Modeling of protein-peptide interactions using the CABS-dock web server for binding site search and flexible docking

1. INTRODUCTION

• Protein-peptide interactions contribute to signaling, localization, immune response, and protein degradation, and some are implicated in cancer and autoimmune disease.Up to 40% of protein-protein interactions in higher eukaryotes are mediated by peptides.Source: §1 · p. 2
• Transient interactions, significantly weaker affinity than other PPIs, and peptide flexibility make structural characterization challenging.The paper identifies computational approaches such as molecular docking as complementary methods for this problem.Source: §1 · p. 2
• Common workflows identify a binding site before local peptide docking and may add high-resolution refinement, while combined search methods often yield unsatisfactory peptide models.Sequence-pattern-only binding-site prediction can produce many false positives.Source: §1 · p. 2
• CABS-dock had previously produced high- or medium-accuracy models for over 80% of bound and unbound benchmark cases.The cited benchmark was described as the largest non-redundant protein-peptide interaction set available at the time.Source: §1 · p. 3
• CABS-dock requires neither a known binding site nor peptide conformation and allows full peptide flexibility and significant receptor conformational changes.Its stated disadvantages include no option to guide docking with known binding-site information and the possibility that the 10 best-scored models omit high-accuracy models present in the larger set.Source: §1 · p. 3
• This work evaluates CABS-dock through default and advanced-option examples and explores external scoring using reconstructed, locally optimized models and short explicit-solvent all-atom molecular-dynamics simulations.The paper also provides a tutorial for displaying and analyzing models and trajectories with VMD.Source: §1 · p. 3

2. METHOD

• CABS represents residues with CA, B, and S pseudo-atoms and uses knowledge-based potentials, lattice-restricted CA coordinates, asymmetric Metropolis sampling, and Replica Exchange.The CABS model had also been applied to protein structure prediction, dynamics, and interaction modeling.Source: §2 · pp. 3–4
• The automated server accepts a receptor 3D structure and peptide sequence, then returns 10 top-scored models through flexible docking, filtering, clustering, and reconstruction.The pipeline's generated model sets are also available for download.Source: §2 · p. 4, Figure 1 · p. 5

2.1 Flexible docking with the CABS model

• Ten random peptide starts undergo Replica Exchange Monte Carlo within 20 Å of the receptor, yielding 10,000 models with full peptide and small receptor-backbone flexibility.The simulation produces 10 trajectories of 1,000 models each.Source: §2.1 · p. 5, Figure 2 · p. 6

2.2 Initial filtering

• Filtering rejects unbound states and selects up to 100 models per trajectory by lowest peptide-receptor interaction energy, retaining up to 1,000 models.Source: §2.2 · p. 6

2.3 Clustering

• The 1,000 filtered models are clustered by repeated k-medoid runs with k=10; consensus medoids are ranked by cluster density using ligand-RMSD to differentiate cluster elements.Source: §2.3 · p. 7

2.4 Reconstruction and local optimization of the final models

• The 10 medoids are reconstructed with MODELLER and optimized using the DOPE potential to produce energy-minimized all-atom models.Source: §2.4 · p. 7

3.1 Docking without prior knowledge of the binding site

• Default CABS-dock yielded high- or medium-accuracy models for over 80% of bound and unbound benchmark cases.Accuracy was assessed by ligand-RMSD, and the benchmark comprised 103 bound and 68 unbound cases.Source: §3.1 · p. 8, Figure 4 · p. 8

3.1.1 An example when the accurate model is top ranked

• In default-mode examples, a top-ranked pose had ligand-RMSD 2.22 Å, while another case’s top 10 included a 4.16 Å pose and its 10,000-model trajectory contained a 2.33 Å pose.The latter case shows that high-accuracy models may be absent from the top 10 yet present among 10,000 trajectory models.Source: Figure 5 · p. 9, §3.1.1 · p. 11, Figure 7 · p. 12

3.2.1 Increasing the flexibility of receptor fragments

• Giving full flexibility to the 10-residue loop (residues 45–54) uncovered the obstructed binding site and produced a first-ranked high-accuracy model at 2.03 Å.Moderate flexibility weakens the default distance restraints, whereas full flexibility removes them for selected residues.Source: §3.2.1 · p. 13, Figure 8 · p. 14

3.2.2 Excluding binding modes from docking search

• Excluding initial poses and resubmitting the job shifted predictions to the native site, where the first-ranked model had a ligand-RMSD of 2.89 Å.Users can mark residues as unlikely to bind or exclude selected models from a previous run to search for additional sites.Source: §3.2.2 · pp. 14–15, Figure 9 · p. 15

3.3.1 Accessing CABS-dock models in coarse-grained representation

• The downloadable coarse-grained archive includes trajectories, top 1,000 models, cluster representatives, and final models; its model sets can be reconstructed to all-atom structures.It contains 10 trajectories of 1,000 models each, and reconstructed structures can support all-atom analyses and visualization.Source: §3.3.1 · pp. 15–16

3.3.2 Strategies for reconstruction and local optimization

• Model reconstruction adds backbone atoms before side chains and requires local optimization; the server implements combined reconstruction and optimization for multichain complexes using Modeller and DOPE minimization.Backbone reconstruction should tolerate small local C-alpha distortions, and SCWRL is recommended for side-chain reconstruction.Source: §3.3.2 · p. 16

3.3.3 Database of large sets of CABS-dock models in an all-atom representation

• All-atom benchmark sets containing 100 or 1,000 top-scored models per case were rebuilt for 103 bound and 68 unbound complexes and made available for download.The sets support all-atom scoring of protein-peptide interactions and development of scoring potentials.Source: §3.3.3 · pp. 16–17

3.4.1 Results of scoring exercises using all-atom MD with explicit solvent

• Across five unbound cases, all-atom force fields generally scored better than CABS-dock; GROMOS43a1 performed better in two cases, comparably in one, and worse in two.For 1N83, all-atom force fields were more effective than the coarse-grained force field at distinguishing the lowest ligand-RMSD models from near-native models; GROMOS and OPLS ranked the best-quality model among their 10 top-scored models.Source: §3.4.1 · pp. 17–18, Table 3 · p. 18, Figure 10 · p. 19

3.4.2 Details of the all-atom MD scoring procedure

• The scoring exercises used OPLS, CHARMM, AMBER99SB, and GROMOS41a, with SPC or TIP3P water models.The passage assigns SPC to GROMOS43a1, AMBER99SB, and OPLS, and TIP3P to CHARMM; it advises checking different schemes because results may depend on the force field and water model.Source: §3.4.2 · p. 19
• Production simulations ran for 20 ps after minimization and two 10-ps equilibration runs at 300 K; ligand restraints were released while the receptor remained restrained.The equilibration stages were NVT followed by NPT, with receptor and ligand restraints during both stages.Source: §3.4.2 · p. 19
• The score was the receptor–ligand Coulomb and Lennard-Jones interaction-energy sum, averaged over saved configurations; one 20-ps model run required about one CPU hour.Coordinates were saved every 0.2 ps; parallel runs could reach speeds up to 5–6 times higher.Source: §3.4.2 · p. 20

4. SUMMARY

• The demonstrated capabilities include efficient docking of fully flexible peptides to flexible proteins without prior knowledge of the binding site.The paper presents these capabilities as making CABS-dock a promising approach to protein-peptide interaction modeling.Source: §4 · p. 20
• Proposed extensions combine CABS-dock with high-resolution refinement, exact scoring, experimental data, and automatic identification of flexible receptor fragments.The current advanced option permits user-selected receptor flexibility; automatic selection is suggested as a future extension.Source: §4 · p. 20
• CABS-dock’s potential applications include searching for interaction hot spots and binding areas, supporting peptide design, and computational epitope mapping.The paper links these applications to protein-peptide and protein-protein interaction characterization and modern drug design.Source: §4 · pp. 20–21

APPENDIX

• Michal Jamroz supported modeling-script programming, online documentation, continued CABS-dock system improvement, and discussions.Source: APPENDIX · p. 21

2. CABS-dock result files and reference structure

• The tutorial covers loading files, visualizing trajectories, aligning models to a known receptor structure, calculating peptide RMSDs, and plotting RMSD against energy.Source: §1 · p. 27
• The tutorial uses VMD with its RMSD Trajectory Tool plug-in and gnuplot for molecular-system visualization and plotting.The RMSD Trajectory Tool plug-in is included in VMD versions 1.8.8 and higher.Source: §1 · p. 27
• Both required programs are free to use, available for major operating systems, and described as fairly easy to install.Source: §1 · p. 27

2. CABS-dock result files and reference structure

• The example uses peptide SSRFESLFAG, secondary structure CHHHHHHHHC, and unbound human androgen receptor structure 2AM9.Source: §2 · p. 28
• The experimentally determined peptide-bound 1T7R androgen-receptor structure serves as the RMSD reference; when a bound complex is unknown, a resulting model may be used instead.Source: §2 · p. 28
• The archive includes 10 final all-atom models, cluster representatives, 10 trajectories, and a top1000.pdb set selected for further clustering and analysis.Source: §2 · p. 28
• The energy.txt file records trajectory and frame indices, temperature, and receptor, peptide, interaction, and total energy values for resulting models.Source: §2 · pp. 28–29

• Italics identify interface elements such as buttons, menu positions, fields, lists, and sections.Source: §3 · p. 29
• Lucida Console marks console input, values, and files, while underlining marks VMD window panels.Source: §3 · p. 29
• Start VMD, confirm the console, OpenGL Display, and VMD Main windows appear, then load files through the graphical interface or console.Source: §3 · p. 29

• In VMD, open File > New Molecule, browse to trajectory_1.pdb, choose PDB and Load all at once, then load the file.Source: From the graphical interface · pp. 29–30
• Add trajectories 2 through 10 to the trajectory_1.pdb molecule in numerical order for plotting RMSDs against energy.Source: From the graphical interface · p. 30
• The frame indicator should reach 9999 for 10,000 frames; load 1T7R.pdb as a new molecule and mark the trajectories as top.Source: From the graphical interface · p. 30

• Use VMD’s Tk Console as an alternative, faster way to load all files.Source: Via the console · p. 30
• Change to the folder containing server results with the cd command, then run the trajectory-loading command.Source: Via the console · p. 30
• A loop using mol addfile loads trajectory_1.pdb through trajectory_10.pdb and waits for each file to finish.Source: Via the console · p. 30
• The next section explains how to change VMD’s default graphical representation.Source: Via the console · p. 31

Chain identification

• Chain IDs appear in the Graphical Representations keyword list, and should be recorded for fragment selection; the peptide has one chain, usually the shortest.The example receptor and peptide both contain chains A and B, though receptors may have more chains.Source: Chain identification · pp. 31–32

Creating trace representation

• Trace representations color each chain separately, allowing trajectory models to be viewed with the frame slider; aligned receptor structures expose peptide conformational sampling.Source: Creating trace representation · pp. 32, 34
• The RMSD Trajectory Tool aligns receptors using chain A alpha carbons, then calculates peptide RMSDs using chain B alpha carbons and saves frame IDs and RMSDs.Source: Creating trace representation · pp. 33–34
• VMD can play trajectories as movies, navigate to individual frames, and export movies using Movie Maker with an external encoder.Source: Creating trace representation · pp. 34–35

4. Creating plots

• Combine energy.txt with the header-stripped trajrmsd.txt; gnuplot can plot RMSD against interaction energy or plot either measure against frame.Source: §4 · pp. 35–36
• The sample scatter plot uses RMSD on the X axis and interaction energy on the Y axis; sorting columns 9 and 6 finds the lowest RMSD and interaction-energy records, respectively.Source: §4 · pp. 35–36