How to Dock a Peptide to a Protein (HPEPDOCK, HADDOCK and AutoDock CrankPep): A Step-by-Step Guide

Answer: To dock a peptide to a protein, do not use standard small-molecule docking. Use a dedicated peptide method: HPEPDOCK for fast blind docking from a sequence, HADDOCK when you know which residues contact each other, and AutoDock CrankPep for local docking that folds the peptide inside the binding site.

Why can’t you just dock a peptide in AutoDock Vina?

Because a peptide has far too many rotatable bonds for a small-molecule search to handle. AutoDock Vina was designed around drug-like ligands, and its search becomes unreliable well before you reach the flexibility of even a 7-residue peptide. A peptide also has no single starting conformation to dock: its backbone folds as it binds.

The authors of HPEPDOCK put the problem plainly in their Nucleic Acids Research paper: peptides “do not have a stable conformation before binding to a receptor”. That is the whole difficulty. A small-molecule docking program samples ligand torsions inside a rigid pocket. A peptide docking program has to sample backbone conformation and binding pose at the same time.

Two consequences follow, and they shape everything below:

• You need a conformational ensemble or a folding search. Either the method pre-generates many peptide conformations, or it folds the peptide during docking.
• Scoring is harder. Peptide interfaces are larger and flatter than drug pockets, so a single score rarely separates the right answer from the near-misses. You will judge poses on clustering and on contacts, not on one number.

If you have not yet run a standard docking job, start with the molecular docking pillar guide and the walkthrough on preparing a protein and ligand for docking. Peptide docking assumes you already know what a binding pose and an RMSD are.

Which peptide docking method should you choose?

Choose by what you already know about the binding site and by whether you can install software. Three routes cover almost every student project.

A practical sequence for a thesis project: run HPEPDOCK first to find candidate sites, then re-dock the best site locally with ADCP or with HADDOCK using restraints. Reporting agreement between two independent methods is far more convincing than one score from one server.

How do you prepare the receptor and the peptide?

Preparation decides whether the run is meaningful. Do these four things before you touch any docking tool.

1. Clean the receptor. Remove waters, crystallization additives and any co-crystallized peptide you plan to redock. Keep structural metal ions and cofactors that sit in or near the site.
2. Fix the structure. Missing side chains and gaps in loops near the interface will distort the pose. Repair them before docking, not after.
3. Add hydrogens at the right protonation states. Peptide binding is dominated by salt bridges and hydrogen bonds, so histidine and terminal charges matter more here than in a hydrophobic drug pocket.
4. Decide the peptide termini. A synthetic peptide is often acetylated at the N-terminus and amidated at the C-terminus. A free peptide carries a positive and a negative charge that do not exist in the real construct. Match the construct you actually study, and say which you used in your methods.

If you do not know where the peptide binds, run a pocket detection pass first. Our guide on finding a protein’s binding site with fpocket, CASTp and PrankWeb covers the tools, and the predicted pockets give you the box coordinates that ADCP needs.

How do you run a blind peptide docking with HPEPDOCK?

HPEPDOCK is the fastest way to get a defensible answer with no installation. It builds an ensemble of peptide conformations with the MODPEP program, docks that ensemble hierarchically, and ranks the results with an iterative knowledge-based scoring function.

The workflow on the HPEPDOCK server:

1. Submit the receptor. Upload a PDB file, give a PDB ID and chain such as 3BFW:A, or paste a sequence in FASTA format. Sequence input triggers modelling, so a real structure is always preferable.
2. Submit the peptide. Paste the one-letter sequence or upload a structure. The server recommends sequences under 30 amino acids for accuracy.
3. Choose global or local docking. Global searches the whole surface. Local requires binding site residues and is substantially more accurate, so use it whenever you have evidence for a site.
4. Set the output size. The default is the top 100 binding models, adjustable from 1 to 500.
5. Wait. The paper reports an average of 29.8 minutes for a global job and 14.2 minutes for a local job.

What the numbers mean. In the benchmark reported by Zhou and colleagues, global blind docking reproduced the correct binding mode for 33.3% of 57 unbound cases within the top 10 models, against 21.1% for the pepATTRACT server. Local docking reached 72.6% on 62 unbound cases within the top 10. Read those figures honestly: a blind run gives you a roughly one-in-three chance in the top 10, which makes it a hypothesis generator, not a result.

One useful detail that students miss: HPEPDOCK also docks peptides against DNA and RNA, so the same server covers nucleic-acid targets.

How do you run information-driven peptide docking with HADDOCK?

HADDOCK is the right choice when you can point at the interface. It converts your knowledge into ambiguous interaction restraints and drives the docking with them, which is why it outperforms blind methods when the information is real and misleads badly when the information is wrong.

The peptide protocol comes from Trellet and colleagues in PLoS ONE (2013). It starts from an ensemble of three peptide conformations: alpha-helix, polyproline-II and extended. Docking that ensemble produced high-quality models in 79.4% of bound/unbound cases and 69.4% of unbound/unbound cases on the benchmark used in that study.

Steps for a peptide run:

1. Build the three-conformation ensemble. Generate the helical, polyproline-II and extended forms of your peptide and concatenate them into one multi-model PDB file. This ensemble is the single most important peptide-specific setting.
2. Define active residues. On the protein, the residues you have evidence for, filtered by solvent accessibility. On a short peptide, the usual practice is to treat the whole peptide as active because every residue is at the interface.
3. Increase sampling. The HADDOCK3 protein-peptide scenario generates 3000 rigid-body models and carries the top 400 into refinement, well above protein-protein defaults.
4. Lengthen the refinement. The Bonvin Lab peptide best-practice guide recommends 2000 MD steps for rigid-body high-temperature TAD, 2000 for the first cooling stage, 4000 for the flexible side-chain stage and 4000 for the fully flexible interface stage.
5. Keep secondary structure sane. The peptide scenario applies dihedral restraints with ssdihed = "alphabeta" so refinement does not unravel a genuine helix.

Read the output as clusters, not as single structures. HADDOCK ranks clusters by score, and a cluster of many similar poses with a low score beats a single low-scoring outlier. The same reading applies to protein-protein docking with ClusPro, HDOCK and HADDOCK.

How do you run local flexible docking with AutoDock CrankPep?

AutoDock CrankPep folds the peptide inside the receptor’s potential field using a Monte Carlo search, so no conformational ensemble is needed. Zhang and Sanner reported in Bioinformatics (2019) an 85.7% success rate within the top 10 on the LEADS-PEP dataset, and for peptides of 16 to 20 residues, re-docking success of 64% for the top solution and 91% within the top 5.

ADCP ships with the ADFRsuite. The workflow has two commands.

1. Prepare the receptor and build the target file. Convert the protonated receptor to PDBQT with prepare_receptor from the ADFRsuite (the suite also bundles reduce, which usually adds hydrogens better). Then define the box and compute the affinity maps with agfr. The official redocking tutorial uses:

agfr -r 3Q47_recH.pdbqt -l 3Q47_pepH.pdbqt -asv 1.1 -o 3Q47

Here -r is the receptor, -l is a reference ligand whose bounding box defines the docking box (agfr adds 4 Angstrom padding on every side), -asv selects the AutoSite version, and -o names the output target file. If you have no reference peptide, define the box directly with the -b / --boxMode option using coordinates from your pocket detection run.

2. Dock the peptide. The same tutorial runs:

adcp -t 3Q47.trg -s npisdvd -N 20 -n 1000000 -o 3Q47_redocking -ref 3Q47_pepH.pdb

-t is the target file from agfr, -s is the peptide sequence in lower case one-letter code, -N is the number of independent searches, -n is the number of steps per search, -o is the output prefix, and -ref supplies a reference structure so ADCP reports RMSD directly. That published example finished in 148.19 seconds.

For production runs, the ADCP documentation recommends roughly 100 replicas with about one million steps per amino acid, starting from a mixture of extended and helical conformations. A 10-residue peptide therefore wants -N 100 -n 10000000, which is a long job on a laptop and a short one on a workstation.

How do you score and validate a peptide pose?

No peptide docking score is a binding affinity. Treat every number as a ranking device and validate the pose independently. Four checks, in order of how much reviewers care:

1. Redock a known complex. Take a crystal structure of your receptor with any peptide bound, strip the peptide, dock it back, and measure backbone RMSD to the crystal pose. Under 2 Angstrom is the usual criterion. It is the single most persuasive control you can run, and it pairs with the settings discipline described in our guide on making docking runs reproducible.
2. Check convergence. Cluster the top models. If independent runs converge on one site and one backbone path, that is evidence. If the top 10 poses sit in 10 different places, your peptide has no preferred site under that scoring function, and you should say so.
3. Check the contacts, not the score. List the hydrogen bonds and salt bridges in the pose and compare them with published mutagenesis. A pose that buries a residue known to be dispensable, and ignores one known to abolish binding, is wrong regardless of its rank. Our guide on interpreting docking results and binding scores covers how to read these numbers.
4. Run a short MD simulation. Peptide poses are the classic case where docking looks fine and 50 nanoseconds of MD shows the peptide sliding out of the groove. Stability under MD is the strongest evidence a student project can produce, and it is the natural next step on the computational biology skills roadmap.

What goes wrong, and how do you fix it?

Frequently asked questions

Can AutoDock Vina dock a peptide at all?

It can produce output for very short peptides of two or three residues, and the result is still unreliable because the search was not designed for backbone flexibility. Use a peptide-specific method for anything longer. Vina’s flexible side-chain docking makes the receptor flexible, not the ligand backbone, so it does not solve this problem.

How long can the peptide be?

HPEPDOCK recommends under 30 amino acids. ADCP was benchmarked on peptides up to 20 residues, with success rates reported separately for the 16 to 20 residue range. Beyond roughly 30 residues you are no longer doing peptide docking; treat the partner as a small protein and use protein-protein docking.

Do I need the peptide’s 3D structure before docking?

No. HPEPDOCK and ADCP both accept a sequence, since both generate or fold the conformation themselves. HADDOCK needs structures, which is why its peptide protocol builds an ensemble of three standard conformations first.

Which method should I report in my thesis?

Report the one that matches your evidence, and report the control. If you had no site information, say the pose came from blind HPEPDOCK docking and give the redocking validation. If you had mutagenesis data, HADDOCK with explicit restraints is the stronger claim because your method reflects your evidence.

Is peptide docking accurate enough to publish?

Yes, with controls. The published success rates are honest about the limits: roughly one in three for blind global docking in the top 10, and much higher once the site is known. Pair the docking with a redocking control and a short MD stability run, and describe the pose as a model rather than a determined structure.

Can I dock a cyclic peptide?

ADCP supports cyclic peptides, and a dedicated follow-up study on docking flexible cyclic peptides with AutoDock CrankPep describes the procedure. Cyclization reduces the conformational search space, which usually helps accuracy.

Written by the StemSkills Lab team, with over ten years of combined work in sequence and structural bioinformatics, drug discovery and design, and multiscale molecular modeling.