Bioinformatics Advance Access originally published online on June 28, 2007
Bioinformatics 2007 23(17):2226-2230; doi:10.1093/bioinformatics/btm336
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Prediction of Ras-effector interactions using position energy matrices
EMBL-CRG Systems Biology Unit, CRG-Centre de Regulacio Genomica, Dr Aiguader 88, 08003 Barcelona, Spain
*To whom correspondence should be addressed.
 |
ABSTRACT |
|---|
 TOP ABSTRACT 1 INTRODUCTION2 METHODS AND RESULTS3 DISCUSSIONACKNOWLEDGEMENTREFERENCES |
|---|
Motivation: One of the more challenging problems in biology is to determine the cellular protein interaction network. Progress has been made to predict protein–protein interactions based on structural information, assuming that structural similar proteins interact in a similar way. In a previous publication, we have determined a genome-wide Ras-effector interaction network based on homology models, with a high accuracy of predicting binding and non-binding domains. However, for a prediction on a genome-wide scale, homology modelling is a time-consuming process. Therefore, we here successfully developed a faster method using position energy matrices, where based on different Ras-effector X-ray template structures, all amino acids in the effector binding domain are sequentially mutated to all other amino acid residues and the effect on binding energy is calculated. Those pre-calculated matrices can then be used to score for binding any Ras or effector sequences.
Results: Based on position energy matrices, the sequences of putative Ras-binding domains can be scanned quickly to calculate an energy sum value. By calibrating energy sum values using quantitative experimental binding data, thresholds can be defined and thus non-binding domains can be excluded quickly. Sequences which have energy sum values above this threshold are considered to be potential binding domains, and could be further analysed using homology modelling. This prediction method could be applied to other protein families sharing conserved interaction types, in order to determine in a fast way large scale cellular protein interaction networks. Thus, it could have an important impact on future in silico structural genomics approaches, in particular with regard to increasing structural proteomics efforts, aiming to determine all possible domain folds and interaction types.
Availability: All matrices are deposited in the ADAN database (http://adan-embl.ibmc.umh.es/).
Contact: christina.kiel{at}crg.es
Supplementary information: Supplementary data are available at Bioinformatics online.
|
1 INTRODUCTION |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS AND RESULTS3 DISCUSSIONACKNOWLEDGEMENTREFERENCES |
|---|
Prediction of protein–protein interactions based on structural information is an important tool in systems biology (Aloy and Russell, 2006; Beltrao et al., 2007). The prediction method is based on the finding that structural similar proteins usually interact in a similar way (Aloy and Russell, 2005; Aloy et al., 2005). One way of structure-based prediction is to generate homology models of proteins which have a similar sequence, and calculate the interaction energy of the modelled complex using protein design algorithms (Kiel et al., 2005). In order to generate a homology model, the amino acid side chains of a protein complex are replaced by the corresponding amino acid side chains of two other proteins which belong to the same families, but where no structural information is available. Prediction of protein interactions using homology modelling and energy calculations has been successfully done for protein family members sharing sequence homology (Kiel et al., 2005; Kiel et al., 2007), with high levels of accuracy (Kiel et al., 2007).
Recently we have determined a genome-wide Ras-effector interaction network, for 20 Ras-like proteins in complex with 50 putative Ras-binding domains, based on homology modelling and energy calculations (Kiel et al., 2007). Predicting protein interactions based on homology modelling although accurate (Kiel et al., 2005; Kiel et al., 2007) is a computer time-consuming method, taking 
20 min per model, since it requires full side chain reconstruction of the complex. This makes in silico structural genomics a time-consuming task. Here, we develop a faster method which implies the calculation of position energy matrices, and we applied this method to the prediction of Ras-effector interactions for which complete models have been made (Kiel et al., 2007). These complexes are an example of a conserved interaction type: the interface of Ras proteins in complex with the Ubiquitin-like domain of effector proteins is mainly formed by two β-sheets (β2 and β3 of Ras and β1and β2 of the effector protein), and by the first helix of the Ubiquitin-like domain.
Ras proteins belong to the superfamily of small GTP-binding proteins where more than 150 proteins have been identified up to now (Colicelli et al., 2004;Takai et al., 2001; Vetter and Wittinghofer, 2001). Members of the Ras-subfamily play an important role in various signal transduction pathways, like proliferation and differentiation. Similar to all other guanine nucleotide binding proteins, Ras proteins have the ability to cycle between an inactive GDP- and an active GTP-bound form (Bourne et al., 1990; Bourne et al., 1991). In the active form the interaction with effectors is possible, which preferentially bind to the GTP-bound form of Ras, and thereby communication into different signaling pathways is achieved.
Effector proteins are usually multi domain proteins and the interactions with Ras-like proteins involves a domain with a topology similar to Ubiquitin, the Ubiquitin-like fold. Thus, they are members of the UB domain superfamily (Orengo et al., 1994). The overall structure and topology of Ubiquitin-like domains is very similar, but large changes in the loop length and slightly different orientations of some secondary structure elements might occur. The sequence identity is rather low and based on a minimal sequence homolog UB domain are classified into five subfamilies in databases like SMART (Letunic et al., 2002) or Pfam (Bateman et al., 2004): the RA (Ras-association) (Ponting and Benjamin, 1996), the RBD (Ras-binding domain), the PI3K_rbd (Ras-binding domain of PI3 kinase), the UBQ (Ubiquitin) and the B41/ERM (ezrin/radixin/moesin) domain families.
|
2 METHODS AND RESULTS |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS AND RESULTS3 DISCUSSIONACKNOWLEDGEMENTREFERENCES |
|---|
2.1 Generation of 120 template structures
A general flow scheme for the generation of template structures and matrices is shown in Figure 1.
|
All Ras-binding domains of effector proteins which have been structurally characterized up to now show a similar topology, the ubiquitin-like fold and complex formation between Ras and effector proteins is mainly mediated by the interaction of two β-sheets (β2 and β3 of Ras and β1and β2 of the effector protein), and by the first helix of the effector proteins (Bunney et al., 2006; Huang et al., 1998; Nassar et al., 1995, 1996; Pacold et al., 2000; Scheffzek et al., 2001; Vetter et al., 1999) (Supplementary Fig. 1a). Although the general binding mode is similar there are small changes in the details of how the interface is formed. Thus, in order to account for this conformational flexibility, all available template structures are used. Out of the seven available X-ray Ras-effector complex structures, we have selected five [which are of good quality (<3 Å resolution) and are complete] to be modified to be used as template structures (similar as done in Kiel et al., 2007). From the Ras-RalGDS structure (PDB entry: 1LFD) (Huang et al., 1998), we selected both complexes in the crystallographic unit, molecules AB (template T1a) and molecules CD (template T1b). The Ras-PI3K complex (PDB entry: 1HE8) (Pacold et al., 2000) was selected as template T2, the Ras-spByr structure (PDB entry: 1K8R) (Scheffzek et al., 2001) as template T3, the Rap-RafRBD complex (PDB entry: 1GUA) (Nassar et al., 1996) as template T4 and the Ras-PLCeRA2 complex (PDB entry: 2C5L) (Bunney et al., 2006) as template T5.
Since the main contribution to complex affinity in Ras-effector complexes originates from residues in the first three secondary structure elements of the effector domain (β1, β2 and 
1), the six selected template structures (T1a, T1b, T2, T3, T4, T5) have been further modified by deleting all secondary structure elements and connecting loops except β1, β2, and 1 (Kiel et al., 2007). An overlay of all template Ras-effector structures used in this study is shown in Supplementary Figure 1b. The lengths of all template structures are summarized in Supplementary Table 1.
For comparison purposes we have selected the same 20 Ras-like proteins as in the previous study (Kiel et al., 2007). In order to generate the Ras-like proteins, we have only mutated residues which are in the interface or at the edge of the interface (Supplementary Fig. 2a) using version 2.7 of FoldX. The rest of the protein was kept unchanged. The reason been that all Ras members so far are very similar in sequence and structure [in our previous work (Kiel et al., 2007) we checked that the interface residues of other Ras members were compatible with the rest of the WT Ras structure]. All indicated positions were mutated according to the alignment of the selected Ras family members as shown in Supplementary Figure 2b. Thus we have generated 120 modified template structures, which are then used for generating the position scan energy matrices.
2.2 Energy matrices
The 120 templates structures are then further modified by mutating all residues in the Ras-binding domain to alanine. The reason for this is that RBD domains are quite different in sequence and therefore there could be incompatibility of certain amino acids within a certain sequence context. On each of the 120 Ras-alanine effectors all residues at the interphase of the effector are mutated independently to all other 20 amino acid residues and the energy difference in binding calculated using FoldX (Guerois et al., 2002; Schymkowitz et al., 2005a, Schymkowitz et al., 2005b). An example of the output for a position energy matrix is shown in Supplementary Figure 3. Depending on the position and the type of amino acid in the Ras-binding domain, the side chain residues contribute either favourable (positive values) or non-favourable (negative values) to complex formation. All matrices are deposited in the ADAN database (http://adan-embl.ibmc.umh.es/).
2.3 Calculation of energy sum values
The alignment for all Ras-binding domains is similar to that from our recent publication (Kiel et al., 2007) (Supplementary Fig. 4). In order to get the energy sum values based on the energy matrices for the different Ras effectors, the energy value for a specific amino acid at a specific position is added up according to the alignment of Ras-binding domains. In order to see how accurate the energy sum reflects the binding affinity of the complexes, we compared the fully homology modelled 120 template Ras-effector structures complex energies with the corresponding energy sum values (Supplementary Fig. 5). We find a good correlation of energy sum values derived from matrices and the interaction energies of template structures (R = 0.74) (Supplementary Fig. 5a). In order to take into account ‘intrinsic’ energetic properties of each template structures, like backbone interactions, we added to the energy sum the complex energy of the corresponding poly-alanine template structure (Supplementary Table 2). By adding this energy to the energy sum values derived from matrices, we observe a very good correlation (R = 0.81) (Supplementary Fig. 5b). Thus, for all energy sum values the interaction energy of the corresponding poly-alanine template was added (Supplementary Table 2), which we refer to as energy sum values here.
2.4 Thresholds for predicting non-binding and binding domains
Similar as to the concept of homology modelling previously published (Kiel et al., 2007), we selected the lowest energy sum value for a particular interaction, calculated based on the six matrices of one template Ras-effector template (Supplementary Table 2). The reason being that we do not know based on sequence what will be the best effector template for a particular sequence. In Figure 2, we show the distribution of energy sum values after selecting the best energy sum values between the six different template structures. In order to define thresholds for each of the six template structures to decide if an effector domain will bind a particular Ras protein, we used the experimental information currently available (Supplementary Table 3). Experimental information was taken from published data, but only when true biophysical methods, like ITC (isothermal titration calorimetry) or fluorescence-based methods were used. The experimental information was added into the diagrams where the binding energies of all Ras and effector combinations tested on a particular template were displayed (Fig. 2). In red we show experimental binding information for complexes of non-template sequences, and in blue we show experimental binding information for complexes of template sequences (e.g. RalGDS, Raf, PLCeRA2 and PI3K). The size of the squares correlates with the magnitude of interaction.
|
Using this information a threshold of –16.59 kcal/mol was found for predicting non-binding sequences, if one false positive is excluded (Fig. 2). Sequences below this threshold are predicted to be non-binding domains and can be excluded quickly. Sequences which have energy sum values above this threshold are considered to be potential binding domains, and could be further analysed using homology modelling. Energy sum values above –22 kcal/mol have a high probability to be binding domains (81%). Thus, we take this threshold to test how good our prediction success is to predict binding domains.
2.5 Accuracy of the prediction method and saved time compared to homology models
The accuracy of predicting non-binding Ras-binding domains was calculated by using a set of 70 pull-down experimental binding information from literature and our previous publication (Kiel et al., 2007). For 12 complexes where non-binding was predicted based on matrices, and where pull-down results are available, 10 complexes showed non-binding in pull-down experiments or gel-bands with less then 2-fold intensity compared to the control, and in only 2 cases binding was found experimentally. Thus, the accuracy of excluding non-binding is good (0.67), not counting the cases with less than 2-fold intensity (Fig. 3). In comparison, the accuracy of predicting non-binding domains using homology modelling was 0.90 (Kiel et al., 2007), where for 30 non-binding predicted complexes only two showed binding in experiments (Fig. 3). Using the binding threshold for energy sum values of –22, a prediction accuracy of predicting binding domains was found to be good as well (0.73), but more non-binding complexes are found here, where binding was predicted experimentally, and thus the accuracy is better using homology modelling (Fig. 3).
|
In our previous study, we have predicted the interaction of 50 potential Ras-binding domains in complex with 20 different Ras proteins using homology modelling and energy calculations (Kiel et al., 2007). Out of these 1000 interactions, 409 domains are predicted not to bind [including the 120 template structures we used in this study to generate the matrices (Kiel et al., 2007)]. Using energy matrices, 309 complexes (from 1000) are below the threshold of –16.59 (non-binding). Total 174 of these are predicted to be non-binding using homology modelling, 130 are predicted to be in the twilight zone and only 3 are predicted to bind in homology modelling (Kiel et al., 2007). It is important to mention that many of the energy values of homology models in the twilight zone are close to the threshold of non-binding, and that also one third of the homology models predicted to be in the twilight zone did non-bind, as shown by pull-down experiments.
|
3 DISCUSSION |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS AND RESULTS3 DISCUSSIONACKNOWLEDGEMENTREFERENCES |
|---|
We have used position energy matrices in order to predict Ras-effector interactions. Based on energy sum values and calibration with experimental binding information, potential Ras-binding domains can be scanned quickly and non-binding domains can be sorted. Sequences which have energy sum values above this threshold are considered to be potential binding domains, and could be further analysed using homology modelling. The accuracy of predicting binding and non-binding domains is not so high compared to accurate homology models (Kiel et al., 2007). However, it is much faster and thus it is a useful tool to quickly test the binding of potential new Ras-binding domains.
All matrices are deposit in the ADAN database (http://adan-embl.ibmc.umh.es/), which opens the possibility to screen further sequences, which are predicted to have a Ubiquitin-like fold, and predict whether they could interact with one of the 20 different Ras proteins. In our previous study, 50 sequences belonging to one of the five Ubiquitin-like fold subfamilies, mainly sequences of the RA, RBD and PI3K_rbs subfamilies. However, there are many more sequences with a predicted Ubiquitin-like topology. Some of them we have excluded them so far, since the sequence cannot be modelled reliable with one of the available template structures (Kiel et al., 2007). But with new template X-ray Ras-effector complex structures available, more sequences could be modelled. Further it is important to mention, that many more domains with a Ubiquitin-like topology might exist, but which are not recorded in one of the five Ubiquitin-subfamilies in the SMART/Pfam databases, due to very low sequence homology, as found for the Ubiquitin-like domain of the PlexinB1 receptor (Kiel and Serrano, 2006; Tong and Buck, 2005). New methods in domain predictions, structure-based alignments, secondary structure prediction tools, etc. (Kiel and Serrano, 2006), might reveal new Ubiquitin-like sequences, which could be screened for binding with the available matrices.
Prediction of protein–protein interactions based on energy matrices could be very important in future in silico approaches in structural genomics and systems biology, which aim to fully predict, model, and understand the protein interaction network of a cell. After the complete sequencing of several genomes, the completion of 3D structures of all possible domain folds and interaction types would open the possibility to fully model and understand biological systems (Aloy and Russell, 2004; Banci et al., 2007; Bork and Serrano, 2005). Structural proteomics efforts have already found 700–800 different folds (Aloy and Russell, 2004) of the predicted 1000 domain folds in nature (Chotia, 1992) and 2000 of the predicted 10 000 domain–domain interaction types (Aloy and Russell, 2004). Matrices of all interaction types can be stored and sequences belonging to a similar domain family can be scanned quickly, in order to filter out clear non-binding domains.
Using matrices will probably be even more important as a tool in predicting protein–peptide interactions in silico. The interface of protein–peptide complexes is usually more flexible within a similar domain family, and therefore more template structures are needed to model all sequences (Fernandez-Ballester and Serrano, 2006). Thus, if many matrices of protein–peptide databases are already stored in a database, potential binding peptides can be scanned quickly. In addition, large scale screening for possible target peptides does not involve any domain prediction, alignments or loop modelling, and therefore, whole genomes can be scanned quickly (Fernandez-Ballester et al., in preparation).
|
ACKNOWLEDGEMENT |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS AND RESULTS3 DISCUSSIONACKNOWLEDGEMENTREFERENCES |
|---|
We thank the EU for financial support (INTERACTION PROTEOME, grant-No. LSHG-CT-2003-505520).
Conflict of Interest: none declared.
|
FOOTNOTES |
|---|
Associate Editor: Anna Tramontano
Received on May 11, 2007; revised on May 11, 2007; accepted on June 16, 2007
|
REFERENCES |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS AND RESULTS3 DISCUSSIONACKNOWLEDGEMENTREFERENCES |
|---|
Aloy P, Russell RB. Ten thousand interactions for the molecular biologist. Nat. Biotechnol, ( (2004) ) 22, : 1317–1321.[CrossRef][ISI][Medline].
Aloy P, Russell RB. Structure-based systems biology: a zoom lens for the cell. FEBS Lett, ( (2005) ) 579, : 1854–1858.[CrossRef][ISI][Medline].
Aloy P, Russell RB. Structural systems biology: modelling protein interactions. Nat. Rev. Mol. Cell. Biol, ( (2006) ) 3, : 188–197..
Aloy P, et al. Protein complexes: structure prediction challenges for the 21st century. Curr. Opin. Struct. Biol, ( (2005) ) 1, : 15–22..
Banci L, et al. Structural proteomics: from the molecule to the system. Nat. Struct. Mol. Biol, ( (2007) ) 14, : 3–4.[CrossRef][ISI][Medline].
Bateman A, et al. The Pfam protein families database. Nucleic Acids Res, ( (2004) ) 32, : D138–D141.
Beltrao P, et al. Structures in systems biology. Curr. Opin. Struct. Biol, ( (2007) ) in press..
Bork P, Serrano L. Towards cellular systems in 4D. Cell, ( (2005) ) 121, : 507–509.[CrossRef][ISI][Medline].
Bourne HR, et al. The GTPase superfamily: a conserved switch for diverse cell functions. Nature, ( (1990) ) 348, : 125–132.[CrossRef][Medline].
Bourne HR, et al. The GTPase superfamily: conserved structure and molecular mechanism. Nature, ( (1991) ) 349, : 117–127.[CrossRef][Medline].
Bunney TD, et al. Structural and mechanistic insights into ras association domains of phospholipase C epsilon. Mol. Cell, ( (2006) ) 21, : 495–507.[CrossRef][ISI][Medline].
Chothia C. One thousand families for the molecular biologist. Nature, ( (1992) ) 357, : 543–544.[CrossRef][Medline].
Colicelli J. Human RAS superfamily proteins and related GTPases. Sci. STKE, ( (2004) ) 250, : RE13–RE31..
Fernandez-Ballester G, Serrano L. Prediction of protein-protein interactions based on structure. Methods Mol. Biol, ( (2006) ) 340, : 207–234.[Medline].
Guerois R, et al. Predicting changes in the stability of proteins and protein complexes: a study of more than 1000 mutations. J. Mol. Biol, ( (2002) ) 320, : 369–387.[CrossRef][ISI][Medline].
Huang L, et al. Structural basis for the interaction of Ras with RalGDS. Nat. Struct. Biol, ( (1998) ) 5, : 422–426.[CrossRef][ISI][Medline].
Kiel C, Serrano L. The ubiquitin domain superfold: structure-based sequence alignments and characterization of binding epitopes. J. Mol. Biol, ( (2006) ) 355, : 821–844.[CrossRef][ISI][Medline].
Kiel C, et al. Recognizing and defining true Ras binding domains II: in silico prediction based on homology modelling and energy calculations. J. Mol. Biol, ( (2005) ) 348, : 759–775.[CrossRef][ISI][Medline].
Kiel C, et al. A genome-wide Ras-effector interaction network. J. Mol. Biol, ( (2007) ) 370, : 1020–1032.[CrossRef][Medline].
Letunic I, et al. Recent improvements to the SMART domain-based sequence annotation resource. Nucleic Acids Res, ( (2002) ) 30, : 242–244.
Nassar N, et al. The 2.2 Å crystal structure of the Ras-binding domain of the serine/threonine kinase c-Raf1 in complex with Rap1A and a GTP analogue. Nature, ( (1995) ) 375, : 554–560.[CrossRef][Medline].
Nassar N, et al. Ras/Rap effector specificity determined by charge reversal. Nat. Struct. Biol, ( (1996) ) 3, : 723–729.[CrossRef][ISI][Medline].
Orengo CA, et al. Protein superfamilies and domain superfolds. Nature, ( (1994) ) 372, : 631–634.[CrossRef][Medline].
Pacold ME, et al. Crystal structure and functional analysis of Ras binding to its effector phosphoinoside 3-kinase gamma. Cell, ( (2000) ) 103, : 931–943.[CrossRef][ISI][Medline].
Ponting CP, Benjamin DR. A novel family of ras-binding domains. Trends Biochem. Sci, ( (1996) ) 21, : 422–425.[CrossRef][ISI][Medline].
Scheffzek K, et al. The Ras-Byr2RBD complex: structural basis for Ras effector recognition in yeast. Structure, ( (2001) ) 9, : 1043–1050.[Medline].
Schymkowitz J, et al. The FoldX web server: an online force field. Nucleic Acids Res, ( (2005) ) 33, : W382–W388.
Schymkowitz JW, et al. Prediction of water and metal binding sites and their affinities by using the Fold-X force field. Proc. Natl Acad. Sci. USA, ( (2005) ) 102, : 10147–10152.
Takai Y, et al. Small GTP binding proteins. Physiol. Rev, ( (2001) ) 81, : 153–208.
Tong Y, Buck M. 1H, 15N and 13C Resonance assignments and secondary structure determination reveal that the minimal Rac1 GTPase binding domain of plexin-B1 has a ubiquitin fold. J. Biomol. NMR, ( (2005) ) 31, : 369–370.[CrossRef][ISI][Medline].
Vetter IR, Wittinghofer A. Signal transduction – the guanine nucleotide-binding switch in three dimensions. Science, ( (2001) ) 294, : 1299–1304.
Vetter IR, et al. Structural and biochemical analysis of Ras-effector signaling via RalGDS. FEBS Lett, ( (1999) ) 451, : 175–180.[CrossRef][ISI][Medline].
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||