Flexible peptides that fold upon binding to another protein molecule mediate a large number of regulatory interactions in the living cell and may provide highly specific recognition modules. We present Rosetta FlexPepDock ab-initio, a protocol for simultaneous docking and de-novo folding of peptides, starting from an approximate specification of the peptide binding site. Using the Rosetta fragments library and a coarse-grained structural representation of the peptide and the receptor, FlexPepDock ab-initio samples efficiently and simultaneously the space of possible peptide backbone conformations and rigid-body orientations over the receptor surface of a given binding site. The subsequent all-atom refinement of the coarse-grained models includes full side-chain modeling of both the receptor and the peptide, resulting in high-resolution models in which key side-chain interactions are recapitulated. The protocol was applied to a benchmark in which peptides were modeled over receptors in either their bound backbone conformations or in their free, unbound form. Near-native peptide conformations were identified in 18/26 of the bound cases and 7/14 of the unbound cases. The protocol performs well on peptides from various classes of secondary structures, including coiled peptides with unusual turns and kinks. The results presented here significantly extend the scope of state-of-the-art methods for high-resolution peptide modeling, which can now be applied to a wide variety of peptide-protein interactions where no prior information about the peptide backbone conformation is available, enabling detailed structure-based studies and manipulation of those interactions.
Citation: Raveh B, London N, Zimmerman L, Schueler-Furman O (2011) Rosetta FlexPepDock ab-initio: Simultaneous Folding, Docking and Refinement of Peptides onto Their Receptors. PLoS ONE 6(4): e18934. https://doi.org/10.1371/journal.pone.0018934
Editor: Vladimir N. Uversky, University of South Florida College of Medicine, United States of America
Received: February 20, 2011; Accepted: March 12, 2011; Published: April 29, 2011
Copyright: © 2011 Raveh et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: This work has been supported by the Israel Science Foundation, funded by the Israel Academy of Science and Humanities ISF Grant 306/6, the USA-Israel Bi-national Science Foundation Grant 2009418 (to OSF), and in part by the 7th Framework Programme for Research of the European Commission, under FET-Open grant number 255827, CGL, Computational Geometry Learning (to BR). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Peptide-mediated interactions with globular proteins play a prominent role in signaling and regulatory networks of the living cell , . It has been estimated that between 15%-40% of all protein-protein interactions are mediated by flexible peptide that fold upon binding to a globular receptor . These peptides often form a modular binding motif, which can be embedded in intrinsically unstructured protein regions and within flexible loops, in order to confer desired interactions , , , .
Due to their cardinal role in regulatory mechanisms (e.g. ), flexible peptides are implicated in human disease and cancer , and therefore provide attractive leads for the design of inhibitory peptides and small molecule drugs , , , , , . A large-scale in-silico survey that we conducted recently suggests that peptides derived from globular proteins often have the potential to disrupt interactions of their origin domains by competitive inhibition , as was previously shown in experiment in several physiological interactions , . Hence, peptide molecules and their derivates hold great potential for targeted modulation of the cellular network of protein interactions.
Available structural models of peptide-protein interactions obtained by X-ray and NMR experiments have contributed significantly to our understanding of the mechanisms underlying key cellular interactions , , , , and enabled the structure-based redesign of both the peptide and receptor sequence at the binding site to inhibit specific cellular interactions altogether , . However, the number of available models solved in experiment represents only a small fraction of known peptide-protein interactions, while high-throughput methods for screening of peptide libraries such as peptide arrays  combinatorial phage-display , ,  and yeast surface display , ,  continue to produce an ever increasing flux of data about new peptide-mediated interactions whose structural basis is mostly poorly understood.
We have recently introduced Rosetta FlexPepDock , a protocol for the refinement of coarse models of peptide-protein complex structures. We benchmarked FlexPepDock thoroughly to define an effective basin of attraction of 5.5 Å RMSD from which the protocol can reliably recover near native peptide conformations. This covers a wide range of real world biological problems, for which an initial approximate structure is available. Indeed, we have used this protocol to model the structure of different peptide-protein interactions and to learn about their functional role (e.g. , , ). However, refinement is effective only if the approximate peptide backbone conformation within the receptor-binding site is given. Other methods dedicated to peptide docking have recently been developed but seem to be rather local as well , , , or restricted to very short peptides , .
In this study, we introduce Rosetta FlexPepDock ab-initio, which is designed to address the subset of problems where the approximate location of the peptide binding site is known, but no information about the peptide backbone conformation is available. The approximate binding site of the peptide can often be obtained from cross-linking experiments, mutational analysis, NMR shifts or any other experimental evidence , , and from computational predictions of increasing quality . FlexPepDock ab-initio borrows from existing Rosetta protocols, and attempts to ‘fold’ the peptide at the binding site, using fragment-based sampling to detect the overall conformation of the peptide in a reduced representation space (centroid mode) , coupled with efficient scanning of peptide orientations over the protein surface. This step is then followed by all-atom refinement of the peptide-protein conformation with fine backbone modeling and side-chain placement, based on the Rosetta FlexPepDock refinement protocol .
In the following sections, we present the Rosetta FlexPepDock ab-initio protocol, its usage and workflow and its performance on a selected benchmark of peptide-protein complexes. We anticipate that the new protocol will significantly contribute to the study of peptide-protein interactions, both for the purpose of basic research and for the increasing use of peptides for pharmaceutical applications.
Fig. 1 shows a schematic view of the FlexPepDock ab-initio protocol. Each step is described in more detail below.
Preliminary steps prior to running the protocol
The input to the protocol is an initial model of the peptide-protein complex. It is assumed that the receptor backbone is approximately correct, and that the peptide is initially positioned close to the correct binding site, albeit with arbitrary backbone conformation (the present study is based on results starting from extended peptide backbone conformations superimposed on a randomly selected anchor residue, but the protocol is designed to work from any arbitrary peptide starting conformation; see below and Table S3). Initial side-chain coordinates (such as the crystallographic side-chains of an unbound receptor) can be optionally provided as part of the input model, but are not required. In our analysis, we discarded the input side-chains from native complexes in the bound runs, but kept the side-chains from the free receptor structures in the unbound runs.
Preparation of fragment library for peptide.
For efficient sampling of the peptide backbone, we generate a library of trimer, pentamer and nonamer backbone fragments, which are extracted from solved protein structures in the Protein Data Bank (PDB at www.pdb.org ), using the protocol described in this issue by Gront et al. The library is constructed based on sequence similarity to the query peptide and on the secondary structure predicted for the peptide by PSIPRED , resulting in 500 fragments from each category of secondary structure type, i.e., α helix, extended β strand and coiled-coil loop (with a total of 1,500 fragments for a given query peptide). We note that although PSIPRED was not optimized for peptides, the resulting fragment libraries showed in practice good coverage of the peptide conformational space (in concordance with a recent report by Vanhee et al. that most peptide-protein interactions can be represented by fragment sets derived from single, non-related monomer structures ). We also note that fragments used in this analysis were not taken from the native peptide structure or any of its homologues (in fact, only fragments from globular monomer proteins were considered).
See the FlexPepDockAbInitio protocol capture archive:
Prepacking step -remove internal clashes in receptor.
A first preliminary step in our protocol involves the packing of the side-chains in each monomer to remove internal clashes that are not related to inter-molecular interactions, as described in Raveh et al. .
The pre-packing stage guarantees a uniform conformational background in non-interface regions, prior to molecular docking . We refer to the pre-packed input structure as the starting structure.
See the FlexPepDockAbInitio protocol capture archive:
$PATH_TO_EXE/FlexPepDocking.linuxgccrelease -database $PATH_TO_DB –s start.pdb -native native.pdb -ex1 -ex2aro -use_input_sc -unboundrot native.pdb -flexpep_prepack -nstruct 1
Step I: generation of models
We generate a large number of models beginning from the starting structure by repeating the procedure described below over multiple independent runs (which can be performed in parallel on a CPU cluster). In the present study we generated 50,000 models from each starting structure. More models can improve sampling and results further.
Step Ia – fast low-resolution modeling.
In a first step, the peptide is folded and docked over the surface of the receptor protein using a low-resolution representation of the complex, in which the side-chains are represented as unified spheres (Rosetta centroid mode ). The peptide is alternately folded and docked for 10 outer cycles. Each such outer-cycle consists of two internal Monte-Carlo simulations. The temperature term of the Metropolis criterion in the internal simulations is gradually decreased from 2.0 in the first outer cycle to 0.6 (arbitrary units) in the last outer cycle, such that large perturbations are favored in the first rounds. The inner simulations consist of: (1) Optimization of the rigid-body orientation: The peptide rigid-body orientation is optimized by a Monte-Carlo simulation consisting of 50 random rigid-body transformations (translational magnitude of 1 Å and rotational magnitude of 10° on average in each step). (2) Optimization of the peptide-backbone: The ab initio sampling protocol of the peptide backbone is performed based on moves described in detail in Rohl et al. . In short, the peptide backbone conformation is perturbed over 50 random Monte-Carlo moves while the peptide rigid body orientation remains fixed. Each move is selected randomly. In 60% of the moves, the φ/ψ torsion angles of random residues are perturbed using the so-called ‘small’ and ‘shear’ random moves described in Rohl et al. , to random magnitude, corresponding for the ‘small’ moves to random changes in φ/ψ angles. In the shear moves, the φ angle is rotated with equal magnitude but opposite direction relative to the preceding ψ angle, thereby reducing the perturbation to the rest of the chain. For both moves, perturbations to non-favorable Ramachandran angles are discriminated against, using a Metropolis criterion. In the remaining 40% of the moves, a trimer (30% of cases), pentamer (7.5% of cases) or a nonamer (2.5% of cases; for peptides with nine residues or more) fragment from the fragment library is inserted in a random position within the peptide.
Step Ib – refinement of low-resolution model.
The low-resolution modeling step results in a coarse-grained model of the peptide-protein complex. This model is further optimized using high-resolution refinement with the Rosetta FlexPepDock refinement protocol . This protocol was shown to be effective when the initial peptide conformation lies up to 5.5 Å from the native conformation. In brief, it consists of alternating optimization of the peptide rigid-body and backbone and orientation using the Monte-Carlo with Minimization approach  and a set of small-scale perturbations. To allow significant perturbations within the binding pocket while preventing the peptide and protein to separate during energy minimization, the refinement step begins with decreased and increased weights for the repulsive and attractive van der Waals term in the energy function, respectively. During refinement, these terms are gradually ramped back towards their original values.
See the FlexPepDockAbInitio protocol capture archive:
$PATH_TO_EXE/FlexPepDocking.linuxgccrelease -database $PATH_TO_DB
#If using multiple processes and no silent file:
#number of structures to produce
#for production run:
#fragment picker flags:
-flexPepDocking:frag5 frags/frags.5mers. offset
Step II: selection of models
In addition to sampling the conformational energy landscape efficiently, the challenge of modeling includes also the selection of the correct model among all the created models, which can be illustrated by the notion of finding a needle in a haystack. In order to do so, we first cluster our top-scoring models, and subsequently select top-scoring clusters as a model for the interaction. The top scoring 500 models are clustered using the Rosetta Cluster application, as described in Gray et al. , with a cluster radius cutoff of 2 Å peptide backbone atom RMSD. From each cluster, a representative model is subsequently selected according to the best energy score. The clusters are then ranked according to the energy of their representative models.
See the FlexPepDockAbInitio protocol capture archive:
scripts/clustering/cluster.sh pdb-id topXrms-radius scorefile reference-pdb models-silent-file score-type
Benchmarking of protocol
Datasets of peptide–protein interactions.
Since the FlexPepDock ab-initio simulations are time intensive, we selected a small but representative subset of complexes on which to assess the protocol. The bound dataset used in this study includes 26 peptide–protein complex structures (Table S1) chosen from the peptiDB dataset , a non-redundant set of high-resolution peptide–protein complex structures (below 70% sequence identity between receptor proteins; structures solved at resolution of 2 Å or better). These peptide-protein complexes represent a wide range of biological contexts. Besides a few interactions with a known and well-defined motif (PDB ids: 1SSH, 1W9E, 1Z90 and 2P1K), the interactions in this dataset were selected randomly. The length of peptides in the dataset varies between 5 and 13 amino acids, with up to 52 rotatable bonds. More details about these complexes can be found in Table S1.
For 14 out of the 26 complexes tested in this study, a high resolution (<2 Å) free receptor structure has been solved (or that of a protein with >90% sequence identity). The unbound structures were also extracted from the peptiDB dataset  and their interface residues were superimposed onto their bound counterparts as described in London et al. , to evaluate the difference between the free and bound receptor (Table S1). We note that in our benchmarking analysis, we discarded the input side-chains from native complexes only for docking to bound receptor structures, but not to free receptor structures (as this information is available in a real world scenario): including side-chains of unbound receptors was shown in our previous docking studies to improve protocol performance , .
Measure of success.
We define a docking model as near-native if the interface backbone atoms of the predicted peptide conformation deviate by ≤2 Å RMSD. For a docking simulation of a given interaction, we define successful sampling as the cases where a near-native model is sampled, and successful ranking as the cases where a near-native model is ranked among the ten lowest-energy clusters.
Extending the peptide from a random anchor.
For testing the protocol starting from an extended peptide conformation, a random peptide position was selected as an anchor (see Table S1 for details), and its coordinates were extracted from the native complex. The peptide's φ/ψ angles were then set to canonical ideal extended conformation (+135°/−135°, respectively), from which the docking simulations were initiated. We note that the random anchor was not specified, and the peptide was completely free to move during the simulation.
Random rigid-body perturbations.
In order to evaluate the robustness of FlexPepDock ab-initio, we repeated the run from a different orientation, created by perturbation of the rigid-body orientation of the extended peptide by random Gaussian translations and rotations of magnitudes 3 Å and 30°, respectively. Docking experiments were then initiated from the perturbed extended conformation.
The protocol and tests described in this manuscript follow the FlexPepDock protocol as implemented within revision 39664 of the Rosetta repository.
As for the original FlexPepDock refinement protocol, the FlexPepDock ab-initio protocol is fully implemented within the Rosetta modeling framework . Rosetta provides well-calibrated energy functions, efficient energy calculations and a battery of established conformational sampling protocols. In particular, we use the Rosetta library of protein fragments extracted from solved protein structures , the Monte-Carlo sampling with Energy Minimization first proposed by Li andScheraga , the Rosetta side-chain repacking protocol  and the Dunbrackrotamer library . For energy scoring, we use a modified version of the Rosetta full-atom energy function (Rosetta score12 , , see below) and the coarse-grained energy function, which employs a unified spheres side-chains model (Rosetta centroid score4 ).
Energy function used for model selection
In our original FlexPepDock refinement study  we used the standard Rosetta scoring function - score12 to rank and select the top-scoring models. However, several previous docking studies with Rosetta have indicated that the interface score, i.e. the energy score across the interface (where only atom-atom contacts between the partners are included in the energy evaluation) provides better estimates by removing effects outside from the actual interface. In addition, studies in our group on the prediction of binding specificity have shown that the score of the peptide (i.e. the internal peptide energy together with the interface energy) provides the best estimate of binding (unpublished data). We therefore investigated different energy terms for their ability to identify the near-native models among the set of created models. The total energy, interface energy, peptide energy, and a weighted sum of all were assessed (see Table S2).
See the FlexPepDockAbInitio protocol capture archive:
Results and Discussion
I. General outline of Rosetta FlexPepDock ab-initio protocol
The main components of our protocol for simultaneous ab-initio folding and docking of flexible peptides are outlined in Fig. 1. In short, after removing internal clashes of the receptor structure (by repacking), we place the peptide into the binding site (here as an extended conformation), and compile a corresponding library of peptide backbone fragments. We then apply FlexPepDock ab-initio, starting with low-resolution optimization of the peptide backbone conformation and its rigid-body orientation, followed by high-resolution refinement with full flexibility for all peptide and receptor side-chains. The high-resolution step corresponds to our previously developed Rosetta FlexPepDock refinement protocol . More details can be found in the Methods section.
II. Large-scale assessment of protocol performance
We assessed the ability of FlexPepDock ab-initio to sample and identify the correct conformation of peptides on a benchmark of 26 different peptide-protein interactions extracted from the PeptiDB dataset  (Table S1 and Methods). We created n = 50,000 conformations, starting from an extended peptide within the binding site. The resulting models were clustered and the clusters were ranked based on a reweighted version of the Rosetta generic full-atom energy score, in which interface and peptide residuesare given additional weight, and which improves the performance of the protocol compared to the standard Rosetta score (see Table S2 and Methods). Finally, we assessed how many runs succeeded to identify near-native models (defined as models with ≤2 Å peptide interface backbone RMSD in the top 10 scoring clusters; see Methods). The performance of FlexPepDock ab-initio is summarized in Table 1 (bound docking), Table 2 (unbound docking) and Fig. 2, and specific examples are presented in Fig. 3.
The FlexPepDock ab-initio protocol (Ab initio runs) samples a near-native structure in most of the Bound (A) and Unbound (B) simulations (height of bars), and in a significant fraction of the complexes a near-native structure is identified by the top-ranking cluster, or among the top-2 or top-10 (shaded parts of the bar; models clustered according to 2 Å peptide backbone RMSD cutoff). This significantly increases the scope FlexPepDock when compared to the original refinement protocol (Refinement runs), for both the Bound and Unbound simulations.
Bound (left panel) and Unbound (right panel) docking runs are shown for three successful and one failed simulation (see Table 1 for the full dataset): (A) 2O9V (Ponsin SH3 domain -paxillin proline rich region ); (B) 1N7F (GRIP1 PDZ6-liprin α c-terminal peptide complex ); (C) 2A3I (Mineralocorticoid ligand receptor domain - LXXLL motif of steroid receptor coactivator-1 (SRC1-4) ); and (D) 2FGR (Omp32-PAP peptide ). For 2O9V, 1N7F and 2A3I, a near-native model was ranked among the top 10 clusters for both the bound and unbound cases (the bound run of 2A3I is a borderline case, with 2.1A backbone interface RMSD). For each complex, the receptor is shown in either lavender or gray shade, for the Bound or Unbound form, respectively. The cartoonre presentation shows the peptide starting orientation (magenta), final model (red), and native structure (green). The corresponding energy landscape plots are shown for (A) (see Fig. S1 for additional plots for all examples). Each model created by FlexPepDock ab-initio is plotted as a red cross according to its peptide interface backbone RMSD (x-axis) and its energy score (y-axis; reweighted score; see Methods). The top 10 lowest energy clusters created from the top 500 scoring models are shown as blue circles. The black line indicates the maximal energy of the top 500 models. The energy plots show a sub-Ångstrom lowest-energy cluster for both Bound and Unbound simulations. The black circles in the left panel show results from the previous FlexPepDock refinement protocol , demonstrating how increased sampling allows the detection of conformations that have better energy scores and are also more similar to the native structure.
Simulations on the bound receptor conformation (Bound docking).
We first assessed the performance of the modeling protocol on the bound backbone conformation of the receptor. Successful modeling within this setup validates our strategy for sampling peptide backbone conformations and rigid-body orientations, and our energetic ranking ability within the setting of an accurate receptor backbone structure. This setting is a prerequisite for more realistic and challenging simulations described below. The results of the bound docking benchmark are summarized in Table 1 and Fig. 2A. In all but one case (25/26; 96%), a near-native model of the interaction (as defined above) was sampled by our protocol, and in most cases (18/26; 69%), the near-native model was also ranked within the ten top-ranking clusters (top-ranked in 12/26; 46%). In half of the cases (12/26) the backbone atoms of the top-ranking modeled peptides were located within only one Ångstrom RMSD of the native peptide backbone. This is remarkable, as the peptide backbones in the benchmark adopt diverse secondary structures and backbone conformations, including unusual kinks, turns and coils that are particularly hard to predict de-novo: The starting backbone configurations include challenging cases where the initial peptide backbone isup to 17 Å away from the native conformation, and the RMSD of the initially extended peptide backbone from the native exceeds 125° in φ/ψ torsion space.
Robustness of the FlexPepDock ab-initio protocol to changes in starting conformation.
In our evaluation described above, we started from an extended peptide initially positioned at the correct binding site, by aligning one of the peptide residues to its native coordinates. Even though the peptide was free to change its orientation and was not constrained in any way during the simulation, it is important to validate the robustness of our docking protocol to this initial orientation. Therefore we reassessed performance by repeating the simulation starting from a different conformation (see Methods). The results indicate that our protocol is indeed robust to the precise initial orientation of the extended peptide at the binding site, since the results did not differ significantly among repeated runs (Table S3). Even in the case where two simulations that from peptides oriented in opposite directions, they converge onto one final structure of the peptide-protein complex (see Fig. 4), leading the way for fully blind peptide docking.
The simulation of the structure of the CheZ-derived peptide bound to CheY (PDB-id 2FMF ) from two opposite starting orientations converge onto the same final conformation. (A) General view of the receptor structure (in gray; interface residues colored in light brown), the two initial extended peptide conformations (in cartoons), and the final helical conformation (in transparent cartoon). Peptide conformations are colored in rainbow. (B) Detailed atomic view of the top predictions from the two simulations (yellow and orange), highlighting the striking similarity of both predictions to the structure of the native peptide (green).
Comparison to peptide refinement protocol.
The FlexPepDock ab-initio protocol uses our previously reported refinement protocol  as a sub-module, but significantly extends its scope thanks to a preceding low-resolution peptide ab-initio sampling step. In particular, the new protocol does not assume any prior information about the peptide backbone conformation, whereas the original refinement protocol works well mainly when the initial peptide conformation is within 5.5 Å Cartesian- and 50° torsion-space RMSD of the native conformation . Indeed, the refinement protocol is able to sample a near-native conformation in only 10 of the 26 interactions (compared to 25/26 for theFlexPepDock ab-initio protocol described here; Fig. 2A), and these 10 are the easier cases where the native backbone conformation is already extended-like. Hence, the low-resolution stage of the ab-initio protocol is able to turn the harder cases into approximate models that can be refined to high-resolution.
Simulations on the free receptor conformation (Unbound docking).
In realistic scenarios, the bound receptor is obviously not part of the input when we try to dock a peptide to its protein receptor, and only the receptor in its free form can be used. The task of unbound docking is much more challenging, as the backbone conformation of the receptor protein may change upon binding, even though these changes are often very small for peptide-protein interactions . We repeated the previous test for all cases where a structure of free receptor structure was available (Table S1). Again, in nearly all cases at least one near-native model was sampled by our protocol (12/14; 86%), and in half of the cases this model was ranked among the ten top-ranking clusters (7/14; 50%) (see Table 2; Fig. 2B and right panels in Fig. 3), indicating that the presented protocol is well suited for de-novo folding and docking in many practical settings.
III. Partial success
Our protocol succeeded in modeling many of the complexes in our benchmarks, but also failed in some cases. Manual inspection of modeling failures revealed that in several of those, only part of the peptide has been modeled at high resolution, resulting in overall inaccurate models that do not meet the formal success criterion. In some cases, these partially accurate models may still be considered useful for practical applications. For instance, key motif residues may be modeled particularly well, (see our previous analysis in Raveh et al ), and other inaccuracies may be introduced by intrinsic motility of flexible peptide tails, and effects due to symmetry related contacts to the peptide that are not taken into account in our simulations (see below). Here we describe another case that involves partial conformational changes of the receptor upon binding of the peptide.
The interaction of the C-terminal region of FEN-1 with PCNA involves a considerable conformational change that redefines part of the receptor peptide binding site: upon peptide binding, an intermolecular β-sheet interface is created between the two partners (contributing to overall 1.5 Å RMSD for interface Cα atoms between the bound and free receptor conformation). This conformational change has been suggested to explain how PCNA stimulates FEN-1 activity. In the structure of PCNA bound toa FEN-1 C-terminal peptide (PDB id 1RXZ ), the peptide consists of two parts: a β-strand (residues 1–4) connected to a short 310 helix (residues 7–11; note that the 310 helix is defined as a turn by STRIDE ). The β-strand forms the intermolecular β sheet, while the helix interacts with a region in the receptor that does not change upon binding. Not surprisingly, while we succeed in modeling the peptide conformation on the bound receptor conformation at high accuracy (Table 1), the models based on the free receptor conformation are only partly accurate (Table 2; Fig. 5): they describe the helical part at high accuracy, including most of the side-chain atoms (1.9 Å all-atom RMSD over the five C-terminal peptide residues for one of the top-10 models; Fig. 5), while the strand is not formed due to the lack of the corresponding partner strand in the receptor. This case demonstrates that even though the overall quality of the peptide model does not pass our ‘formal’ threshold for success, there is still a substructure that is modeled very accurately. Within the context of blind docking of the peptide conformation onto a free receptor conformation, this substructure could still serve as a useful starting point for subsequent peptide-based design and manipulation of the interaction.
Partial accuracy due to a conformational change in the receptor in the interaction of the C-terminal region of FEN-1 with PCNA (PDB id 1RXZ ). (A) Structural view of this interaction in cartoon view: The free and bound conformations of the receptor are shown on the left in blue and gray ribbon, respectively. The native peptide is shown in green cartoon, and two of the top-scoring models are shown in red and pink. The C-terminal region of the receptor (highlighted) undergoes a conformational change upon binding to create a β-sheet pairing with the native peptide (green cartoon). Therefore, while the helical region of the peptide (residues 7–11, see inset) is well modeled in top-ranking models, the strand-part (residues 1–4) is not accurate. The inset on the right highlights the accurate recapitulation of certain atomic details in the helical peptide part. (B) Energy plots of this interaction: Left panel: The plot of the full peptide demonstrates that no near-native conformations are sampled or selected when considering the entire peptide sequence (peptide interface backbone RMSD, x-axis, vs. energy score, y-axis; same depiction as in Fig. 3). Right panel: The plot for the best pentamer substructure of the peptide demonstrates high accuracy, which can be attributed to the helical part of the peptide (note that this plot shows allatom RMSD).
IV. Identification of new challenges
(1) Cases of accurate bound but inaccurate unbound docking highlight importance of receptor flexibility for peptide-protein interactions.
While our protocol performs well in half of the unbound cases, examples like 1RXZ (described above and in Fig. 5) demonstrate the importance of including receptor flexibility in peptide docking, which is an ongoing work in our group. Even though the conformational changes of the receptor upon binding of peptides are usually fairly small (Table 2 and ), 5/14 peptide-protein complexes were modeled and ranked accurately only based on the bound receptor conformation, but not based on the free receptor conformation. These cases will be analyzed in more detail in a follow-up study and used to calibrate a protocol that includes receptor backbone flexibility.
(2) Cases of inaccurate bound docking might be due to crystal contacts that define the peptide conformation.
We observed that in several structures (in particular where we fail in the simulations starting from the bound receptor structure), crystal symmetry operations reveal additional contacts to the peptide that are contributed by symmetry related molecules. These can represent a considerable fraction of the overall contacts that the peptide forms with its surrounding, and thus influence the structure that the peptide will adopt. Since we do not account for those contacts, our simulations might fail to accurately model, or select, peptide models that resemble the native crystal structure (e.g. we completely fail to identify the correct conformation for PDB-id 2J6F ). This intriguing finding suggests that solved peptide-protein complex structures might sometimes actually represent non-biological conformations. We are currently investigating in more detail how prevalent the influence of symmetry-related molecules on the peptide conformation is, and to what degree it affects our ability to identify and accurately model the key features in the peptide-protein interface of biological interactions.
(3) Towards blind peptide docking – integration with binding site prediction tools.
We demonstrated that the present protocol is robust to the precise starting orientation of the peptide near the binding site (Fig. 4 and Table S3). Therefore, we foresee that it can be integrated with emerging techniques for identifying ligand and peptide binding sites, based on the chemical and the statistical features that characterize these sites (e.g. , , ). This would enable blind docking of peptides without any prior knowledge about either the binding site location or the peptide backbone conformation. We anticipate that such a tool will be applicable on a proteome-wide scale, and are working towards this direction.
We have presented a Rosetta protocol to efficiently model the structure of a peptide bound to its receptor, using an optimization scheme that involves simultaneous full ab initio sampling of the peptide backbone conformation and its orientation on the receptor protein. This computationally intensive protocol samples a considerable conformation space, and consequently is able to identify near-native models within the top-ranking clusters for many challenging cases. These candidate structures provide an excellent starting point for the subsequent characterization and modulation of a peptide-mediated interaction: the atomic details of the interaction are revealed, and further refinement with the FlexPepDock refinement protocol described previously  can identify peptide residues that contribute significantly to binding affinity and specificity. This will significantly increase the number of peptide-mediated interactions that can be accurately characterized and manipulated.
Energy landscape plots for models created by the FlexPepDockab initio protocol. Energy plots for Bound (left panel) and Unbound (right panel) docking runs are shown for the three successful and one failed simulations shown in Fig. 3 (see Table 1 for the full dataset). From top to bottom: 2O9V; 1N7F; 2A3I; and 2FGR. See Legend to Fig. 3 for more details.
The benchmark of peptide-protein interactions used in this study.
Number of successful predictions based on different scoring functions. Reweighted score (blue) performs best. Results refer to detection of top-scoring models, prior to clustering (see Methods for more details).
The FlexPepDockab-initio protocol is robust to changes in starting conformation. Similar results are obtained for two repeats of the protocol fromdistinct starting structures. The two starting structures are the extended conformation reported in Table 1 (in italics and parentheses), and an initial peptide orientation obtained by random translation and rotation of 3A and 30°, respectively).
We thank the members of the Furman lab and the Rosetta developer community for stimulating discussions.
Conceived and designed the experiments: BR NL OS-F LZ. Performed the experiments: BR LZ NL. Analyzed the data: BR NL LZ OS-F. Wrote the paper: BR OS-F NL.
- 1. Pawson T, Nash P (2003) Assembly of cell regulatory systems through protein interaction domains. Science 300: 445–452.T. PawsonP. Nash2003Assembly of cell regulatory systems through protein interaction domains.Science300445452
- 2. Petsalaki E, Russell RB (2008) Peptide-mediated interactions in biological systems: new discoveries and applications. Curr Opin Biotechnol 19: 344–350.E. PetsalakiRB Russell2008Peptide-mediated interactions in biological systems: new discoveries and applications.Curr Opin Biotechnol19344350
- 3. Wright PE, Dyson HJ (2009) Linking folding and binding. Curr Opin Struct Biol 19: 31–38.PE WrightHJ Dyson2009Linking folding and binding.Curr Opin Struct Biol193138
- 4. Neduva V, Linding R, Su-Angrand I, Stark A, de Masi F, et al. (2005) Systematic discovery of new recognition peptides mediating protein interaction networks. PLoS Biol 3: e405.V. NeduvaR. LindingI. Su-AngrandA. StarkF. de Masi2005Systematic discovery of new recognition peptides mediating protein interaction networks.PLoS Biol3e405
- 5. Vacic V, Oldfield CJ, Mohan A, Radivojac P, Cortese MS, et al. (2007) Characterization of molecular recognition features, MoRFs, and their binding partners. J Proteome Res 6: 2351–2366.V. VacicCJ OldfieldA. MohanP. RadivojacMS Cortese2007Characterization of molecular recognition features, MoRFs, and their binding partners.J Proteome Res623512366
- 6. Kondo T, Plaza S, Zanet J, Benrabah E, Valenti P, et al. (2010) Small peptides switch the transcriptional activity of Shavenbaby during Drosophila embryogenesis. Science 329: 336–339.T. KondoS. PlazaJ. ZanetE. BenrabahP. Valenti2010Small peptides switch the transcriptional activity of Shavenbaby during Drosophila embryogenesis.Science329336339
- 7. Vanhee P, van der Sloot AM, Verschueren E, Serrano L, Rousseau F, et al. (2011) Computational design of peptide ligands. To appear in Trends Biotechnol 29: 231–239.P. VanheeAM van der SlootE. VerschuerenL. SerranoF. Rousseau2011Computational design of peptide ligands.To appear in Trends Biotechnol29231239
- 8. Audie J, Boyd C (2010) The Synergistic Use of Computation, Chemistry and Biology to Discover Novel Peptide-Based Drugs: The Time is Right. Current Pharmaceutical Design 16: 567–582.J. AudieC. Boyd2010The Synergistic Use of Computation, Chemistry and Biology to Discover Novel Peptide-Based Drugs: The Time is Right.Current Pharmaceutical Design16567582
- 9. London N, Movshovitz-Attias D, Schueler-Furman O (2010) The structural basis of peptide-protein binding strategies. Structure 18: 188–199.N. LondonD. Movshovitz-AttiasO. Schueler-Furman2010The structural basis of peptide-protein binding strategies.Structure18188199
- 10. Vlieghe P, Lisowski V, Martinez J, Khrestchatisky M (2010) Synthetic therapeutic peptides: science and market. Drug Discov Today 15: 40–56.P. VliegheV. LisowskiJ. MartinezM. Khrestchatisky2010Synthetic therapeutic peptides: science and market.Drug Discov Today154056
- 11. Rubinstein M, Niv MY (2009) Peptidic modulators of protein-protein interactions: progress and challenges in computational design. Biopolymers 91: 505–513.M. RubinsteinMY Niv2009Peptidic modulators of protein-protein interactions: progress and challenges in computational design.Biopolymers91505513
- 12. Parthasarathi L, Casey F, Stein A, Aloy P, Shields DC (2008) Approved drug mimics of short peptide ligands from protein interaction motifs. J Chem Inf Model 48: 1943–1948.L. ParthasarathiF. CaseyA. SteinP. AloyDC Shields2008Approved drug mimics of short peptide ligands from protein interaction motifs.J Chem Inf Model4819431948
- 13. London N, Raveh B, Movshovitz-Attias D, Schueler-Furman O (2010) Can Self-Inhibitory Peptides be Derived from the Interfaces of Globular Protein-Protein Interactions? Proteins 78: 3140–3149.N. LondonB. RavehD. Movshovitz-AttiasO. Schueler-Furman2010Can Self-Inhibitory Peptides be Derived from the Interfaces of Globular Protein-Protein Interactions?Proteins7831403149
- 14. Burnier M (2001) Angiotensin II type 1 receptor blockers. Circulation 103: 904–912.M. Burnier2001Angiotensin II type 1 receptor blockers.Circulation103904912
- 15. Hashemzadeh M, Furukawa M, Goldsberry S, Movahed MR (2008) Chemical structures and mode of action of intravenous glycoprotein IIb/IIIa receptor blockers: A review. Exp Clin Cardiol 13: 192–197.M. HashemzadehM. FurukawaS. GoldsberryMR Movahed2008Chemical structures and mode of action of intravenous glycoprotein IIb/IIIa receptor blockers: A review.Exp Clin Cardiol13192197
- 16. Vanhee P, Reumers J, Stricher F, Baeten L, Serrano L, et al. (2010) PepX: a structural database of non-redundant protein-peptide complexes. Nucleic Acids Res 38: D545–551.P. VanheeJ. ReumersF. StricherL. BaetenL. Serrano2010PepX: a structural database of non-redundant protein-peptide complexes.Nucleic Acids Res38D545551
- 17. Vanhee P, Stricher F, Baeten L, Verschueren E, Lenaerts T, et al. (2009) Protein-peptide interactions adopt the same structural motifs as monomeric protein folds. Structure 17: 1128–1136.P. VanheeF. StricherL. BaetenE. VerschuerenT. Lenaerts2009Protein-peptide interactions adopt the same structural motifs as monomeric protein folds.Structure1711281136
- 18. Stein A, Aloy P (2008) Contextual specificity in peptide-mediated protein interactions. PLoS One 3: e2524.A. SteinP. Aloy2008Contextual specificity in peptide-mediated protein interactions.PLoS One3e2524
- 19. Katz C, Levy-Beladev L, Rotem-Bamberger S, Rito T, Rudiger SG, et al. (2011) Studying protein-protein interactions using peptide arrays. C. KatzL. Levy-BeladevS. Rotem-BambergerT. RitoSG Rudiger2011Studying protein-protein interactions using peptide arrays.In press: ChemSoc Rev. In press: ChemSoc Rev.
- 20. Molek P, Strukelj B, Bratkovic T (2011) Peptide phage display as a tool for drug discovery: targeting membrane receptors. Molecules 16: 857–887.P. MolekB. StrukeljT. Bratkovic2011Peptide phage display as a tool for drug discovery: targeting membrane receptors.Molecules16857887
- 21. Ladner RC, Sato AK, Gorzelany J, de Souza M (2004) Phage display-derived peptides as therapeutic alternatives to antibodies. Drug Discov Today 9: 525–529.RC LadnerAK SatoJ. GorzelanyM. de Souza2004Phage display-derived peptides as therapeutic alternatives to antibodies.Drug Discov Today9525529
- 22. Sidhu SS, Lowman HB, Cunningham BC, Wells JA (2000) Phage display for selection of novel binding peptides. Methods Enzymol 328: 333–363.SS SidhuHB LowmanBC CunninghamJA Wells2000Phage display for selection of novel binding peptides.Methods Enzymol328333363
- 23. Gai SA, Wittrup KD (2007) Yeast surface display for protein engineering and characterization. Curr Opin Struct Biol 17: 467–473.SA GaiKD Wittrup2007Yeast surface display for protein engineering and characterization.Curr Opin Struct Biol17467473
- 24. Dutta S, Gulla S, Chen TS, Fire E, Grant RA, et al. (2010) Determinants of BH3 binding specificity for Mcl-1 versus Bcl-xL. J Mol Biol 398: 747–762.S. DuttaS. GullaTS ChenE. FireRA Grant2010Determinants of BH3 binding specificity for Mcl-1 versus Bcl-xL.J Mol Biol398747762
- 25. Fowler DM, Araya CL, Fleishman SJ, Kellogg EH, Stephany JJ, et al. (2010) High-resolution mapping of protein sequence-function relationships. Nat Methods 7: 741–746.DM FowlerCL ArayaSJ FleishmanEH KelloggJJ Stephany2010High-resolution mapping of protein sequence-function relationships.Nat Methods7741746
- 26. Raveh B, London N, Schueler-Furman O (2010) Sub-angstrom modeling of complexes between flexible peptides and globular proteins. Proteins 78: 2029–2040.B. RavehN. LondonO. Schueler-Furman2010Sub-angstrom modeling of complexes between flexible peptides and globular proteins.Proteins7820292040
- 27. Crawley SW, Samimi Gharaei M, Ye Q, Yang Y, Raveh B, et al. (2010) Autophosphorylation activates Dictyostelium myosin II heavy chain kinase A by providing a ligand for an allosteric binding site in the α-kinase domain. Journal of Biological Chemistry 286: 2607–2616.SW CrawleyM. Samimi GharaeiQ. YeY. YangB. Raveh2010Autophosphorylation activates Dictyostelium myosin II heavy chain kinase A by providing a ligand for an allosteric binding site in the α-kinase domain.Journal of Biological Chemistry28626072616
- 28. Buch I, Fishelovitch D, London N, Raveh B, Wolfson HJ, et al. (2010) Allosteric regulation of glycogen synthase kinase 3beta: a theoretical study. Biochemistry 49: 10890–10901.I. BuchD. FishelovitchN. LondonB. RavehHJ Wolfson2010Allosteric regulation of glycogen synthase kinase 3beta: a theoretical study.Biochemistry491089010901
- 29. Belitsky M, Avshalom H, Yelin I, London N, Sperber M, et al. (2011) The Escherichia coli extracellular death factor EDF induces the endoribonucleolytic activity of MazF. To appear in Molecular Cell 41: 625–635.M. BelitskyH. AvshalomI. YelinN. LondonM. Sperber2011The Escherichia coli extracellular death factor EDF induces the endoribonucleolytic activity of MazF.To appear in Molecular Cell41625635
- 30. Antes I (2010) DynaDock: A new molecular dynamics-based algorithm for protein-peptide docking including receptor flexibility. Proteins 78: 1084–1104.I. Antes2010DynaDock: A new molecular dynamics-based algorithm for protein-peptide docking including receptor flexibility.Proteins7810841104
- 31. Niv MY, Weinstein H (2005) A flexible docking procedure for the exploration of peptide binding selectivity to known structures and homology models of PDZ domains. J Am Chem Soc 127: 14072–14079.MY NivH. Weinstein2005A flexible docking procedure for the exploration of peptide binding selectivity to known structures and homology models of PDZ domains.J Am Chem Soc1271407214079
- 32. Prasad PA, Gautham N (2008) A new peptide docking strategy using a mean field technique with mutually orthogonal Latin square sampling. Journal of Computer-Aided Molecular Design 22: 815–829.PA PrasadN. Gautham2008A new peptide docking strategy using a mean field technique with mutually orthogonal Latin square sampling.Journal of Computer-Aided Molecular Design22815829
- 33. Abe K, Kobayashi N, Sode K, Ikebukuro K (2007) Peptide ligand screening of alpha-synuclein aggregation modulators by in silico panning. BMC Bioinformatics 8: 451.K. AbeN. KobayashiK. SodeK. Ikebukuro2007Peptide ligand screening of alpha-synuclein aggregation modulators by in silico panning.BMC Bioinformatics8451
- 34. Hetenyi C, van der Spoel D (2002) Efficient docking of peptides to proteins without prior knowledge of the binding site. Protein Sci 11: 1729–1737.C. HetenyiD. van der Spoel2002Efficient docking of peptides to proteins without prior knowledge of the binding site.Protein Sci1117291737
- 35. Morrison KL, Weiss GA (2001) Combinatorial alanine-scanning. Curr Opin Chem Biol 5: 302–307.KL MorrisonGA Weiss2001Combinatorial alanine-scanning.Curr Opin Chem Biol5302307
- 36. Mandell JG, Falick AM, Komives EA (1998) Identification of protein-protein interfaces by decreased amide proton solvent accessibility. Proc Natl Acad Sci U S A 95: 14705–14710.JG MandellAM FalickEA Komives1998Identification of protein-protein interfaces by decreased amide proton solvent accessibility.Proc Natl Acad Sci U S A951470514710
- 37. Petsalaki E, Stark A, Garcia-Urdiales E, Russell RB (2009) Accurate prediction of peptide binding sites on protein surfaces. PLoS Comput Biol 5: e1000335.E. PetsalakiA. StarkE. Garcia-UrdialesRB Russell2009Accurate prediction of peptide binding sites on protein surfaces.PLoS Comput Biol5e1000335
- 38. Rohl CA, Strauss CE, Misura KM, Baker D (2004) Protein structure prediction using Rosetta. Methods Enzymol 383: 66–93.CA RohlCE StraussKM MisuraD. Baker2004Protein structure prediction using Rosetta.Methods Enzymol3836693
- 39. Berman HM, Westbrook J, Feng Z, Gilliland G, Bhat TN, et al. (2000) The Protein Data Bank. Nucleic Acids Res 28: 235–242.HM BermanJ. WestbrookZ. FengG. GillilandTN Bhat2000The Protein Data Bank.Nucleic Acids Res28235242
- 40. Jones DT (1999) Protein secondary structure prediction based on position-specific scoring matrices. J Mol Biol 292: 195–202.DT Jones1999Protein secondary structure prediction based on position-specific scoring matrices.J Mol Biol292195202
- 41. Gray JJ, Moughon S, Wang C, Schueler-Furman O, Kuhlman B, et al. (2003) Protein-protein docking with simultaneous optimization of rigid-body displacement and side-chain conformations. J Mol Biol 331: 281–299.JJ GrayS. MoughonC. WangO. Schueler-FurmanB. Kuhlman2003Protein-protein docking with simultaneous optimization of rigid-body displacement and side-chain conformations.J Mol Biol331281299
- 42. Li Z, Scheraga HA (1987) Monte Carlo-minimization approach to the multiple-minima problem in protein folding. Proc Natl Acad Sci U S A 84: 6611–6615.Z. LiHA Scheraga1987Monte Carlo-minimization approach to the multiple-minima problem in protein folding.Proc Natl Acad Sci U S A8466116615
- 43. Wang C, Schueler-Furman O, Baker D (2005) Improved side-chain modeling for protein-protein docking. Protein Sci 14: 1328–1339.C. WangO. Schueler-FurmanD. Baker2005Improved side-chain modeling for protein-protein docking.Protein Sci1413281339
- 44. Das R, Baker D (2008) Macromolecular modeling with rosetta. Annu Rev Biochem 77: 363–382.R. DasD. Baker2008Macromolecular modeling with rosetta.Annu Rev Biochem77363382
- 45. Kuhlman B, Baker D (2000) Native protein sequences are close to optimal for their structures. Proc Natl Acad Sci U S A 97: 10383–10388.B. KuhlmanD. Baker2000Native protein sequences are close to optimal for their structures.Proc Natl Acad Sci U S A971038310388
- 46. Dunbrack RL Jr, Cohen FE (1997) Bayesian statistical analysis of protein side-chain rotamer preferences. Protein Sci 6: 1661–1681.RL Dunbrack JrFE Cohen1997Bayesian statistical analysis of protein side-chain rotamer preferences.Protein Sci616611681
- 47. Chapados BR, Hosfield DJ, Han S, Qiu J, Yelent B, et al. (2004) Structural basis for FEN-1 substrate specificity and PCNA-mediated activation in DNA replication and repair. Cell 116: 39–50.BR ChapadosDJ HosfieldS. HanJ. QiuB. Yelent2004Structural basis for FEN-1 substrate specificity and PCNA-mediated activation in DNA replication and repair.Cell1163950
- 48. Frishman D, Argos P (1995) Knowledge-based protein secondary structure assignment. Proteins 23: 566–579.D. FrishmanP. Argos1995Knowledge-based protein secondary structure assignment.Proteins23566579
- 49. Moncalian G, Cardenes N, Deribe YL, Spinola-Amilibia M, Dikic I, et al. (2006) Atypical polyproline recognition by the CMS N-terminal Src homology 3 domain. J Biol Chem 281: 38845–38853.G. MoncalianN. CardenesYL DeribeM. Spinola-AmilibiaI. Dikic2006Atypical polyproline recognition by the CMS N-terminal Src homology 3 domain.J Biol Chem2813884538853
- 50. Brenke R, Kozakov D, Chuang GY, Beglov D, Hall D, et al. (2009) Fragment-based identification of druggable ‘hot spots’ of proteins using Fourier domain correlation techniques. Bioinformatics 25: 621–627.R. BrenkeD. KozakovGY ChuangD. BeglovD. Hall2009Fragment-based identification of druggable ‘hot spots’ of proteins using Fourier domain correlation techniques.Bioinformatics25621627
- 51. Gehmlich K, Pinotsis N, Hayess K, van der Ven PF, Milting H, et al. (2007) Paxillin and ponsin interact in nascent costameres of muscle cells. J Mol Biol 369: 665–682.K. GehmlichN. PinotsisK. HayessPF van der VenH. Milting2007Paxillin and ponsin interact in nascent costameres of muscle cells.J Mol Biol369665682
- 52. Im YJ, Park SH, Rho SH, Lee JH, Kang GB, et al. (2003) Crystal structure of GRIP1 PDZ6-peptide complex reveals the structural basis for class II PDZ target recognition and PDZ domain-mediated multimerization. J Biol Chem 278: 8501–8507.YJ ImSH ParkSH RhoJH LeeGB Kang2003Crystal structure of GRIP1 PDZ6-peptide complex reveals the structural basis for class II PDZ target recognition and PDZ domain-mediated multimerization.J Biol Chem27885018507
- 53. Li Y, Choi M, Cavey G, Daugherty J, Suino K, et al. (2005) Crystallographic identification and functional characterization of phospholipids as ligands for the orphan nuclear receptor steroidogenic factor-1. Mol Cell 17: 491–502.Y. LiM. ChoiG. CaveyJ. DaughertyK. Suino2005Crystallographic identification and functional characterization of phospholipids as ligands for the orphan nuclear receptor steroidogenic factor-1.Mol Cell17491502
- 54. Zachariae U, Kluhspies T, De S, Engelhardt H, Zeth K (2006) High resolution crystal structures and molecular dynamics studies reveal substrate binding in the porin Omp32. J Biol Chem 281: 7413–7420.U. ZachariaeT. KluhspiesS. DeH. EngelhardtK. Zeth2006High resolution crystal structures and molecular dynamics studies reveal substrate binding in the porin Omp32.J Biol Chem28174137420
- 55. Guhaniyogi J, Robinson VL, Stock AM (2006) Crystal structures of beryllium fluoride-free and beryllium fluoride-bound CheY in complex with the conserved C-terminal peptide of CheZ reveal dual binding modes specific to CheY conformation. J Mol Biol 359: 624–645.J. GuhaniyogiVL RobinsonAM Stock2006Crystal structures of beryllium fluoride-free and beryllium fluoride-bound CheY in complex with the conserved C-terminal peptide of CheZ reveal dual binding modes specific to CheY conformation.J Mol Biol359624645