Get Started

Home Domains Molecular Docking

Molecular Docking

Molecular docking explained end to end: preparing proteins and ligands, reading PDB structures, choosing a binding site, running AutoDock Vina and judging a pose.

Structure-based designDomain guideFree to readUpdated 2026-10-13

Docking produces hypotheses about how a molecule may sit in a binding site. A score is not a measurement of affinity and is never evidence that a compound is safe or effective.

Docking is the cheapest way to ask a structural question: could this molecule sit in this pocket, and how? It is also one of the easiest methods to misuse, because a ranked list of numbers looks like a result even when the preparation, the box or the scoring function made it meaningless.

This guide is written around the checks that separate the two. Every stage ends in something you can inspect — a prepared file, a control RMSD, a pose that either matches the known chemistry of the site or does not.

If you are here to run your first docking, read the protein and ligand preparation sections before the workflow section. The workflow is the short part; the preparation is the science.

What docking actually computes

Docking places a ligand into a binding site and ranks the placements with a fast approximate function. Two separate programs are doing the work: a search algorithm that proposes poses, and a scoring function that ranks them. Understanding that split is the difference between using docking and trusting it blindly.

  • Search samples ligand position, orientation and (in flexible docking) rotatable bonds inside a defined box.
  • Scoring is empirical or knowledge-based. It approximates interaction energy quickly, which is why it is wrong in absolute terms but useful for ranking.
  • The output is a set of poses with scores in kcal/mol, not a dissociation constant.
  • Docking says nothing about cell permeability, metabolism or whether the site is reachable in a living system.

Preparing the protein

Most docking failures are preparation failures. A receptor file is a curated artefact: chains chosen, waters decided on, protonation states set, missing atoms rebuilt and every decision written down.

  • Choose a structure by resolution and fit to experimental density, not by how recent the deposition is.
  • If you are validating the method, use a structure that already contains the ligand or partner bound in that site.
  • Decide consciously about crystallographic waters — some are structural or directly mediate binding.
  • Set protonation states and histidine tautomers for the pH you are modelling, and record what you set.
  • Reject structures with large unresolved regions in or near the site unless you can rebuild them defensibly.

Preparing the ligand

A SMILES string is not a docking input. The molecule has to be converted to a 3D structure with the protonation state, tautomer, stereochemistry and formal charges that are plausible under your conditions.

  • Generate 3D coordinates and check that bond lengths and ring geometry came out sane.
  • Pick the protonation state for the target pH — a charged carboxylate and its neutral form are different ligands to the scoring function.
  • Enumerate the tautomers that matter, and say which one you docked in the methods.
  • Keep stereochemistry explicit, because enantiomers can score differently in the same site.
  • Reuse the exact same prepared file across a series of related compounds, or the comparison stops meaning anything.

PDB structures and finding the binding site

The Protein Data Bank is a working tool, not an archive you link to once. Reading an entry properly — chains, ligands, mutations, assembly, resolution — is a core docking skill, and choosing the box is part of the scientific argument.

  • Start from the co-crystallised ligand or a known partner when one exists; that geometry defines the site.
  • Use cavity detection such as fpocket when the site is unknown, then check the result against the literature.
  • Check UniProt annotations for domain boundaries, active-site residues and known variants.
  • Define the grid box around the site with a small margin, and log its centre and size — a box that covers half the protein invites junk poses.

Running a docking workflow

A reproducible docking run is a short script with recorded parameters, not a series of clicks. AutoDock Vina, PyRx, Smina and GNINA cover the common use cases, and each exposes the settings that decide whether the run means anything.

  • Set exhaustiveness and the random seed explicitly so a rerun can reproduce the same poses.
  • Keep a redocking control in the same run — the same ligand you removed from the structure — to measure whether the protocol can find a known answer.
  • For a series of ligands, keep the receptor, box, protonation strategy and parameters identical.
  • Store inputs, logs and the exact tool versions with the results; versions change scores.

Reading the results honestly

The top-ranked pose is a candidate, not a conclusion. Interpretation is where docking becomes science: what the pose implies chemically, whether the control worked, and what the score can and cannot support.

  • Check the pose against known interactions — hydrogen bonds, salt bridges, hydrophobic contacts with named residues.
  • Use redocking RMSD against the experimental pose as your protocol quality check; below roughly 2 Å is the common working threshold for a validated setup.
  • Compare scores only within one run and one receptor. Cross-target score comparisons are meaningless.
  • In virtual screening, confirm that known actives rank above decoys before you report any hit.
  • Treat every predicted binder as a hypothesis to test experimentally, never as a result.

Learning path

Four stages, each ending in something you can show — a prepared file, a figure, a reproducible run. The order matters more than the speed.

Start

Protein structure literacy

Read a PDB entry until you can name the chains, the bound ligands, the resolution and the gaps — and explain what each means for the site you care about.

Build

Prepare one complex, dock it

Prepare a receptor and a ligand with written decisions, dock with AutoDock Vina, and report the top three poses with their scores.

Validate

Redock a known answer

Remove a co-crystallised ligand, dock it back, and measure the RMSD to the experimental pose. If the protocol cannot reproduce a known pose, stop and fix it.

Apply

Screen a focused library

Dock a small set of related compounds against one site, triage by pose plausibility before score, and state the limits of what you found.

Tools this domain uses

Open-source and freely available — that is what we teach. Commercial platforms exist, but none of them are required to learn the work.

  • AutoDock VinaThe reference open-source docking engine: search plus scoring, with an RMSD table for the reported modes.
  • PyRxA visual front end that wraps Vina and AutoDock for people who want to see the box and the poses without writing scripts first.
  • RDKitLigand preparation in code: reading structures, generating conformers, protonation and property calculation.
  • Open BabelFormat conversion between the many file types you will meet — PDB, SDF, MOL2, SMILES — and quick 3D generation.
  • PyMOL or ChimeraXPose inspection: measuring distances, checking clashes, and making the figure that shows a reviewer what you claim.
  • fpocketCavity detection when the binding site is not already known from a co-crystal structure.

Still being built

Listed openly so nothing here is mistaken for something you can use today. None of these are published yet.

  • Guided AutoDock Vina walkthrough with a downloadable sample complex — In development — not published yet
  • Receptor preparation checklist you can score your own file against — In development — not published yet

Questions learners ask

Is a docking score a binding affinity?

No. It is a fast empirical estimate used to rank poses and, loosely, compounds. Converting a score into a predicted affinity or a dissociation constant is not something the standard scoring functions support.

Do I need a powerful computer to learn docking?

Not for one complex. A normal laptop runs AutoDock Vina on a single protein–ligand pair in seconds to minutes. Throughput becomes the constraint only when you dock libraries of thousands of compounds.

Which PDB structure should I choose?

Prefer a structure with the relevant ligand or partner bound, good resolution, and minimal unresolved regions around the site. If several qualify, compare them and say why you picked one.

How do I know my docking result is meaningful?

Run the positive control. Redock a ligand whose experimental pose is known and measure the RMSD. A protocol that fails to reproduce a known pose cannot be trusted for an unknown one.

Can docking replace laboratory experiments?

No. Docking narrows the list of things worth testing. Every predicted binder still has to be made and measured, and most predictions will not survive that step.

01/Related domains