Modeling of protein binary complexes using structural mass spectrometry data

Abstract

In this article, we describe a general approach to modeling the structure of binary protein complexes using structural mass spectrometry data combined with molecular docking. In the first step, hydroxyl radical mediated oxidative protein footprinting is used to identify residues that experience conformational reorganization due to binding or participate in the binding interface. In the second step, a three-dimensional atomic structure of the complex is derived by computational modeling. Homology modeling approaches are used to define the structures of the individual proteins if footprinting detects significant conformational reorganization as a function of complex formation. A three-dimensional model of the complex is constructed from these binary partners using the ClusPro program, which is composed of docking, energy filtering, and clustering steps. Footprinting data are used to incorporate constraints—positive and/or negative—in the docking step and are also used to decide the type of energy filter—electrostatics or desolvation—in the successive energy-filtering step. By using this approach, we examine the structure of a number of binary complexes of monomeric actin and compare the results to crystallographic data. Based on docking alone, a number of competing models with widely varying structures are observed, one of which is likely to agree with crystallographic data. When the docking steps are guided by footprinting data, accurate models emerge as top scoring. We demonstrate this method with the actin/gelsolin segment-1 complex. We also provide a structural model for the actin/cofilin complex using this approach which does not have a crystal or NMR structure.

Knowledge of the three-dimensional structure of proteins and their structural and functional interactions in protein complexes is critical to understanding a myriad of processes in cell biology. Despite the obvious importance of this structural information, the challenges to obtaining it are considerable using current technologies. It is well known that the number of structural entries for proteins in the Protein Data Bank (PDB) (42,350 protein entries as of September 25, 2007) is 50–100 times lower than the number of open reading frame sequences in GenBank/TREMBL. When these databases are clustered in terms of homologous sequences, the gap between sequence and structure is not lessened at all. However, structural genomics projects in the United States and worldwide are attempting to bridge this gap by solving structures of a selected set of proteins derived from as many protein families as possible (Bonanno et al. 2005; Chandonia and Brenner 2005). These template structures, at least one for each family clustered at 30% identity, provide anchor points for high-throughput homology modeling to provide accurate models for the remaining family members (Baker and Sali 2001; Vitkup et al. 2001). Thus, there is some prospect for providing highly accurate structural models for a large fraction of all known genomes.

For protein complexes, the prospects are daunting by comparison. Data from high-throughput proteomics projects, including both “pulldown” and yeast-2-hybrid experiments, indicate that the total number of protein interactions in the cell may exceed the total number of proteins (von Mering et al. 2002; Reboul et al. 2003; Tyers and Mann 2003; Gavin et al. 2006; Ewing et al. 2007). On the other hand, our knowledge of the details of these structural interactions, as reflected by the number of protein complexes deposited in PDB, is currently limited to 3990 crystal structure entries for heterodimers as of September 25, 2007 (http://pqs.ebi.ac.uk/pqs-doc.shtml). Since the number of authentic interactions is at least as large as the number of proteins, the gap between the number of complexes currently solved and the number of interest for structure elucidation likely exceeds 1000-fold, considering that several million open reading frame sequences currently exist (Chance et al. 2004). Progress toward bridging this gap will involve significant advances in experimental and computational methods. As for the structural genomics projects, where a homology-based approach was used to bridge the gap between the number of structures and the number of sequences, it is likely that the structure of carefully selected “template complexes” must be identified and solved; these structures will be leveraged by computational modeling to provide models of the appropriate orthologous “sibling complexes.” Concerted efforts to solve these template structures will require a hybrid approach to structure solution that combines both high- and medium-resolution structural data (Sali et al. 2003; Chiu et al. 2006).

The barriers to determining the structure of complexes include known limitations in crystallography and nuclear magnetic resonance (NMR) technologies: Issues such as complex size, crystallizability, solubility, and amounts of materials are well known. In recent years, electron microscopy (EM) and tomography techniques, particularly at low temperatures, have substantially improved and are making important contributions to determining the structure of complexes (Sali et al. 2003; Chiu et al. 2006). These approaches have resolution limitations for many samples and are better for larger complexes or cells due to sample dose issues. This leaves a gap in technological progress for the “medium” size complexes; particularly medium-sized binary complexes (50–200 kD). This has spurred the development of a host of computational methods that can fill in the gap and contribute to understanding the relationship between protein structure and function.

The most reliable computer-based technique for generating three-dimensional models of protein–protein complexes is using docking algorithms that compare surface complementarity. The quality of this comparison is highly dependent on the quality of the input structures (Halperin et al. 2002; Kontoyianni et al. 2004; Perola et al. 2004). High accuracy homology modeling (root mean square deviation [RMSD] ∼1Å), which can provide high-quality input structures for docking, can be achieved when the target and template proteins have sequence identities of >50% (Baker and Sali 2001); the accuracy drops significantly when the identity is <30% (Baker and Sali 2001). The performance of docking programs is also highly dependent on the nature of the binding site, e.g., the driving forces for binding and specificity. Several docking procedures have been recently developed that use different docking algorithms and scoring functions, and these have been evaluated in CAPRI (critical assessment of prediction of interactions) (Janin et al. 2003; Mendez et al. 2003). In the CAPRI competition, the unbound three-dimensional coordinates of the binary partners are given to researchers prior to the release of the coordinates of the cocrystallized complex; this serves as a blind test for evaluating the current status of the field for predicting the structural details of protein–protein interactions. In these competitions, the correct structure can generally be predicted but cannot be distinguished from a number of equally good possibilities (e.g., many false positives). Recently, a fast algorithm has been developed, namely, ClusPro, which energy-filters docked conformations having good surface complementarity and ranks them based on their clustering properties (Comeau et al. 2004a,b; Comeau 2007). ClusPro is the first automated server that has participated in the CAPRI experiment; it has made several correct predictions within 24 h of the release of the coordinates (Comeau et al. 2005). The free energy filters provided in ClusPro select docked complexes with minimized desolvation or electrostatic energies. Clustering is then used to smooth the local minima and to select the candidates with the broadest energy wells. The robustness of the method was tested for 48 pairs of interacting proteins. The probability of the correct structure being “top scored” is 27%, while the probability of the correct structure being one of the top 10 models is 65% and within the top 30 models is 81%. Thus, additional experimental information that could reject the false positives in the top 20–30 models would be a valuable addition to the modeling process. In particular, identification of interface residues that could guide the docking step and recognition of the driving forces defining the interface to guide the energy minimization step are crucial to the modeling. With these limitations in mind, docking approaches where ClusPro is constrained with information from mutagenesis (Pons et al. 2006), NMR (Gruschus et al. 2004; Pons et al. 2006) and cryoEM (Agbulut et al. 2007) experiments have been recently reported in the literature.

Recently, we used a combination of radiolytic footprinting coupled to mass spectrometry analysis and docking with the ClusPro server to derive a structure for the actin/cofilin binary complex (Kamal et al. 2007). Actin together with its functionally interacting proteins such as cofilin, gelsolin, profilin, Arp 2/3, and WASP regulate many cellular processes, including cell motility, cell morphology, cell division, exocytosis and endocystosis (Amann and Pollard 2000; Pollard et al. 2000; Paavilainen et al. 2004). Cofilin, upon binding to actin filament causes depolymerization at the pointed end, thereby preventing their reassembly, and severs the barbed end, thereby promoting filament extension, which is the basis for cell motility and cell protrusion (Ghosh et al. 2004; DesMarais et al. 2005). In this article, we outline a general approach using multiple examples of binary complexes that include monomeric actin to illustrate the power and limitations of the current method and to provide a guide to other researchers in the field. Specifically, we show how the ClusPro server performs in predicting the structure of the actin/DNase I and actin/profilin structures in the absence of additional experimental data; we show how the use of ClusPro in conjunction with previously published footprinting can provide a model of the actin/gelsolin segment-1 (GS1) complex that agrees with crystallographic data; and we show how the method is used to derive the structure of the actin/cofilin complex which does not have a crystal or NMR structure.

Results

Our general strategy of protein binary complex structure determination involves conducting radiolytic footprinting experiments with mass spectrometry analysis to derive information on the residues participating in the binding interface and/or those involved in conformational reorganization. This is followed by homology modeling if the bound conformation of either protein significantly deviated from the free form and then protein–protein docking by ClusPro server (Comeau et al. 2004a,b; Comeau 2007). The top 20,000 docked conformations are filtered either by electrostatics or by desolvation free energy filtering provided in the ClusPro. The choice of energy filter is critical in providing appropriate models. The top 2000 energy minimized structures are clustered and ranked according to their cluster sizes. The default value of the clustering radius (9 Å) is used if not stated otherwise. The general approach is illustrated in Scheme 1 and relevant steps are explained in the examples that follow.

Scheme 1.

General strategy of protein binary complex structure determination from footprinting data and computational modeling.

Docking actin/ABP complexes of known crystal structures without experimental constraints: G-actin/DNase I and G-actin/profilin

Using ClusPro, we carried out molecular docking of selected actin-binding proteins (ABPs) to G-actin where the crystal structures are known. This allows us to both develop and judge the effectiveness of the method; the various parameters of the resulting docked models are provided in Table 1. The individual components of the indicated crystal structures of actin and its ABP complex were docked to obtain the respective complexes. The resulting models were compared with the respective crystal structures, and their RMSDs were calculated and interface parameters were derived. For docking of actin/DNase I and actin/profilin, the individual chains from the PDB structures of the two complexes, 1ATN (Fig. 1A) and 2BTF (Fig. 1B), respectively, were used. Actin and DNase I exhibit (overall) negative electrostatic potential surface patterns (Fig. 2A,B) in accordance with their acidic pI values (∼5.2). However, their binding surfaces comprise mostly neutral/hydrophobic residues, suggesting a hydrophobic driven mode of binding. Consistently, the interface of the crystal structure showed a higher percentage for nonpolar atoms over polar atoms for both actin and DNase I (29% and 41% polar, respectively) (Table 1). Based on this analysis of the interface, we chose desolvation free energy filtering of the docked conformations for actin/DNase I from the ClusPro energy filtering options (Comeau et al. 2004a,b). Profilin exhibits an overall positively charged electrostatic surface pattern (pI = 8.4) (Fig. 2C). Its binding site for actin also exhibits many positive and several negatively charged residues. Specifically, the binding interface of actin/profilin complex consists of appropriate positioning of oppositely charged residues from the individual components, suggesting that the mode of binding is primarily driven by electrostatics. In this case, we chose electrostatic free energy filtering for actin/profilin docking.

Table 1.

Summary of the docking and structural interface parameters of the docked models of actin/ABP complexes

Figure 1.

Crystal structures of G-actin/actin-binding protein complexes and cofilin. (A) Structures of G-actin/DNase I (PDB code 1ATN), (B) G-actin/GS1 (1YAG), (C) G-actin/profilin (2BTF), and (D) cofilin (1COF). The red-colored region in the cofilin structure is the G-actin-binding site (G/F site) established by footprinting (Guan et al. 2002), mutagenesis (Lappalainen et al. 1997), and NMR (Pope et al. 2004). The specific footprinting probe residues that form the G-actin-binding site are indicated with stick models. Helices and β-strands are indicated with α and β symbols, including their serial numbers.

Figure 2.

Electrostatic potential surfaces of actin and actin-binding proteins. (A) DNase I in the actin-binding orientation, (B) actin, (C) profilin in the actin-binding orientation, (D) gelsolin S1 (GS1) in the actin-binding orientation, (E) gelsolin S1 (GS1) rotated 180° from the actin-binding orientation, (F) cofilin in the actin-binding orientation, and (G) cofilin rotated 90° to the right from the actin-binding orientation. Red color is for negative electrostatics, blue color is for positive electrostatics, and white is for neutral.

The five top-scoring models for each docking exercise are shown in Fig. 3A,C. The top-ranked model obtained for the actin/DNase I docking (Fig. 3A) does not agree with the crystal structure (Fig. 1A); its RMSD compared with the crystallographic structure is 29.4 Å. Nevertheless, the second-ranked model (Fig. 3B) is very similar to the crystal structure, with an RMSD of the DNase I = 3.2 Å (Table 1). In order to evaluate the geometry of the interface alone, we also calculated the RMSD of the crystallographically established interfacial residues; the residues that showed a minimum of 1 Ų solvent accessible surface area (SASA) decrease upon complex formation. This value for the second-ranked model is 1.6 Å (Table 1). The actin segments identified at the interface of the model are 38–49, 60–64, and 202–207, while for DNase I the interfacial segments are 13–14, 43–69, 91–96, and 112; these overlap closely with those identified in the crystal structure (actin: 38–47, 57–64, 202–207; DNase I: 13–14, 43–69, 79, 91–96, 114). In addition, four of the five predicted hydrogen bonds (H-bonds) at the interface of the model are identical to four of the nine H-bonds seen in the interface of the crystal structure; they are (actin-DNase I numbers) Arg39-Asp53, Gln41-Tyr65, Val43-Tyr65, and Gly63-His44.

Figure 3.

Modeling of G-actin/DNase I and G-actin/profilin. (A) Top five models obtained from the docking strategy for G-actin/DNase I. First-ranked model is pink, second-ranked model is yellow, third-ranked model is gray, fourth-ranked model is orange, and fifth-ranked model is greenish cyan. (B) The model of G-actin/DNase I that is closest to the crystal structure. (C) Top five models obtained from the docking for G-actin/profilin. Color codes follow the previous scheme. (D) The model of G-actin/profilin that is closest to the crystal structure.

For actin/profilin, the first-ranked model (Fig. 3D) is close to the crystal structure (Fig. 1B), with RMSD of 4.3 Å for profilin and 2.2 Å for interface residues (Table 1). The interface segments identified in the model (actin: 113–116, 166–173, 284–290, 354–355, 361–375; profilin: 56–60, 71–90, 99, 119–129) match well with those in the crystal structure (actin: 113–116, 133, 166–173, 283–290, 354–355, 361–375; profilin: 59–69, 71–91, 97–99, 117–129). In addition, all the four predicted H-bonds at the interface of the model are identical to four of the 11 H-bonds at the interface of the crystal structure; they are (actin-profilin numbers) Lys113-Glu82, Glu361-Lys125, Arg372-Arg74, and Arg372-Thr84. Other parameters such as interface SASA, planarity, and gap volume index (GVI) are also close to the values reported for the crystal structure for both actin/DNase I and actin/profilin (Table 1). Next we switched the energy filter options between the two complexes; electrostatic free energy filtering for actin/DNase I and desolvation free energy filtering for actin/profilin, and no structures close to crystal structure within 10 Å RMSD showed up in the top 10 ranked models.

These data suggest two important conclusions about the ClusPro modeling process. First, ClusPro requires that the mode of interaction for the binary complex (either desolvation or electrostatic) be correctly specified if the program is to provide candidate models consistent with the correct structure (e.g., from crystallography) of complexes with one of these energies being the dominant binding force. For complexes with significant contribution from both the binding energies, ClusPro gives reliable models with either the filter or by pooling conformations from both filters (Comeau et al. 2004a,b; Comeau 2007). Second, the presumed correct prediction may not be the top-scoring model even when the correct mode of interaction is specified. Thus, for a successful protocol to be developed for docking, the driving forces for binding must be understood and the structural nature of the interface must be defined to a degree. In the next section, we describe how footprinting data are used to provide both sets of required information.

Radiolytic footprinting: G-actin/GS1 and G-actin/cofilin

We previously conducted radiolytic footprinting experiments for G-actin/GS1 (Goldsmith et al. 2001; Guan et al. 2003), whose crystal structure is known, and for G-actin/cofilin (Guan et al. 2002; Kamal et al. 2007), whose crystal or NMR structure is not yet reported. In the footprinting experiment, protein samples in the radiolysis buffer are exposed to a white synchrotron X-ray beam for intervals from 0–200 ms. Exposed samples are subjected to enzymatic digestion (Trypsin, Asp-N, and Glu-C) followed by quantification of the peptides by liquid chromatography (LC) coupled electrospray ion-source mass spectrometry (ESI-MS). First-order rate constants of modification are derived from dose response curves (Guan and Chance 2005; Takamoto and Chance 2006). These modification rate values are in accordance with the number, type (specific reactivity) and solvent exposure of the reactive amino acids in the respective peptides (Guan et al. 2003, 2004, 2005; Guan and Chance 2005; Takamoto and Chance 2006). Footprinting data on the G-actin/GS1 complex (Goldsmith et al. 2001; Guan et al. 2003) revealed five peptides generated from the trypsin digestion of actin (119–147, 157–178, 292–312, 316–326, and 337–359) and the residues Cys374 and Phe375 of the Asp-N peptide 363–375 that are protected as a function of GS1 binding (Table 2). These peptides are located in the subdomains 1 and 3 and include the hinge helices (Gly137-Ser145 and Arg335-Ser348) at the cleft; this region is the presumed binding interface based on footprinting alone (McLaughlin et al. 1993). Eleven peptides had rates of oxidation that were unchanged within error (within 20%) upon GS1 binding (1–18, 19–28, 40–50, 63–68, 69–84, 85–95, 96–113, 184–191, 197–206, 329–335, and 360–372) (Guan et al. 2003); the footprinting experiments suggest these are located outside the binding interface. Among the three modifiable Asp-N digested GS1 peptides detected (25–49, 66–83, and 96–109) (Goldsmith et al. 2001), only the peptide 96–109 (probe residue is Phe104) was found protected upon actin/GS1 complex formation (Table 2), which corresponds to the peptide 72–84 (probe residue is Phe80) in the crystal structure of actin/GS1 (PDB code 1YAG). Overall, the data indicated specific sites of contact for both actin and GS1 and provided no evidence for any major conformational reorganization of actin or GS1 induced by complex formation.

Table 2.

Modification rates for actin and GS1 peptides that showed protection in the actin/GS1 complex

Table 3 lists the oxidation rates of actin and cofilin peptides in free form and in the bound form in the actin/cofilin complex (Guan et al. 2002; Kamal et al. 2007). Figure 4 depicts the locations of various protected residues on the structure of actin monomer. Structural comparisons (Blanchoin and Pollard 1998; Wriggers et al. 1998; Dominguez 2004) and competitive binding studies (Blanchoin and Pollard 1998; Mannherz et al. 2007) suggested cofilin binds to the cleft formed between subdomains 1 and 3 of actin as preceded by the binding of GS1 and profilin. Since footprinting data did not detect significant protections within the cleft between subdomains 1 and 3 as a function of cofilin binding (Table 3; Fig. 4), this hypothetical binding site of cofilin is ruled out in these experiments (Kamal et al. 2007). On the other hand, footprinting revealed significant protections for a number of nearby peptides (1–18, 51–61, 69–84, 85–95, 96–113, 118–125, 360–372) (Table 3) within subdomains 1 and 2, indicating the putative binding interface (Kamal et al. 2007). Protections afforded by probe residues within and nearby the nucleotide cleft are accounted for allosteric conformational rearrangement causing the closure of the nucleotide cleft (Kamal et al. 2007). Among the 11 modifiable cofilin peptides detected (Table 3) (Guan et al. 2002), only the peptides 4–20, 10–17, 83–96, 91–105, and 106–107 were found to be protected upon actin binding. The probe residues of these protected peptides (Leu13, Pro94, Met99, Leu108, and Leu112) are shown in Figure 1D. This G-actin binding region of cofilin identified by footprinting agreed with the mutational data (Met1-Gly5, Arg96, Lys98, Asp123, and Glu126) (Lappalainen et al. 1997) and were confirmed later by NMR (Ala7, Val95, Lys98, Met99, Tyr101, Ser103, Leu108, Arg109, Gln120, and Asp123) (Pope et al. 2004). The remaining six peptides showed no change in oxidation within the error limit, indicating the absence of any appreciable conformational reorganization of cofilin with actin binding.

Table 3.

Modification rates for actin and cofilin peptides in free and in the actin/cofilin complex

Figure 4.

G-actin structure indicating the protection sites revealed by radiolytic footprinting. (A) Modified amino acids are shown as stick models in the structure of G-actin (1ATN). Residues showing substantial protection are colored in red, moderate protection in yellow, and nearly no protection in cyan. The residue Tyr143 shown in blue-colored stick model shows negative protection (increased modification). (B) Actin structure tilted to left, exposing the substantially protected residues (colored red) clustering in the cleft between subdomains 1 and 2.

Docking of actin/ABP complex of known crystal structure with footprinting constraints: G-actin/GS1

Although the overall electrostatic surface pattern of GS1 is negative (pI = 5.5), its actin-binding site suggested by footprinting exhibits predominantly neutral/hydrophobic and negative electrostatic surface patterns (Fig. 2D,E). This suggests a major contribution for desolvation in the binding. Therefore, we used desolvation free energy filtering of the docked conformations. Without using any experimental constraints, the top-scored structure obtained (Fig. 5A) does not have GS1 located in the cleft between subdomains 1 and 3 and hence is not consistent with the footprinting data or with the crystal structure (the RMSD for GS1 compared with crystallographic data is 22.6 Å). The closest structure among the top 10 ranked models is the sixth-ranked one, which exhibits an RMSD of 7.4 Å for GS1 and 3.8 Å for the crystallographically established interface residues. This model is very different from the crystallographic geometry, especially the geometry of the N terminus; it does not show interactions with the C-terminal regions of actin (Leu349-Gln353 and Lys370-Phe375), which is one of the crystallographic established contact regions (Fig. 1C) and is indicated to be a contact point by footprinting data (Table 2). The interface SASA value for actin is 772 Ų and that for GS1 is 773 Ų, lower than that seen for other actin complexes while the GVI is higher (Table 1).

Figure 5.

Modeling of G-actin/GS1. (A) Top-scored model of actin/GS1 without any experimental constraints. (B) Model closest to the crystal structure with footprinting attracts constraint. (C) Model closest to crystal structure with footprinting blocks constraint.

In order to evaluate the extent of consistency with the interface region identified by the footprinting experiment, we calculated a parameter defined as the footprinting interface consistency score (FICS) (see Materials and Methods). The footprinting experiments defined a set of residues protected as a function of complex formation (24 actin residues and one GS1 residue) (Table 2). The probe residues that were recognized as interface residues in the model numbered six, including the actin residues Tyr143, Tyr166, Tyr169, Leu346, and Leu349 and the GS1 residue Phe80. Thus, the FICS score for actin for this model is 0.21 and for GS1 is 1.0.

Although the use of footprinting to provide the choice of driving force for binding did generate some models that were of interest, this was not sufficient to provide a definitive choice of model. Thus, to further improve the docking we set up “attract” constraints using selected footprinting probe residues from Table 2 that are positioned in the cleft between subdomain 1 and 3 (Tyr143, Tyr166, Tyr169, Leu349, Phe352, and Met355). The rank of the structure closest to the crystal structure improved to two from six, cluster size increased to 69 from 33, and the RMSD of the GS1 is decreased to 6.1 Å from 7.4 Å (Fig. 5B). The N terminus is found interacting with C-terminal regions of actin in this model consistent with crystal structure and footprinting data. The interface segments identified for actin in the model are 23–25, 133, 143–149, 166–169, 345–355, and 373–375 which overlap with those identified by footprinting data (119–147, 157–178, 292–312, 316–326, 337–359, and 374–375) and with those in the crystal structure (23–25, 143–148, 167–169, 292–296, 334, 341–355, and 373). The interface segments identified for GS1 in the model are 1–8, 24–27, 63, 71–94, and 125, which span the peptide 72–84 detected by footprinting data and overlap with all of the interface segments of the crystal structure (1–8, 25–26, 63–65, 71–86, 94–96, 124–125). One out of the two predicted H-bonds at the interface (actin-GS1 numbers), Thr350-Phe25, is conserved in the crystal structure. Three additional actin probe residues (Phe352, Met355, and Phe375) became buried in the attract model as compared with the “no constraint” model due to the appearance of the GS1 N-terminal interaction with the actin C terminus as discussed above. For this model, the FICS score for actin improved to 0.33 from 0.21.

Analysis of the FICS scores for the structural models suggests the importance of solution dynamics in interpreting the footprinting data. Typically, we detect more probe residues as being protected upon complex formation by footprinting than are seen to be structurally buried in the interface of the model. Probe residues from within and around the interface are likely to experience oxidation events due to their dynamic fluctuations but are observed as inaccessible in the free and bound forms of the static crystal structure. For example, the actin probe residues Met119, Met123, and Phe124 (from the peptide 119–147), which are found protected upon the complex formation, are completely buried in the free (SASA = 8.6, 0, and 0 Ų, respectively) and bound (SASA = 8.8, 0, and 0 Ų, respectively) forms and are also located in the proximity of the interface region, precisely at the backside of the cleft, which is not entirely overlapping with the GS1 binding site as GS1 binds at the front half of the cleft. Complex formation likely suppresses this dynamic behavior, such that oxidation is reduced although the precise residue in question does not form a contact. Another probe residue from the same peptide, Tyr143 (SASA = 63.9 Ų), is recognized as within the interface region in the model (SASA = 10 Ų) and is located within one of the hinge helices. Thus, our hypothesized interface may include closely lying accessible residues, and their protection arises due to a binding induced restriction of their conformational flexibility, without affecting the overall backbone conformation. For example, the actin probe residues Met305-Pro307 (from peptide 292–312), Pro322, and Met325 (from peptide 316–326), which are found protected upon actin/GS1 complex formation, lie just outside of, but in close proximity to, the interface (precisely at the front side of subdomain 3). Their protection can be attributed to differences in the side-chain dynamics, which are not reflected in the crystal structure or in the rigid body docking strategy. Therefore we expect FICS values not to approach 1.0 and use them only for comparing different models.

To further develop and test the method, we set up “block” constraints (without attract constraint) in the docking step in a separate run. From the footprinting probe residues that were found unchanged upon complex formation, several residues that are sufficiently exposed and somewhat remote from the hypothesized binding interface were selected for block constraints (these included 21, 40, 67, 69, 73, 79, 87, 88, 91, 101, 102, 190, 200, 201, and 202). The model closest to the crystal structure is now top-scored with an RMSD of 6.7 Å (Fig. 5C). The RMSD of the interface is 2.9 Å, and the value of the FICS is 0.29 for actin and 1.0 for GS1. As compared to the attract model, the footprinting residue Tyr166 of actin is not found at the interface of the block model due to the location of the residue being near the boundary region of the interface. A small orientation difference of the block model from the attract model causes this residue to fall out of the interface. The interface segments are very similar to those for the attract model (actin: 23–25, 143–148, 167–169, 341–356, 372–375; GS1: 1–8, 24–27, 63, 71–86, 94, 125). The two H-bonds predicted at the interface are Ser145-Gln71 and Thr351-Phe25, which exhibit a high degree of conservation with two H-bonds in the crystal structure, Ser144-Gln71 and Thr350-Phe25. The cluster size is increased to 65 from 33 of the no constraint model. Upon incorporating both attract and block constraints together, the model obtained has an increased cluster size. Slight variations in cluster size (84–93) and rank (1–3) are observed run to run due to overpopulation of other interfaces when both constraints are used together and the filtering is retaining hits at a relatively proportional rate (Comeau 2007). When the attract and/or block constraints are used, a larger number of docked conformations near attract residues and nonblocked residues are retained in the initial docking stage which results in a high ranking for that particular conformation. These data reveal that docking alone cannot distinguish between models. When models generated using attract, block, or both are analyzed, a single candidate in overall agreement with the crystal structure is provided.

Last, we examined the effect of clustering radius on the simulations. The clustering radius is dependent on the size of the ligand (Comeau et al. 2004a,b; Comeau 2007). However, there is no clear statistical data on the precise correlation between residue length of the ligand and clustering radius (Comeau 2007). We have used the default clustering radius of 9 Å for docking actin with DNase I, whose residue length is 260, and with profilin, whose residue length is 139, and obtained reliable results (RMSDs with crystal structures are 3.2 Å and 4.3 Å, respectively). The residue length of GS1 is 125, which is lower than those of DNase I and profilin. Therefore we decided to observe the effect of reducing the clustering radius for docking actin/GS1. In addition, as discussed earlier, footprinting revealed the involvement of the hinge helices of actin (from the cleft between subdomains 1 and 3) and the long actin-binding helix of GS1 in the binding. The geometry of the groove shaped by these hinge helices of actin provides a very narrow space for the positioning of the actin-binding helix of GS1. Thus from the perspective of the available information on the binding geometry, there is indication for the need of a lower clustering radius in order to sample only the structurally close neighbors for clustering. Therefore, we decreased the clustering radius from the default value of 9 Å to 8, 7, and 6 Å. In all cases, the closest model to crystal structure is found top-scored with RMSD of GS1 falling below 5 Å. The RMSD is 3.4 Å with 8 Å radius, 4.6 Å with 7 Å radius, and 2.5 Å with 6 Å radius. Thus, in carrying out the docking, the clustering radius must be tailored to the modeling problems in question. When a small binding partner is used a smaller clustering radius is required.

We also looked at the effect of using the incorrect energy filter (electrostatic free energy filter as opposed to desolvation free energy filter) with no, attract, and block constraints for docking actin/GS1. In these docking simulations no models with GS1 at the cleft between subdomain 1 and 3 showed up in the top 10 models without using any constraints. With attract or block constraints, eighth- and tenth-ranked models obtained have GS1 located somewhat close to the subdomains 1 and 3 cleft; however, their configurations are very different from the native conformation; their RMSDs are 17.0 Å and 25.5 Å, respectively, using attract constraints and 46.5 Å and 41.6 Å using block constraints. This indicates that although we target specific regions to have more ligand clustering in the docking step through attract and block constraints, use of an improper energy filter causes rejection or poor rankings for these conformations owing to a high-energy score calculated for the specified energy type. Therefore, the correct energy filter is required for the correct prediction to be top scoring, which is feasible only if there is experimental information about the binding surfaces. Overall, the results indicate that when the interface can be defined using attract and block constraints and the energy filtering is correctly defined. The model from the docking calculation that is most consistent with the experimental data, which has received a high scoring and often being placed as the top-most model, is the one which is actually most consistent with the crystallographic structure.

Docking of actin/ABP complex of unknown crystal structure with footprinting constraints: G-actin/cofilin

Since footprinting data revealed an allosteric conformational change for actin as a result of cofilin binding that could be attributed to the closure of its nucleotide cleft, we constructed at first, a model of G-actin in the putative bound form (closed nucleotide cleft conformation) by homology modeling using SWISS-MODEL (Schwede et al. 2003). The details are provided elsewhere (Kamal et al. 2007). Since footprinting data did not detect observable conformational changes for cofilin as a result of binding to G-actin, we used the available crystal structure of cofilin (1COF) (Fedorov et al. 1997) for docking to the modeled G-actin structure. The interface regions of actin and cofilin as revealed by the footprinting indicate oppositely charged electrostatic surface patterns (Fig. 2B,F,G). Therefore, we chose electrostatic free energy filtering of the docked structures for actin/cofilin compared with desolvation; this is consistent with the known pH-dependent binding of the proteins. We first docked actin and cofilin without incorporating any experimental constraints. The first-ranked structure (Fig. 6A) (cluster size 43) had neither the cofilin side of the interfacial residues nor the actin side of the interfacial residues as identified by footprinting. The second-ranked structure (cluster size 30) is consistent with the experimental data of the interface residues (Fig. 6B, see discussion below). The FICS value obtained is 0.10 for actin and 0.2 for cofilin. Actin probe residues His87 and Tyr91 and cofilin probe residue Met99 from the 26 protected residues shown bolded in Table 3 (21 actin residues and 5 cofilin residues) are recognized as interface residues in the model. A single H-bond (actin-cofilin numbers), Glu99-Lys105, is predicted to be at the interface. The third-ranked structure (Fig. 6C) has cofilin at the previously assumed binding location, e.g., cleft between subomains 1 and 3 (cluster size 23).

Figure 6.

Three top-ranked G-actin/cofilin models from the computational modeling strategy without incorporating experimental constraints. (A) First-ranked model. (B) Second-ranked model. (C) Third-ranked model. Red-colored region in actin marks the interface segments identified by footprinting (Kamal et al. 2007), while that in cofilin marks the interface segments identified by footprinting (Guan et al. 2002), mutagenesis (Lappalainen et al. 1997), and NMR (Pope et al. 2004).

Next we incorporated attract constraints in the docking step using seven proposed interfacial residues of G-actin from the footprinting data that are solvent exposed and/or located at the surface including Trp79 (SASA = 53 Ų), His87 (89 Ų), His88 (33 Ų), Tyr91 (72 Ų), Tyr362 (13 Ų), Pro367 (46 Ų), His371 (28 Ų). A similar conformation to the no constraint model seen in Figure 6B but with a slightly different orientation (RMSD of the cofilin of the attract model with that of the no constraint model is 3.8 Å), is top-ranked (cluster size 57) (Fig. 7), and the conventional model (Fig. 6C) ranks seventh (cluster size 20). The similarity of the top-ranked attract model (Fig. 7A) with the second-ranked no constraint model (Fig. 6B) suggests that this particular binding mode has high propensity for shape complementarity, such as would be found in a native interface structure. The FICS value of actin improved to 0.19 from 0.10 for this model, while that of cofilin remained the same at 0.20 (actin residues Trp79, His87, His88, and Tyr91; cofilin residue Met99). The actin segments identified at the interface of the model are 4–5, 48–54, 79–100, 118–128, and 359–363, and the cofilin segments identified are 59–60, 87–110, 117–126 and 134–138. These interface segments overlap with those identified by the footprinting (actin:1–18, 51–61, 69–84, 85–95, 96–113, 118–125, 360–372; cofilin:4–20, 10–17, 83–96, 91–105, 106–117). Although the involvement of cofilin N terminus in the G-actin binding is predicted by footprinting (Leu13) (Kamal et al. 2007), mutation (Met1-Gly5) (Lappalainen et al. 1997), and NMR (Ala7) (Pope et al. 2004), the model does not reflect this fact. This is due to the missing N-terminal segment of 1–6, due to which the nearby residues such as Val7 and Leu13 of the N terminus do not see a net change in SASA value from free to the model. However, the location and directionality of this segment clearly indicate its involvement in the binding (Figs. 6B, 7). In addition to the N terminus, a second segment of cofilin involved with the G-actin binding spans the long α3 helix (Ser89–Leu112), which is considered the actin-binding motif of the cofilin/ADF family of proteins, in common with other homologous proteins such as GS1. Footprinting identifies Pro94, Met99, Leu108, and Leu112 from this region. Mutation (Lappalainen et al. 1997) and NMR (Pope et al. 2004) recognizes residues Val95, Arg96, Lys98, Met99, Tyr101, Ser103, Leu108, and Arg109. The residues recognized by the model are Thr87, Ser89, Asp91, Thr92, Ala93, Val95, Arg96, Lys98, Met99, Ala102, Ser103, Lys105, Asp106, Arg109, and Arg110. This is the center of cofilin's actin-binding surface and interacts with the maximum number of actin segments (Gly48–Ser52, Trp79–Asn92, Glu125–Asn128, and Lys359–Asp363). Some of the charged cofilin interface residues identified by mutation (Lappalainen et al. 1997) and NMR (Pope et al. 2004), which could not be detected by the footprinting (possibly due to the decay of the modified charged residues), namely, Gln120, Asp123 and Glu126, from the β6-α4 loop, are recognized as interface residues in the model. However, footprinting does detect a hydrophobic residue Phe124 from this loop as not belonging to the interface. Consistently, this residue is not recognized as an interface residue in the model due to its pointing away from the surface, deeply buried within the protein matrix. Additional interface residues recognized from this region in the model are Thr117, Asp118, Val119, Thr122, and Ser125. These residues are followed by several residues from the α4 helix (C terminus), namely, Glu134, Arg135, and Arg138, that are also found to be part of the interface in the model. These regions (Thr117–Glu126 and Glu134–Arg138) are found interacting with the Glu93–Glu100 segment of actin. An additional region of interaction was suggested by the model as part of the G-actin-binding site that is not identified by footprinting, mutation or NMR; residues Glu59 and Asn60 from the tip of the β4-α2 loop, which interacts with the lower side of the actin's D-loop, especially with the Gln49 residue. The number of H-bonds predicted at the interface is nine and are Gln49–Asn60, His87–Ser89, Asn92–Asp91, Arg95–Asp123, Arg95–Ser125, Glu100–Arg135, Glu100–Arg138, Asn128–Lys105, and Lys359–Asp106. This substantial number of H-bonds at the interface that are comparable to those seen in the crystal structures of actin/DNase I, actin/profilin, and actin/GS1 suggest a high degree of accuracy of this model of actin/cofilin to be the native structure.

Figure 7.

Three-dimensional model of G-actin/cofilin complex. (A) Top-scored G-actin/cofilin model from the docking constrained with the footprinting data. (B) Side view (rotated 90° to the left) of the model.

With the incorporation of block constraint for regions within subdomain 3 (peptides 284–290 and 323–327) and subdomain 4 (peptides 198–203, 224–234 and 241–248) that are unlikely to be at the binding surface as shown by the footprinting data, a similar structure that ranked first with the attract constraint (RMSD between the two structures is only 1.4 Å) ranks again first with a cluster size of 48. The FICS value of actin is 0.14 and that of cofilin is 0.20. The lower actin FICS value for the block structure compared with attract structure (0.19) is because of the exclusion of Trp79 from the interface, which is located at the rear end of the cleft helix and belongs to the boundary region of the interface. The interface regions of the block model are very similar to the attract model (actin: 4, 48–52, 80–100, 125–128, 359–363; cofilin: 59–60, 89–110, 118–126, 135–138). The number of H-bonds predicted at the interface is 11. All of the H-bonds predicted for the attract model except Gln49-Asn60 are conserved in this model. In addition, it has two extra H-bonds: His87 has an additional bond with Thr122 apart from Ser89, and Glu100 has an additional bond with the second nitrogen atom of Arg138. Upon incorporating both attract and block constraints together, the same overall structure (RMSD compared with the attract structure is 1.0 Å and compared with that of the block structure is 1.7 Å) ranks first with an increased cluster size, 68. SASA calculations of actin and cofilin that engage in the binding interface of the putative model (attract and block models) derived are in the range reported for other actin/ABP structures (Table 1). Other interface parameters are also within the range reported (Table 1).

Next we tried the effect of the clustering radius on the docking results using the attract constraints. Decreasing the clustering radius to 8 Å did not change the ranking or the RMSD to any appreciable extent relative to the initial model (1.4 Å with the model of a 9 Å clustering radius), although the cluster size varied. Upon decreasing the clustering radius still further to 7 Å, the RMSD of the model increased to 4.6 Å. Decreasing the clustering radius to 6 Å not only increased the RMSD further (5.4 Å) but also changed the orientation of the bound cofilin such that the residues at the C terminus of actin do not interact with cofilin, which is not consistent with footprinting data and would reduce the FICS score. Cofilin's size is similar to that of profilin (residue length is 139); in the latter case, we obtained reliable results with a 9 Å clustering radius (4.3 Å RMSD) (Comeau et al. 2004a,b). In addition, the putative model of actin/cofilin shows a broad convex/concave interface, and such surfaces provide better results with a clustering radius of 8–9 Å in general (Comeau 2007). In such cases, a narrower clustering radius results in a lower sampling of structural neighbors for clustering, which may result in a model having appreciable deviation from the native structure.

Discussion

Due to technical limitations of crystallography, NMR, and cryoEM, bridging the gap between the protein sequence and structure requires the application of computational methods. However, a main drawback of the computational methods lies in distinguishing correct models from false positives. Incorporation of experimental data with computational modeling methods reduces the computational search time and facilitates correct predictions. There are many computational methods reported in the literature that incorporate experimental data, such as NMR, mutation, and cryoEM, in the structure determination of individual proteins as well as protein binary complexes. In the present study, we have developed a method to generate a three-dimensional model of protein binary complexes using structural mass spectrometry data. The project was originally conceived in order to provide a model for the actin/cofilin complex; here we extend it to the examination of a wide range of complexes. By using radiolytic footprinting approaches, we monitor the surface accessibility of the reactive residues in proteins; this accessibility can change due to allosteric effects of ligand binding or due to solvent exclusion at macromolecular interfaces.

Without the constraints provided by footprinting data, we were able to generate accurate models of actin/DNase I and actin/profilin using docking alone. For actin/profilin, the top-scoring model is consistent with the crystal structure, whereas for actin/DNase I, the second-ranked structure matched with the crystal structure. In the latter case, the top-scoring model deviates from the crystal structure by ∼30 Å RMSD. In addition, these models could only be provided if the correct energy filtering (desolvation or electrostatic) was chosen. When incorrect energy filtering was used, the models generated were >10 Å RMSD from the respective crystal structures. Overall, the probability of the correct structure being top-scored by ClusPro is only 27%, based on a study involving 48 pairs of interacting proteins. Therefore, picking the correct model from the multiple models generated by docking algorithms is very difficult without additional experimental information.

We used radiolytic footprinting data as experimental constraints to derive a model for actin/GS1 complex consistent with the known crystal structure and of actin/cofilin complex of unknown crystal or NMR structure. First, the binding (interface) region of the larger protein (actin in this case) identified by footprinting is examined and surface residues that are sufficiently solvent exposed in that binding partner are chosen as attract residues to feed into the docking step. Surface residues from the regions outside the binding are chosen as block residues to feed into the docking step. Attract and block constraints are set up in separate runs or in combination. Attract constraint provides positive weights to the chosen residues in computing the van der Waals energy during the docking step; block constraints prevent docking in these regions by providing a high positive value of the energy parameter in the docking step. Second, the binding regions of individual proteins identified by the footprinting are examined for the nature of the electrostatic surface pattern; based on these data, either electrostatic or desolvation free energy filter options are chosen in the docking program. This step retains the top 2000 energy-minimized conformations for the complex out of the 20,000 docked conformations that have the best shape-complementarity scores. For actin/GS1 docking, there is a low probability of finding the correct structure as top scoring when constraints in the docking step are not used or when the incorrect energy filter was used. However, when the correct energy filter was used, and attract, block, or a combination of both constraints are used, the models that ranked second, first, and first, respectively, were those that showed an interface structure similar to that of the crystal structure.

We also derived a model for G-actin/cofilin complex of unknown crystal or NMR structure based on footprinting data. A conformation consistent with footprinting data was top-scored with “attract” or “block” or “attract + block” constraints and with the correct energy filter. Observation of a substantial number of H-bonds (nine to 12 in number) at the interface is consistent with the reported charge-dependent binding of cofilin to G-actin (Pavlov et al. 2006). Our model also satisfies the findings of mutational and NMR data on the involvement of cofilin's charged residues Gln120, Asp123, and Glu126 (from β6-α4 loop) in G-actin binding.

Overall, the method outlined in Scheme 1 provides a rational and successful approach to generating and selecting the most accurate structural models when docking binary complexes. Similar to radiolytic footprinting data, data from hydrogen/deuterium exchange mass spectrometry (Truhlar et al. 2006) and cross-linking (Orlova et al. 2001) can also be used as constraints in the docking algorithms to derive accurate structural models. The current capabilities of the ClusPro server allow receptor molecule heavy atom (non H-atoms) sizes of 9700 and ligand molecule heavy atom sizes of 3700. Also, attract and block constraints are allowed only for the receptor molecule residues. However, improvements in this and other docking algorithms are likely in the near future (Tovchigrechko and Vakser 2006).

Materials and Methods

Radiolysis

Radiolysis experiments, in which protein samples are exposed to a white synchrotron X-ray beam for intervals from 0–200 ms are performed at the X-28C beamline of the National Synchrotron Light Source, Brookhaven National Laboratory (Upton, NY) as previously described (Guan et al. 2002, 2003). On these timescales, oxidative modifications dominate the chemistry compared with cross-linking events and cleavage (Davies and Dean 1997), and effects on the global protein structure are minimal.

Mass spectrometry

Radiolyzed samples are subjected to enzymatic proteolysis (Guan et al. 2002, 2003), and the resulting peptide mixtures are separated and analyzed using a coupled, high pressure LC (HPLC)–ESI-MS equipped with quadrupole ion trap (Waters 2690 microflow HPLC-Finnigan LCQ classic ESI-MS, Dionex nanoflow HPLC-Finnigan LCQ DecaXP nanospray MS). The oxidized probe sites within each peptide are confirmed by tandem mass spectrometry (MS/MS) (Guan and Chance 2005; Takamoto and Chance 2006). First-order rate constants of modification are derived from dose response curves (Guan and Chance 2005; Takamoto and Chance 2006). The detailed procedures have been previously described (Guan and Chance 2005; Takamoto and Chance 2006).

SASA calculation

The SASA of amino acids were calculated using the computer program GETAREA 1.1 (http://www.pauli.utmb.edu/cgi-bin/get_a_form.tcl) from the crystal structures with a probe radius of 1.4 Å with additional atomic parameter database and residue type library entries. SASA is defined as the surface mapped out by the center of the probe as if it were rolled around the van der Waals surface of the protein.

Homology modeling

Actin in the cofilin-bound form, precisely in the closed nucleotide cleft conformation, was obtained by homology modeling using the SWISS-MODEL protein modeling server (http://swissmodel.expasy.org/) (Schwede et al. 2003) inputting the rabbit actin sequence as “target” and the three-dimensional structure of bovine actin (2BTF chain A, 2.55 Å resolution) as “template.”

Protein–protein docking

Docking of ligand protein to receptor protein, energy filtering, clustering, and ranking were done using the ClusPro Web server (http://nrc.bu.edu/cluster) (Comeau et al. 2004a,b). Docking with DOT (Daughter of Turnip) program (Katchalski-Katzir et al. 1992; Ten Eyck et al. 1995), provided in the ClusPro, is executed using a surface complementarity fit of the rotating molecule (ligand protein, which is ABP in the present study) on the fixed molecule (receptor protein, which is actin in the present study) (Katchalski-Katzir et al. 1992), which is based on van der Waal's contact energy minimization. This employs a 128 × 128 × 128 Å grid with a spacing of 1 Å. The receptor protein is placed on the center of the 128 ų cube. Each heavy atom of the protein is surrounded by an inner repulsive layer of 1.5 Å and an outer attractive layer that extends 4 Å beyond the repulsive core. Technically, each heavy atom of the receptor protein is given an energy score of +1000 to within the repulsive layer and of −1.0 within the attractive layer. The ligand atoms are viewed as point charges with a value of +1.0 and cannot be modified. The goal, then, is to minimize the “energy” score by maximizing the number of attractive layer–ligand interactions while minimizing the number of repulsive layer–ligand interactions. Using a predefined list of 13,000 rotations of ligand protein within the 128 ų cube around the receptor protein, 2.7 × 10¹⁰ structures are evaluated, retaining the 20,000 structures with the best shape complementarity scores. These structures are then energy filtered (electrostatics or desolvation) to retain the top 2000 structures, which are then clustered (9 Å clustering radius) and ranked according to their cluster sizes. For details of the desolvation and electrostatics energy function, refer to Comeau et al. (2004a^,b). Clustering is accomplished based on the work of Shortle et al. (1998), where the native conformation is found to be that with the highest number of structural neighbors (e.g., highest cluster sizes).

Docking with experimental constraints

Constraints, namely, attract and block, can be set up within the docking protocol (Comeau et al. 2004a,b) individually in separate runs or in combination. The attract constraint provides positive weighting to docking at the experimentally derived interfacial residues of receptor protein, and the block constraint blocks the docking outside this interfacial area. Technically, when attracting to certain regions, the repulsive layer is decreased by 1 Å, leaving only a 0.5 Å repulsive core on the atom and changing the energy score of the attractive layer from −1.0 to −2.0. When blocking certain regions of the protein, the attractive layer of all selected atoms are removed by changing the value from −1.0 to 0.0. The details of these constraints are provided in the documentation of the ClusPro (Comeau et al. 2004a,b). Finally, the data from models are compared with biological and biochemical data to understand the functional implications of the model.

Electrostatic potential surface mapping and calculation of interface parameters

Electrostatic potential surfaces of proteins were mapped using DeepView. The RMSD of the actin binding protein (ABP) and of the interface of the models with the crystal structures or with other structural models were calculated using DeepView. Actin molecules are aligned at first using the option “fit molecules from selection,” followed by calculating the RMSD of the C_α atoms of the ligand proteins and of the interface residues, respectively. Interface SASA, percentage of polar atoms, hydrogen bonds/salt bridges, planarity, and GVI are derived using protein–protein interaction server V 1.5. (http://www.biochem.ucl.ac.uk/bsm/PP/server/) (Jones and Thornton 1996). Interface residues are all those residues that showed a SASA decrease of 1 Ų or more upon complexation (Jones and Thornton 1996). Planarity of the interface is defined as the RMSD of the atoms form the interface plane, which is the best fit plane through the three-dimensional co-ordinates of the atoms in the interface using principal component analysis (Jones and Thornton 1996). GVI is defined as the ratio of the volume of the gaps between the two interacting proteins at the interface as calculated using a program SURFNET, to the interface SASA. The GVI is a measure of the complementarity of the interacting surfaces.

Footprinting interface consistency score

FICS is the fraction of the total number of footprinting probe residues that are hypothesized to form the binding surface based on an independent analysis of the footprinting data compared (denominator) to the number that are actually found in the interface of the docked model (numerator). As such, this score can vary between zero (no common residues) and one (all residues match).