus, our analyses provide evidence that variation in the abundance of proteins that compete for binding to XOR nodes could contribute to context-specific signaling plasticity.


A key question in signal transduction biology is how a limited number of proteins act together to coordinate a multitude of biological processes and cell type–specific responses. For example, treating PC12 pheochromocytoma cells with epidermal growth factor (EGF) or nerve growth factor results in either transient or sustained ERK (extracellular signal–regulated kinase) activation and distinct cell fates (1). Similarly, treating MCF-7 breast cancer cells with EGF or heregulin (HRG) induces either transient or sustained ERK activation (2). In addition, there are examples where the same growth factor can elicit different downstream signaling events depending on the cell type (35). Several mechanisms have been invoked to explain these differences (5) including clustered receptors in microdomains (6, 7), binding to scaffolding proteins (8), localized signaling by proteins in different cell compartments (9), changes in local protein concentration (10), cooperative binding on multidomain proteins (11), alternative splicing (1214), and the interplay among feedback loops (15).

In signaling networks, central proteins act as hubs by binding to multiple partner proteins. In some cases, partner proteins do not bind simultaneously to the hub because of steric hindrance at a common binding domain or linear sequence motif, also known as a mutually exclusive binding site (16, 17), referred to here as an XOR node. Recent works demonstrate that competitive binding at these sites is important for cellular functions. For example, competition among G protein [heterotrimeric guanosine triphosphate (GTP)–binding protein]–coupled receptor kinases (GRK1, GRK5, and GRK6) to phosphorylate G protein–coupled receptors affects the amount of G protein–mediated (transient) or β-arrestin–mediated (sustained) activation of ERK (18). Likewise, a structural analysis of the human-virus protein interaction network revealed that viral protein domains mimic those of endogenous human proteins and compete for binding to hub proteins and thus inhibit protein complex formation and function (19).

We reasoned that if a hub protein at a critical network branch point is present at a limiting concentration, then the formation of a specific protein complex is determined by the concentrations of its binding partners, and that changes in relative concentrations of the partners could result in the flow of information through distinct signaling pathways. To test this hypothesis, we focused on the ErbB signaling network. The EGF receptor (EGFR) is the founding member of the ErbB receptor tyrosine kinase family, which also includes ErbB2, ErbB3, and ErbB4, and regulates various cellular processes including cell growth, metabolism, survival, apoptosis, proliferation, and differentiation (2, 2022). We predicted putative competitive interactions using three-dimensional structural modeling to identify XOR nodes on hub proteins where steric effects could prevent simultaneous binding to multiple partner proteins (16) and where abundance of the hub protein was limiting relative to that of the binding partners. We then used this information to annotate a high-confidence protein interaction network for ErbB signaling. Computational modeling of theoretical XOR nodes with partner proteins arranged in different patterns (XOR motifs) showed that changes in the abundance of partner proteins produced distinct signaling responses from each motif. Modeling of the Ras [a guanosine triphosphatase (GTPase)] XOR motif led to the prediction that RIN1 (Ras and Rab interactor 1) competed with CRAF (v-raf–leukemia viral oncogene 1) for binding to Ras. Experimentally, we found that overexpression of RIN1 resulted in the inhibition of growth factor–induced Ras-mediated phosphorylation of CRAF and its downstream targets MEK [mitogen-activated protein kinase (MAPK) kinase] and ERK. Thus, changes in protein abundance at XOR nodes produce an array of outputs for remodeling downstream signaling by forming alternate protein complexes.


Construction and structural annotation of a high-confidence binary ErbB signaling network

To determine the set of ErbB signaling-related proteins, we extended the most recent comprehensive EGFR network (20). This network contains 77 proteins to which we added 122 proteins from the NetPath (23), Reactome (24), and manual curation of the literature (Fig. 1, fig. S1, and table S1). Confidence scores were assigned to protein-protein interactions (PPIs) using the STRING database (25). Among 33 high-confidence ErbB signaling PPIs from the Reactome data set, the lowest STRING score was 0.96 (table S1). PPIs among the list of 198 proteins that fell below this threshold were excluded from further analysis, resulting in a network of 162 proteins connected by 590 PPIs. Not all PPIs are direct (binary) interactions; some PPIs are discovered using indirect methods (for example, coimmunoprecipitation) (17). Thirty-seven proteins, previously described as being involved in ErbB signaling (20), were not connected to the network with a STRING score equal to or greater than 0.96 and were among those excluded (table S1). Annotation with Gene Ontology (GO) terms demonstrated that proteins in this network functioned in a variety of cellular and receptor-specific functions connected to ErbB signaling (26, 27) (table S1).

Fig. 1 Construction of a high-confidence, structurally annotated model of the ErbB protein interaction network.

(Top) Venn diagram depicting 198 proteins from Human Protein Reference Database (HPRD) (177 proteins), the Kitano laboratory (20) (77 proteins), Reactome (35 proteins), and those manually curated from the literature (7 proteins). These proteins were used as input to generate a list of 590 PPIs using the STRING database with stringent criteria. (Middle) The binary PPI network was annotated by manually curating experimental DDI and DPI information and structural data (177 interactions) or by manually curating phosphorylation and dephosphorylation reactions (122 interactions). (Bottom) The resultant 162 proteins and 269 annotated interactions were assembled using Cytoscape. The color of nodes corresponds to their functional annotation: yellow, receptor tyrosine kinases; blue, adaptors; light blue, adaptors with enzymatic function; green, messenger-modifying enzymes; red, Ser-Thr kinases; orange, phosphatases; magenta, small GTPases; pink, transcription factors; white, other. The binary ErbB network is also available as Cytoscape file (data file S1).

To further annotate the network and to distinguish stable physical interactions from transient kinase-substrate and phosphatase-substrate interactions, we deconstructed the 162 proteins into the protein domains that govern PPIs using the Pfam protein family database (28) (table S2). We parsed the 590 PPIs into those mediated by domain-domain interactions (DDIs) or by domain–linear sequence motif interactions (DPIs) using manual annotation based on the literature. We annotated PPIs with binding affinities, kinetic constants, phosphorylation sites, and the predicted PPI directionality using a variety of resources (table S2). We annotated experimental evidence for physical interactions (DDI or DPI) for 179 of 590 PPIs and for phosphorylation or dephosphorylation interactions for 122 of 590 PPIs (Fig. 1 and data file S1). For 90 PPIs, there was no evidence of DDIs or DPIs, and only transient kinase-substrate or phosphatase-substrate interactions were annotated. The remaining 321 PPIs were removed from the network, resulting in 34 unconnected proteins, which were also removed. The final binary network contained 128 proteins, 179 physical and 90 phosphorylation or dephosphorylation PPIs, predicted directionality for 71% of PPIs, and quantitative affinities or kinetic constants for 30% of PPIs (Fig. 1 and table S2), representing an additional 51 proteins and 115 PPIs as compared to the published network (data file S1) (20).

Distinguishing between mutually exclusive and compatible interactions

Proteins that bind to the same or neighboring domains on a hub protein could disrupt each other’s binding through steric hindrance. To determine whether binding partners for hub proteins in the ErbB network showed steric interference, we determined structural information for 175 of 346 domains in 128 proteins using the Protein Data Bank (PDB) (29). One hundred sixty-four domains were modeled using nonhuman orthologs or by using domains of the same family, for example, substituting an SH3 (SRC homology 3) domain for an SH2 domain. We assessed the structure of protein complexes between hubs and partner proteins. For 26% of physical PPIs, a structure of the protein complex was available in the database. For the remaining PPIs, we inferred the structure of the complex on the basis of sequence homology to a known structure [as in (30)]. In contrast to less than 30% of proteins in the entire human proteome (31), we were able to annotate structural information for 66% of multidomain proteins in the ErbB network (fig. S2, A and B). The majority of proteins (76%), including adaptor and membrane-associated proteins, interacted by using more than one domain or linear motif to bind multiple other proteins (table S2). For example, SOS1 (Son of sevenless homolog 1) bound to the SH3 domain of GRB2 (growth factor receptor–bound protein 2) with its proline-rich region, to GTP-bound Ras with a distal binding site in its guanine nucleotide exchange factor domain (32), and to membrane phospholipids with its pleckstrin homology domain (table S2).

To identify XOR nodes, we analyzed the structure of the domains of hub proteins and their binding partners (figs. S2C and S3). If two different domains of the hub protein bound to two partner proteins, we assumed that binding was compatible (AND). However, if the hub protein used the same domain to bind to different proteins, then we performed superimposition of the structures of the binding domains to determine whether there was overlap between the resulting protein complexes (17), as shown for Ras (Fig. 2). We defined nodes on hub proteins where binding partners overlapped as mutually exclusive (XOR) and those where binding partners did not overlap as compatible (AND). Analyses of 92 hub proteins revealed that 74% of nodes mediated by physical PPIs were XOR and 89% of nodes mediated by phosphorylation or dephosphorylation PPIs were XOR (Fig. 3, A to C, and data file S2), suggesting that these proteins could compete for common binding sites and that this competition could be important for ErbB signaling.

Fig. 2 Defining XOR nodes with structural superimposition and domain annotation.

(A) Superimposition of Ras in a complex with the Rho-binding domain (RBD), the Ras GTPase–activating domain (GAP), and the Ras guanine nucleotide effector domain (GEF). All domains used a similar surface to bind to Ras, and therefore, the PPIs were mutually exclusive (XOR). (B) Ras depicted as an example for a hub protein with an XOR node. Gray circles represent the partner proteins, and small white dots represent the domain that mediates binding to Ras. The thickness of the lines indicates the binding affinities. The dotted line indicates that no binding affinity for the interaction of the Ras with the RasGEF domain was available.

Fig. 3 Diagram of the ErbB network with annotation of DDI- and DPI-type PPIs showing XOR and AND nodes.

(A) Gray circles represent proteins, and small circles represent domains or linear motifs that mediate PPIs. Red indicates XOR nodes, and blue indicates AND nodes. See data file S2 for a high-resolution diagram. (B and C) Proportion of PPIs for XOR and AND nodes for physical interactions (BIND, B) including DDIs and DPIs or phosphatase-substrate and kinase-substrate interactions (PD, C).

Protein concentration and competition at XOR nodes governing physical PPIs

Our finding that physical PPIs in the ErbB network involved a large number of XOR nodes implies that differences in protein concentrations among binding partner proteins could result in competitive interactions. We used published mass spectrometry–based measurements of protein abundance in NIH3T3, U2OS, and HeLa cells (3335) to analyze 53 XOR node containing hub proteins and their binding partners. To reduce the effect of artifacts due to technical and biological variability of measured protein abundance in these studies, we performed Monte Carlo simulations, varying protein concentration by two- or fivefold (fig. S4). If in 95% of the simulations the sum of the abundance of partner proteins exceeded that of hub protein, we defined this as a competitive XOR node, and we found that more than 64% of XOR nodes on hub proteins were competitive (Fig. 4 and table S3). More than 60% of XOR nodes were competitive in all three cell lines, and 26% of XOR nodes were not competitive in any cell line (table S3). This result suggests that competition at XOR nodes is conserved across cell types and could modulate output responses through the formation of different protein complexes.

Fig. 4 Analysis of competition at XOR nodes from abundance determined with Monte Carlo simulations.

The left graph indicates the total number of XOR nodes analyzed. The right graph shows the number of nodes for which the abundance of the interaction partners was greater than that of the XOR hub protein as determined in at least 95% of Monte Carlo simulations assuming either a two- or fivefold cumulative error relative to absolute protein abundance published for NIH3T3, U2OS, and HeLa cells (3335).

Experimental evaluation of competitive binding at XOR nodes

Our data suggested that variation in the concentration of proteins that bind to XOR nodes could affect protein complex formation and functional output at multiple places in the ErbB network. We modeled different motifs of XOR hub proteins and binding partners and discovered that theoretical variations in binding partner protein concentration could affect the composition of protein complexes in different ways (fig. S4). For example, in an XOR motif with two binding partners of equal affinity for the XOR node, increasing the concentration of one binding partner decreased the formation of a complex between the XOR hub and the other binding partner, suggesting alterations in the functional output (fig. S4A). One such XOR motif in the ErbB network involves Ras, which has only one protein interaction domain and several competitive PPIs with different partner proteins (Figs. 2B and 3A and table S3). Analysis of gene expression data using BioGPS (36) suggested that tissue-specific differences in the expression of Ras and Ras-binding proteins could explain differences in functional outputs (fig. S5 and table S7). Ras and Ras-binding proteins RIN1, RalGDS (Ral guanine nucleotide dissociation stimulator), and CRAF are expressed in the brain (36). Activation of EGFR by ligands such as EGF or HRG promotes Ras-mediated phosphorylation of CRAF (37), and CRAF and RIN1 bind Ras with a similar affinity (38, 39). In forebrain neurons, binding of RIN1 to Ras has been proposed as a mechanism by which RIN1 can inhibit synaptic remodeling and learning and memory, which requires Ras-mediated phosphorylation of CRAF (40).

To address whether increased abundance of RIN1 could decrease Ras-mediated phosphorylation of CRAF and downstream signaling, we modeled the RIN1-CRAF-Ras XOR motif (Fig. 5A, tables S5 and S6, and data file S3). We used the in vitro Ras-binding affinities of RIN1 and CRAF measured with identical methods (38). In addition, we empirically determined the absolute abundance of Ras (sum of H-, K-, and N-Ras), CRAF, MEK (sum of MEK1 and MEK2), GRB2, and ERK (sum of ERK1 and ERK2) in human embryonic kidney (HEK) 293 and MCF-7 cells using quantitative Western blot (figs. S8 to S10 and table S4) and included this information in the model. Model simulation indicated that a 10-fold increase in the abundance of RIN1 relative to CRAF was sufficient to inhibit Ras-mediated phosphorylation of CRAF (Fig. 5B).

Fig. 5 Computational and experimental analysis of the RIN1-Ras-CRAF XOR motif.

(A) Schematic diagram of the network used for simulation of the competition between RIN1 and CRAF for binding to Ras. (B) Simulated abundance of phosphorylated CRAF using different concentrations of RIN1. (C) Western blot for the indicated proteins in MCF-7 and HEK293 cells untransfected (U) or transfected with RIN1 and stimulated with HRG or EGF for 5 min. β-Actin was used as a loading control (LC). Bar graphs show the results from three biological replicates for the abundance of the indicated proteins normalized to the loading control. Blue is untransfected and red is RIN1 transfected. Data are means ± SD. *P < 0.05, **P < 0.005, t test.

Thus, we tested the ability of overexpressed RIN1 to decrease Ras-mediated phosphorylation of CRAF in cell culture. We transiently overexpressed RIN1 in HEK293 and MCF-7 cells, resulting in protein abundances ranging from 8.43 × 105 to 2.08 × 106 molecules per cell. This corresponded to an abundance ratio between 6.6- and 30-fold relative to CRAF (fig. S9 and table S4). We serum-starved the cells, stimulated them with EGF or HRG for 5 min, and measured the phosphorylation state of CRAF and its downstream targets by Western blot (Fig. 5C). We found that overexpressing RIN1 in cells significantly reduced the phosphorylation of CRAF, MEK, ERK, and ribosomal protein S6 kinase (RSK) relative to untransfected control cells (Fig. 5C and fig. S7) but did not affect the total abundance of these proteins (fig. S6).

Analysis of theoretical XOR network motifs also suggested that membrane localization of a binding partner for a membrane-localized hub protein should decrease complex formation with the other binding partner (fig. S4D). Because Ras is membrane-bound (41), we predicted that membrane localization of RIN1 would result in a local increase in abundance and reduce CRAF phosphorylation. Modeling suggested that a fivefold increase in the local membrane concentration of RIN1 was sufficient to decrease CRAF phosphorylation (Fig. 6, A and B). We fused RIN1 to a C-terminal prenylation signal (CAAX box) (42) to promote membrane localization and to yield a predicted local increase in abundance of about fivefold (43). Similar to wild-type RIN1, overexpression of RIN1-CAAX in HEK293 or MCF-7 cells reduced the ability of EGF or HRG to stimulate phosphorylation of CRAF, MEK, and ERK (Fig. 6C) but did not affect the total abundance of these proteins (fig. S6). Thus, our results suggest that competition at XOR nodes and variation in protein abundance or localization can modulate protein complex formation and downstream responses.

Fig. 6 Computational and experimental analysis of the membrane-associated RIN1-Ras-CRAF XOR motif.

(A) Schematic diagram of the network used for simulation of the competition between membrane-localized RIN1 (yellow rectangle, RIN1-CAAX) and CRAF for binding to Ras. (B) Simulated abundance of phosphorylated CRAF using different concentrations of RIN1-CAAX. (C) Western blot for the indicated proteins in MCF-7 and HEK293 cells untransfected (U) or transfected with RIN1-CAAX and stimulated with HRG or EGF for 5 min. β-Actin was used as a loading control (LC). Bar graphs show the results from three biological replicates for the abundance of the indicated proteins normalized to the loading control. Blue is untransfected, red is RIN1 transfected, and green is RIN1-CAAX transfected. Data are means ± SD. *P < 0.05, **P < 0.005, t test.


Using database mining, literature search, and manual curation, we assembled an ErbB signaling network including the domains and linear motifs that mediate binary PPIs. A large proportion (66%) of protein domains of all proteins in the core of the network were involved in physical PPIs, suggesting that our network is nearly complete. Other large-scale structural interaction studies, such as the MAPK PPI network developed with the Coev2Net framework (44), have constructed similar three-dimensional PPI networks. Our data augment these efforts; however, the protein overlap between these studies and our ErbB network was too small to perform direct comparison.

Structural information about protein domains was used to distinguish AND and XOR PPIs among hub proteins and their binding partners. The majority of interactions in the ErbB network were XOR, suggesting that competitive binding plays an important role in ErbB signaling. A structural analysis of the rhodopsin protein interaction network by our group revealed that competitive binding is frequently associated with cellular processes where proteins dynamically connect to different signaling modules (17). In the ErbB network, several proteins were involved in multiple signaling complexes. For example, the SH3 domain of GRB2 binds to SOS1 to activate Ras and MAPK (27), binds to GAB1 and GAB2 to activate phosphoinositide 3-kinase (27), or binds to dynamin-1 to activate endocytic pathways (45). We found that GRB2 was more abundant than the sum of its binding partners in the ErbB network, suggesting that GRB2 was a noncompetitive XOR node. However, competition at the GRB2 XOR node could occur if only a fraction of the total pool of GRB2 was available for ErbB signaling. In addition, theoretical analysis of XOR motifs indicated that in addition to concentration, competition could arise from the relative affinity of XOR node binding proteins. Thus, multiple factors in addition to those considered in our network could influence the ability of hub proteins such as GRB2 to interact dynamically with multiple proteins.

A limitation of our model is that signaling networks do not operate in isolation. As mentioned, hub proteins such as GRB2 also bind proteins in other signaling pathways and those with general cellular functions. The ability of hub proteins and binding partners to participate in various interactions and complexes depends on both the connectivity (compatible and mutually exclusive) and the abundance of all of the proteins in a given cell type or tissue. Mathematical modeling of multidomain PPI networks is limited by combinatorial complexity (46), meaning that the number of differential equations that need to be solved becomes too high for computational implementation [more than 1 × 106 equations even for simple signaling networks (2)]. Recently, rule-based modeling (RBM) (47) was applied to a small, but complete, yeast PPI network (48). RBM uses a set of rules for binding between domains and binding sites and facilitates construction and simulation of mathematical models (47). RBM or similar approaches could ultimately enable simulation of proteome-wide arrays of AND and XOR nodes in human whole-cell PPI networks.

Competitive binding at XOR nodes happens if the hub protein is less abundant than its partner proteins. To determine which XOR nodes in the ErbB network had the potential for competition, we annotated protein abundances using recent proteomic data from mammalian cell lines (3335). Advances in mass spectrometry have overcome technical issues relating to the vast complexity and large dynamic range of the proteome and have enabled robust quantification and deep sequence coverage yielding information for about 10,000 proteins each in NIH3T3 (33), U2OS (34), and HeLa (35) cells. We found that the sum of the concentrations of the binding partners was significantly higher than that of the hub proteins for the majority of XOR nodes and that the status of an XOR node as competitive or noncompetitive was conserved in all three cell types. This could represent a fundamental principle of cellular design. Kim et al. (16) analyzed the yeast structural interaction network and found that hub proteins with a single domain that acted as an XOR node evolved faster than multidomain proteins with AND nodes. Our study further subdivided XOR nodes into competing and noncompeting XOR, and mathematical modeling demonstrated that changes in protein abundance produced different signaling outputs depending on the XOR network motif. On the basis of these observations, we predict that evolutionary gene duplication of a competitive XOR node binding protein would result in a gain of interaction (function) for the duplicate protein and a loss of function for the original protein, whereas at a noncompetitive XOR node, the duplicate protein could develop a new function without perturbing the function of the original protein. Thus, competitive XOR node binding proteins should evolve slower than those that bind noncompetitive XOR nodes, although this remains to be tested. In addition, similar changes in protein isoforms, for example, those due to mRNA splicing, or relative differences in the expression of paralogous genes could give rise to cell type– or other context-specific differences in signal transduction.

Several examples show that variation in the abundance of proteins contributes to context-specific signaling that may be attributable to competitive XOR motifs. For example, in synthetic MAPK cascades in yeast, changing the concentration of kinases or scaffolding proteins enables flexibility in the pattern of activation (from a binary switch to a gradient response) (49). Moreover, small Wnt-induced cyclic increases in β-catenin abundance enhance the efficiency of somatic cell reprogramming mediated by cell fusion (50). In Chinese hamster ovary cells with constitutive expression of the insulin receptor, two proteins—IRS1 (insulin receptor substrate 1) and Shc (SH2 domain–containing transforming protein C)—compete for binding to GRB2, and overexpression of IRS1 inhibits the ability of Shc to promote insulin-dependent activation of MAPK (51). Finally, intercellular variation in the abundance or activation state of proteins involved in receptor-mediated apoptosis correlates with the induction and timing of TRAIL-induced apoptosis (52). How and whether these proteins function as XOR hubs and whether competitive binding plays a role in accomplishing these diverse responses remain to be explored.

Our model and experimental data suggest that competitive binding at the Ras XOR node may govern diverse cellular functions. We found that overexpressing RIN1 in cells decreased the phosphorylation of CRAF and inhibited downstream signaling, presumably through competition for binding to Ras. Using a bioinformatics approach, we established that the mRNA expression of Ras and Ras-binding proteins do not covary across tissues. We found that RIN1, RalGDS, and CRAF were highly expressed in the brain, where RIN1 may modulate Ras signaling to control learning and memory (40). RIN1 was also highly expressed in the liver. Low RIN1 expression is associated with poor prognosis in liver cancer (53), which could be due to increased signaling through CRAF. RalGDS was highly expressed in the early erythroid lineage, and Ral (a GTPase downstream of RalGDS) is important for regulating the balance of differentiation and self-renewal in hematopoietic stem cells (54). RASSF1, an effector of Ras that promotes apoptosis, was highly expressed in cells of hematopoietic origin. Apoptosis plays a role in pruning superfluous hematopoietic stem cells in development (55), suggesting that the interplay between RalGDS and RASSF1 at the Ras XOR node could orchestrate this function. In summary, our work suggests that protein concentration and competition for binding play crucial roles in achieving context-dependent signaling.


Network construction, domain analyses, and structural modeling

One hundred ninety-eight proteins in the initial ErbB network represent the union of 177 proteins from HPRD ( and NetPath (, 77 proteins from the most recent comprehensive EGFR signaling network (20), 35 proteins from Reactome (, and 7 proteins from manual curation of latest literature. Proteins were assigned function using information from a recent review (26) and UniProt ( STRING ( (STRING version 8.1, November 2009) was used to determine confidence scores for PPIs among these 198 proteins. Predicted PPI directionalities were extracted from a recent publication (56). Protein sequences were retrieved from UniProt. Domains were predicted using Pfam ( Experimentally determined phosphorylation sites were obtained from Human Proteinpedia ( and Phospho.ELM ( PDB database ( was used to extract protein structures, and Pfam ( to identify domains. A large domain-peptide interaction screen (57) and manual annotation based on the literature were used to identify DDIs and DPIs. NetPath was used to annotate phosphorylation and dephosphorylation sites and their associated kinases and phosphatases. Structural information for domains was annotated using the protein itself (47%), a homologous domain in a similar protein family (30.5%), or a domain in the same Pfam family (18.6%), and 3.9% of proteins had no available structure. The PDB database ( was searched for protein complexes of known structure whose elements shared at least 70% homology with the query proteins. Structural superimpositions of the domains mediating PPIs were performed using the Swiss-PDB Viewer ( (17). For linear motifs, we assumed that 10 residues were needed to accommodate an SH2 domain, as estimated from the structure of an SH2 domain in complex with a P-Tyr peptide; for example, see PDB entry 1ZFP. If a protein was phosphorylated or dephosphorylated by different partner proteins on the same domain, we assumed the PPI was XOR; otherwise, we assumed that the PPIs were AND. Full details of protein and PPI annotation are available in table S2.

Cell culture and quantitative Western blot analysis

Plasmids were generated using Gateway cloning (Invitrogen) and verified by full-length sequencing. Protein expression was performed in Escherichia coli Rosetta strain in a 24-well plate using 2 ml of autoinduction medium (Novagen). Cells were incubated for 4 hours at 37°C, and then 20 hours at 25°C at 800 rpm. Proteins were extracted with 1 ml of lysis buffer LEW (Protino Kit) and lysozyme (1 mg/ml). After 1-hour incubation at room temperature, the cell lysate was centrifuged for 15 min at 4000 rpm. The supernatant was kept (soluble proteins), and the pellet was resuspended in 1 ml of lysis buffer in denaturing buffer [1× phosphate-buffered saline (PBS), 6 M urea, 300 mM NaCl, 0.5% Triton X-100]. After 1-hour incubation at room temperature, the pellet fraction was centrifuged at 13,000 rpm for 30 min. Supernatant and pellet fractions were probed for the target protein by Western blotting. For all proteins except Ras, the pellet fraction was used for subsequent experiments. Protino Mulit-96 Ni-IDA (Macherey-Nagel) columns were equilibrated with 500 μl of loading buffer [100 mM tris (pH 7.4), 500 mM NaCl, 6 M urea, 5 mM imidazole]. Supernatants were loaded onto the columns, washed twice with 500 μl of loading buffer, and eluted twice with 250 μl of elution buffer under denaturing conditions [100 mM tris (pH 7.4), 6 M urea, 1.5 M imidazole]. When necessary, protein expression and purification were scaled to 50 ml of autoinduction medium under the same expression and purification conditions by adjusting the relative volumes of the solutions. Protein expression was detected by SDS–polyacrylamide gel electrophoresis and Western blotting with the following antibodies: His (Sigma, H1029), Ras (H-, K-, and N-Ras; Abcam, ab52939), RAF1 also known as CRAF (Sigma, R2404), GRB2 (GeneTex, GTX61160), MEK2 (data file S4), ERK1 (Sigma, M5670), and ERK2 (Sigma, M5670). Total protein concentration was determined by measuring the absorbance at 280 nm (A280) using a NanoDrop spectrophotometer (Thermo Scientific). The proportion of protein of interest to total protein was calculated using ImageJ (NIH). A280 values were adjusted for purity, and the concentration of the protein of interest was calculated using the protein-specific extinction coefficient. Ras was purified as described previously (39).

HEK293 and MCF-7 were cultured in Dulbecco’s modified Eagle’s medium (Gibco) supplemented with l-glutamine and 10% (v/v) heat-inactivated fetal calf serum in 10-cm dishes to 80% confluence, washed twice with PBS, and resuspended in 800 μl of lysis buffer [0.1% SDS, 25 mM tris (pH 7.8), 1:1000 protease inhibitor cocktail 1 and 2 (Sigma)]. A parallel dish of cells was trypsinized, and cells were counted using a Neubauer chamber. Cell lysates were loaded on the same gel with a dilution series of purified proteins and processed for Western blot. Blots were incubated with an enhanced chemiluminescence reagent (SuperSignal West Femto, Thermo 34096) and visualized with a LAS-3000 imager (Fujifilm Co.). The intensity of protein bands was quantified with ImageJ in comparison to that of the purified proteins (figs. S8 and S9). Protein concentration was then performed by calculating the total amount protein relative to the number and volume (see below) of cells.

Known amounts of nonhuman proteins [avidin (chicken), protein A (Staphylococcus aureus), ovalbumin (chicken), myoglobin (equine), and lactoglobulin β (bovine) in high purity (>98%); Sigma] were added to MCF-7 cell lysates and quantified to assess accuracy and precision across biological replicates using Western blot (fig. S10) with the following antibodies: avidin (Sigma, A9390), protein A (Sigma, P 6031), ovalbumin (Sigma, A7641), myoglobin (Sigma, M0630), and lactoglobulin (Abcam, ab112893).

Volumetric analysis of HEK293 and MCF-7 cells

To estimate cell volumes, HEK293 and MCF-7 cells were seeded on 16-chamber Lab-Tek slides (Nunc Brand) and grown for 24 hours. Cells were fixed in 4% paraformaldehyde for 20 min, blocked with 4% bovine serum albumin in PBS, permeabilized with 0.1% Triton X-100 in PBS, stained with CellMask Deep Red (0.1 μg/ml; Invitrogen) and 4′,6-diamidino-2-phenylindole (1 μg/ml; Sigma-Aldrich), and mounted in ProLong Gold antifade solution (Invitrogen). Cells were imaged at ×63 magnification and 3× zoom on a TCS SP2 confocal microscope (Leica). High-resolution z-stacks were generated for at least five cells per cell line. The images were deconvoluted (Huygens Essential, SBI), and the surface area and volume of the whole cell and nucleus were measured from reconstructions using Imaris (version 6.2, Bitplane).

Estimation of protein abundance at XOR nodes using Monte Carlo simulations

Protein concentrations were extracted from NIH3T3, HeLa, and U2OS cells (3335) and used as a basis for analysis of XOR nodes. If a protein was not detected in one of the data sets, the assay detection limit provided by the authors was substituted as that protein’s abundance. Cumulative error rates of either two- or fivefold were used to account for biological and technical variability in these measured data. We used the log-normal distribution of abundance for each protein and estimated parameters within a 99.99% confidence interval. We calculated the relative abundance of hub proteins and binding partners using 100,000 Monte Carlo simulations for each error rate (Monte Carlo algorithm implemented in Python programming language). If the sum of the concentration of its partner proteins was larger than that of XOR hub in 95% of the simulations, we considered that there is competition among partners.

Model of RIN1-Ras-CRAF XOR motifs

A simplified mathematical model involving the Ras node was constructed on the basis of mass action kinetics (tables S5 and S6 and data file S3) using the iNA simulation software (58). This model does not include receptor activation, dimerization, or adaptor binding. The GRB2-SOS1 complex (Grb-Sos) moved between the cytosol and membrane at a constant rate (Fig. 6A, reaction 2). Guanosine diphosphate (GDP)–bound Ras (RasD) moved between the cytosol and membrane (R1) with higher rate of association at membrane. At the membrane, Grb-Sos catalyzed the exchange of GDP for GTP on Ras and resulted in active Ras (RasT) (Fig. 6A, reactions 3 and 4). RasT was restricted to the membrane. A positive feedback from RasT to Grb-Sos through a distal binding site in RasT (32) was included (Fig. 6A, reactions 5 to 7). CRAF binding to RasT, CRAF phosphorylation, and subsequent phosphorylation of MEK and ERK were included (Fig. 6A, reactions 8 to 13). A negative feedback from active ERK to Grb-Sos was included, which reduced the membrane-available Grb-Sos fraction (Fig. 6A, reactions 14 to 19). RasGAPs or phosphatases were not included, and instead a slow degradation rate to “deactivate” RasT, CRAF, MEK, and ERK was used to simplify model complexity (Fig. 6A, reactions 20 to 23). Binding of RIN1 to RasT was included (Fig. 6A, reaction 24). Rate constants were used as described (3, 59, 60) except for Ras-CRAF and Ras-RIN1 complexes, which were determined experimentally (38). Protein abundances included in the model were determined in this study using HEK293 and MCF-7 cells and quantitative Western blotting (average values from table S4). See table S6 for the starting protein abundances used to construct the model. The reaction volume was 1 × 10−14 liters.

RIN1 and RIN1-CAAX cloning and transfection

RIN1 complementary DNA was cloned into pT-REX-DEST with an N-terminal His tag using Gateway (Invitrogen) and fully sequenced. For the RIN1-CAAX plasmid, RIN1 was appended on the C terminus with 17 residues of the C terminus of K-Ras-2B using one-step isothermal assembly cloning (61).

HEK293 and MCF-7 cells were seeded on 35-mm dishes and transfected with 3 μg of plasmid, using Lipofectamine 2000 (Invitrogen) according to the manufacturer’s instructions. One day after transfection, cells were stimulated with 50 ng in 3 ml of EGF or HRG for the indicated times, washed twice with PBS, and lysed as above. The abundances of RIN1 and RIN1-CAAX were quantified by Western blot analysis with purified RIN1 protein as a standard (fig. S6). Abundance varied between 8.43 × 105 and 2.08 × 106 molecules per cell in different biological replicates. Blots were probed with the following antibodies: phospho-CRAF Ser338 (Cell Signaling, #9427), phospho-MEK Ser217 and Ser221 (Cell Signaling, #9121), phospho-ERK Thr202 and Tyr204 (Sigma, E7028), phospho-RSK Ser381 (Sigma, R6525), RIN1 (Abcam, ab69897), ERK (Sigma, M5670), CRAF (Sigma, R2404), MEK1 (Sigma, WH0005604M1), MEK2 (data file S4), and β-actin (Thermo, MA5-15739). Blots were developed, imaged, and quantified as described above.


Fig. S1. Overlap of PPIs from different data sets.

Fig. S2. Domains mediating PPIs in the ErbB network and the Ras XOR motif.

Fig. S3. Representation of PPIs with single and multi-interface XOR hub proteins.

Fig. S4. Modeling of variations in protein abundance at XOR motifs.

Fig. S5. Tissue-specific mRNA expression of proteins in the Ras XOR motif.

Fig. S6. Abundance of RIN1, CRAF, and downstream targets in MCF-7 and HEK293 cells overexpressing RIN1 or RIN1-CAAX.

Fig. S7. Phosphorylation of CRAF, MEK, and ERK in MCF-7 and HEK293 cells.

Fig. S8. Recombinant proteins used for quantitative Western blot.

Fig. S9. Analysis of protein abundance in MCF-7 and HEK293 cells by Western blot.

Fig. S10. Analysis of quantitative Western blotting technique.

Table S1. Annotation of the initial ErbB network with PPI and GO terms.

Table S2. Annotation of the ErbB network with binding affinities, structural information, domain predictions, and phosphorylation sites.

Table S3. Analysis of competition at XOR nodes using Monte Carlo simulations.

Table S4. Protein abundances in MCF-7 and HEK293 cells determined using quantitative Western blot.

Table S5. Reactions and rate constants used to model the RIN1-Ras-CRAF XOR motif.

Table S6. Initial protein abundances used to model the RIN1-Ras-CRAF XOR motif.

Table S7. List of tissue categories from BioGPS database.

Data file S1. ErbB network (Cytoscape file).

Data file S2. High-resolution ErbB network (PDF).

Data file S3. Computational model of the Ras XOR motif (XML file).

Data file S4. Information for MEK2 antibody (PDF).


Acknowledgments: We thank B. Lehner, M. Lluch Senar, and T. Ferrar for critical reading and suggestions about the manuscript. Part of the network analysis involved the Bioinformatics Core Facility at the Centre for Genomic Regulation. Funding: The research leading to these results received funding from the European Union Seventh Framework Programme (FP7-HEALTH-2011) PROSPECTS and PRIMES project under grant agreement numbers FP7-HEALTH-F4-2008-201648 and FP7-HEALTH-2011-278568. This work was supported by the Spanish Ministerio de Economía y Competitividad, Plan Nacional BIO2012-39754 and the European Fund for Regional Development. Author contributions: C.K. and L.S. conceived the project and designed the experiments. C.K. performed the experiments and data analysis. C.K., E.V., and J.-S.Y. performed the protein interaction network analysis. C.K. and L.S. wrote the manuscript with inputs from the other authors. Competing interests: The authors declare that they have no competing interests.
View Abstract

Stay Connected to Science Signaling

Navigate This Article