CABS-dock web server for the flexible docking of peptides to proteins without prior knowledge of the binding site

INTRODUCTION

• Protein–peptide interactions are important in cell functions, but their structural details are poorly known because peptide binding is dynamic and transient, making experiments difficult.The introduction identifies computer-aided support such as molecular docking as needed.Source: INTRODUCTION · p. 1
• Docking must handle peptide conformational fluctuations and receptor flexibility, a major problem because even small receptor changes are computationally costly (1,2).The introduction notes that current algorithms are not efficient enough for the high conformational fluctuations of peptides.Source: INTRODUCTION · p. 1
• CABS-dock unifies binding-site prediction, peptide modeling, and complex refinement in an efficient simulation of peptide folding and binding to a flexible receptor, without prior binding-site knowledge.The approach allows full peptide flexibility and small receptor-backbone fluctuations.Source: INTRODUCTION · p. 2
• The benchmark included peptides of 5–15 amino acids and 103 bound plus 68 unbound receptor cases.Bound and unbound cases were determined experimentally in complex with and without a peptide, respectively.Source: INTRODUCTION · p. 2

Previous applications and background

• Earlier CABS-dock studies of disordered peptides (5), antibody-targeting peptides (6), and nuclear-receptor co-activators (7,8) produced arrangements close to native structures without binding-site or peptide-conformation information.Peptides were fully flexible in all the cited validation tests.Source: Previous applications and background · p. 2
• CABS-dock uses a multiscale coarse-grained CABS model, with up to four interaction centers per amino acid, Monte Carlo dynamics, and statistical-potential force fields (9).The model is designed to efficiently treat protein conformational changes while preserving local accuracy for reconstruction to all-atom representation.Source: Previous applications and background · p. 2
• The CABS protein model has been used for reliable simulations of long-term folding mechanisms (10,11) and near-native short-term fluctuations (12,13).These applications support its use for protein dynamics simulations.Source: Previous applications and background · p. 2
• CABS also performed well in protein-structure prediction, especially blind predictions of short globular proteins (14) and large protein fragments (15,16).Together with its interaction model and sampling scheme, these studies support simulations of simultaneous folding and binding.Source: Previous applications and background · p. 2

Protocol overview

• Random peptide structures are placed on a sphere centered at the receptor’s geometric center, with radius equal to its longest dimension plus 20 angstroms.This is the protocol’s starting step.Source: Protocol overview · pp. 2–3
• Ten temperature-spaced replicas undergo Replica Exchange Monte Carlo dynamics with decreasing temperatures, producing 10 trajectories of 1,000 snapshots each, or 10,000 models total.The temperatures decrease as the simulation proceeds toward the bottom of the energy minima.Source: Protocol overview · p. 3
• During simulation, the receptor is kept near-native using distance restraints on C-alpha atom pairs selected from the input-structure distance map, 5–15 Å apart and at least five residues apart in sequence.Violations of less than 1 Å are not penalized.Source: Protocol overview · p. 3
• User-designated semi-flexible residues halve the restraint-penalty slope, while fully flexible residues set it to zero for assigned restraints.The adjustments apply to all restraints assigned to those residues.Source: Protocol overview · p. 3
• Final-model selection first excludes unbound states, then selects up to 100 lowest-binding-energy models from each of 10 trajectories (fewer if under 100 bound states, rarely), yielding 1,000 models for clustering.The 1,000 selected models are clustered by k-medoids 100 times with different initial medoids and k=10.Source: Protocol overview · p. 3
• Final models are reconstructed from the C-alpha trace to all-atom structures and optimized using Modeller (17) with the DOPE statistical potential (18).Source: Protocol overview · p. 3

Docking without prior knowledge of the binding site

• The validation used the largest available non-redundant peptide–protein interaction dataset, with less than 70% receptor sequence identity.The benchmark had been created for FlexPepDock refinement and later extended with unbound cases in a HADDOCK study (3, 19).Source: Docking without prior knowledge of the… · p. 3
• The authors classified high-quality predictions as peptide RMSD < 3 Å and low-quality predictions as RMSD > 5.5 Å after receptor superimposition.They used 5.5 Å as the cutoff between low- and medium-accuracy models.Source: Docking without prior knowledge of the… · p. 4
• Models beyond the 5.5 Å cut-off are not necessarily useless for further refinement: in the FlexPepDock benchmark, refinement from 6.5–7.5 Å structures yielded near-native models in 48% of bound-docking cases (3).Source: Docking without prior knowledge of the… · p. 4
• CABS-dock performed at a similar level for bound and unbound cases in a blind test using neither bound-peptide structure nor binding-site information.For most benchmark cases, bound and unbound interface differences were below 1 Å and were handled with default receptor flexibility.Source: Docking without prior knowledge of the… · p. 4
• Over 80% of bound and unbound benchmark cases yielded high- or medium-accuracy models considered sufficient for practical use or further refinement.The authors note that prediction outcomes may vary between runs because the simulation is stochastic.Source: Docking without prior knowledge of the… · p. 4
• Prediction cases generally showed either consistent or qualitatively distinct outcomes across stochastic runs, so the authors recommend several independent runs for ambiguous cases.The prediction results may differ qualitatively between runs.Source: Docking without prior knowledge of the… · p. 4

Input interface and requirements

• The required receptor input is a PDB structure or PDB code with chain identifier(s), and single- or multi-chain proteins are accepted.Each chain may be up to 500 amino acids long.Source: Input interface and requirements · p. 5
• PDB residues must have complete backbone atoms (N, Cα, C and O), but side-chain atoms may be missing; non-standard amino acids are converted to standard counterparts.Chains must be up to 500 amino acids long.Source: Input interface and requirements · p. 5
• The required peptide input is a sequence in one-letter code, using standard amino acids only, with a maximum length of 30 amino acids.Source: Input interface and requirements · p. 5
• Peptide secondary structure is optional; when omitted, PSI-PRED (20) predicts it, while experimentally derived assignments are recommended when available.Assignments use H for helix, E for extended state (beta strand), and C for coil.Source: Input interface and requirements · p. 5
• For ambiguous peptide secondary-structure predictions, assigning coil (C) is preferable to overpredicting regular helix (H) or extended (E) structure.Overpredictions of regular secondary structure are more harmful to result quality than underpredictions.Source: Input interface and requirements · p. 5
• Optional project names help organize users’ work, email addresses enable job-completion notifications, and advanced options can also be supplied.Source: Input interface and requirements · p. 5

Docking prediction results

• The Docking prediction results tab displays and lets users download 10 final models, plus an archive of models, trajectories, receptor input, and simulation log.The 10 final models represent structural clusters; the archive includes cluster models and complete trajectories in PDB and C-alpha representation.Source: Output interface · p. 5

Clustering details

• Clustering details provides an interactive chart linking cluster composition to trajectory affiliation, model viewing and downloads, and a table of cluster characteristics.Users can select chart points to view models in 3D or download them in PDB format; the table summarizes density and diversity.Source: Output interface · p. 6

Contact maps

• Contact Maps lets users inspect peptide–receptor residue contacts in an interactive chart, set the contact cutoff distance, and view a text list of mapped contacts.Source: Output interface · p. 6

Advanced options

• Users can exclude binding modes by selecting receptor residues unlikely to interact with the peptide or by marking models from a completed job.Residues can be listed when starting a job or marked models can be excluded by resubmitting a completed job.Source: Advanced options · p. 6
• For selected receptor fragments, users can remove restraints that keep the structure near-native and choose moderate or full flexibility.The option is available by marking flexible regions on the main page.Source: Advanced options · p. 6
• Users can raise the Monte Carlo macrocycle count from 50 to at most 200 for difficult cases, such as large receptors or peptides longer than 20 residues.Running independent simulations is presented as an alternative for demanding cases.Source: Advanced options · p. 6

• Documentation is accessible through links in the menu at the top of every CABS-dock server page.Source: Online documentation · p. 6
• It describes the method and includes a tutorial on accessing and interpreting resulting data.Source: Online documentation · p. 6
• The online documentation is updated regularly according to users’ needs or method improvements.Source: Online documentation · p. 6

Server and output data availability

• The server is free to use and requires no login.Submitted results receive a link that can be bookmarked and revisited, but are retained only for a limited time.Source: Server and output data availability · p. 7
• Submitted jobs appear on a queue page unless users select the option to hide them from the results page.The queue page is available from the main page.Source: Server and output data availability · p. 7

Server architecture and run-time

• The website interface and parsers use Python, Flask, and Jinja2; molecular visualization uses 3Dmol.js and JSmol, with Apache2 and SQLite3 supporting the website and queue storage.Source: Server architecture and run-time · p. 7
• The queue is checked every 5 minutes, and new jobs are added to the SGE queue; a typical run takes about 3 hours on a Linux supercomputer cluster with about 100 CPU threads.Job status changes from pending to running when computation starts, then to done or error after completion.Source: Server architecture and run-time · p. 7

SUMMARY

• The CABS-dock protocol had already been used successfully in protein–peptide interaction studies (5-8), and this work developed an easy-to-use web server interface for it.Source: SUMMARY · p. 7
• The authors expect the web server to be applied to new systems and used as part of new modeling procedures.Source: SUMMARY · p. 7
• The authors identify refinement, experimental-data incorporation, and greater flexibility of selected receptor fragments as possible CABS-dock extensions (19,23) (24).Predicted restraints are given as one example for increasing receptor-fragment flexibility.Source: SUMMARY · p. 7

FUNDING

• The work was supported by the Foundation for Polish Science TEAM project, EU European Regional Development Fund, and Polish Ministry of Science and Higher Education.The open-access charge was supported by the Foundation for Polish Science TEAM project.Source: FUNDING · p. 7

TABLE AND FIGURES LEGENDS

• In the 2P1T example, 7 of 10 final models reached the native binding site, and the most accurate model was within 1.37 Å of the native peptide.The figure illustrates the stages from random peptide conformations and positions through 10,000 simulated models, filtering and clustering, and 10 representative final models.Source: Figure 1 · p. 12
• Figure 2 summarizes high-, medium- and low-accuracy model percentages across 103 bound and 68 unbound cases, for all 10,000 generated models and the top 10 selected models.Detailed results are listed in Tables S1 and S2.Source: Figure 2 · p. 13, TABLE AND FIGURES LEGENDS · p. 13
• The web-server screenshots show the ‘Docking prediction results’ and ‘Clustering details’ tabs.Source: Figure 3 · p. 14, TABLE AND FIGURES LEGENDS · p. 14

Supplementary data

• The supplementary materials provide a flow-chart of the CABS-dock protocol and list PDB codes for receptor pairs in bound and unbound forms.The main text directs readers to Figure S1 for a more detailed pipeline flow-chart.Source: TABLE AND FIGURES LEGENDS · p. 12, Figure S1 · p. 15, Table S3 · p. 23
• Table S1 reports lowest peptide ligand-RMSD values for 103 bound cases across three prediction runs and four model sets: all 10,000, top 1k, top 100, and top 10.It also gives the lowest RMSD values across the three runs.Source: Table S1 · p. 16
• Table S2 reports lowest peptide ligand-RMSD values for 68 unbound cases across three prediction runs and four model sets: all 10,000, top 1k, top 100, and top 10.It also gives the lowest RMSD values across the three runs.Source: Table S2 · p. 20