# Molecular docking steps: from protein preparation to validation

> Molecular docking runs in seven steps, from choosing and preparing the receptor to validating the poses. Learn what each step decides, the types of docking, how to prepare a protein for docking, and how one imatinib run moved from 0.7 Å to 12 Å when a single step changed.

## What are the steps of molecular docking?

Molecular docking predicts how a small molecule sits in a protein pocket, and it runs in seven steps:

1. [Choose the receptor structure](#step-1-choose-the-receptor-structure)
2. [Prepare the protein](#step-2-prepare-the-protein)
3. [Prepare the ligand](#step-3-prepare-the-ligand)
4. [Define the binding site and search box](#step-4-define-the-binding-site-and-search-box)
5. [Run the search and scoring](#step-5-run-the-search-and-scoring)
6. [Analyze the poses](#step-6-analyze-the-poses)
7. [Validate the setup](#step-7-validate-the-setup)

Only the fifth step is the docking program itself. The other six decide what the program is asked and how far to believe its answer. The [types of docking](#types-of-molecular-docking) and the [full imatinib test](#one-ligand-fifteen-setups) have their own sections.

The program is the easy part to get right. We docked imatinib into its crystal structure in c-Abl kinase (PDB 1IEP) with [AutoDock Vina](/app/autodock-vina) 15 times, changing one choice per run. The default setup put the top pose 0.76 Å from the crystal pose. A different random seed, a box 7 Å narrower, a box centered on the residue imatinib is known to bind, or an AlphaFold model of the same kinase moved the top pose 10 to 12 Å away. Nothing in any single run said it had failed.

What changes the answer most is the receptor conformation and where the search is allowed to look. Hydrogens, waters and the ligand's charge moved this score by up to 1.5 kcal/mol but left the pose where it was. The rest of this page goes through the steps in order, with what each one decides and the check that catches its usual mistake.

## Types of molecular docking

Docking methods differ in what they let move, how much they know about the site, what the partner is and how they search. Kuntz and colleagues docked rigid shapes in 1982; most current programs keep the protein rigid and let the ligand flex, and newer methods can change the pocket or skip the search box altogether. Table 1 sorts the common types by the question each one answers.

  <caption><strong>Table 1. Types of molecular docking.</strong> Grouped by what moves, what is known about the site and what the partner is. Methods in one row can be combined with those in another: a blind docking run can also be flexible, for example.</caption>
  <thead>
    <tr>
      <th scope="col">Type</th>
      <th scope="col">What changes</th>
      <th scope="col">Use it when</th>
      <th scope="col">Example tools</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Rigid docking</td>
      <td>Nothing internal; ligand and protein keep their shapes</td>
      <td>Fast shape screens and protein-protein first passes</td>
      <td>[LightDock](/app/lightdock)</td>
    </tr>
    <tr>
      <td>Flexible-ligand docking</td>
      <td>Ligand torsions, position and orientation; protein rigid</td>
      <td>The default for small molecules in a known pocket</td>
      <td>[AutoDock Vina](/app/autodock-vina), [smina](/app/smina), [GNINA](/app/gnina)</td>
    </tr>
    <tr>
      <td>[Flexible receptor docking](/use-cases/flexible-molecular-docking)</td>
      <td>Selected side chains, or the pocket itself</td>
      <td>The pocket must move to fit the ligand</td>
      <td>[AutoDock Vina](/app/autodock-vina) flexible residues, [DynamicBind](/app/dynamic-bind)</td>
    </tr>
    <tr>
      <td>[Ensemble docking](/use-cases/ensemble-docking)</td>
      <td>The receptor, picked from several structures</td>
      <td>Several crystal or simulated conformations exist</td>
      <td>Any docking program, run once per structure</td>
    </tr>
    <tr>
      <td>Site-directed docking</td>
      <td>Search limited to a box around a known pocket</td>
      <td>A ligand-bound structure, mutagenesis or a homolog marks the site</td>
      <td>[AutoDock Vina](/app/autodock-vina), [GNINA](/app/gnina)</td>
    </tr>
    <tr>
      <td>[Blind docking](/use-cases/blind-docking)</td>
      <td>Search over the whole protein surface</td>
      <td>The site is unknown</td>
      <td>[DiffDock](/app/diff-dock), [fpocket](/app/fpocket) then Vina</td>
    </tr>
    <tr>
      <td>[AI docking](/use-cases/ai-molecular-docking) and co-folding</td>
      <td>A learned model places the ligand, or predicts protein and ligand together</td>
      <td>No good structure, or a fast first guess at the site</td>
      <td>[DiffDock](/app/diff-dock), [Boltz-2](/app/boltz-2)</td>
    </tr>
    <tr>
      <td>[Protein-protein docking](/use-cases/protein-protein-docking)</td>
      <td>Two proteins, mostly rigid bodies with refinement</td>
      <td>Predicting a complex interface</td>
      <td>[LightDock](/app/lightdock), [DiffDock-PP](/app/diffdock-pp)</td>
    </tr>
    <tr>
      <td>[Peptide-protein docking](/use-cases/peptide-protein-docking)</td>
      <td>A flexible peptide on a protein surface</td>
      <td>Peptide ligands too flexible for small-molecule programs</td>
      <td>[LightDock](/app/lightdock)</td>
    </tr>
  </tbody>

The steps below follow site-directed, flexible-ligand docking of a small molecule, the most common case. The other types skip or change some steps: blind docking replaces step 4 with a pocket search, and co-folding methods replace steps 1, 2 and 4 with a prediction. Steps 6 and 7 apply to all of them.

## Step 1: choose the receptor structure

The receptor decides more than any setting after it. A crystal structure with a ligand like yours in the pocket is the best start, a structure solved with a different ligand is next, and an unbound or predicted structure is last. In the GNINA 1.0 benchmark, the top Vina pose was within 2 Å in 58% of redocking cases, where the receptor came from the ligand's own crystal, and in 27% of cross-docking cases, where it came from another ligand's crystal.

Predicted structures sit lower still, even when they look confident. Holcomb and colleagues redocked 2,474 PDBbind complexes with [AutoDock-GPU](/app/autodock-gpu): the top pose was within 2 Å for 41% against the crystal receptors and 17% against [AlphaFold2](/app/alphafold-2) models. Karelina and colleagues found a similar gap on GPCRs with Glide: 15% correct on AlphaFold2 models, which was not significantly better than the 9% on traditional homology models, against 44% on experimental structures solved with other ligands.

Our run shows why. The AlphaFold model of human ABL1 has a mean pLDDT of 90.9 across the 32 pocket residues and matches 1IEP to 1.96 Å over their Cα atoms. But its DFG phenylalanine, Phe382, has moved 8.5 Å from its crystal position into the space imatinib fills: nine of its atoms lie within 2.5 Å of crystal imatinib atoms, the closest at 0.5 Å. Imatinib binds an inactive form of Abl, as the 1IEP structure showed, and the model shows another state.

None of the nine poses came within 10 Å of the crystal pose, and the best score, −8.99 kcal/mol, was 4 kcal/mol weaker than in the crystal structure. A confident model of a different state is still a different state.

Search the [PDB](/app/pdb-download) for every structure of your target before choosing one, and prefer one whose bound ligand resembles yours. When there is none, an [AlphaFold Database](/app/alphafold-database-download) or [Boltz-2](/app/boltz-2) model is a hypothesis about the pocket, and [ensemble docking](/use-cases/ensemble-docking) into several conformations is safer than trusting one.

## Step 2: prepare the protein

Protein preparation for docking turns a deposited structure into a model the scoring function can read. A PDB entry has no hydrogens in most cases, may have missing side chains, loops and alternate conformations, and carries waters, ions, buffer molecules and other chains that may or may not belong in the pocket. Sastry and colleagues tested each preparation step with Glide on 36 targets and then on DUD, and found that skipping some of them caused a systematic loss of enrichment, large for some targets.

To prepare a protein for docking, work through these decisions in order:

1. Keep the chain or chains that form the pocket and delete other copies in the asymmetric unit.
2. Pick one alternate conformation per residue and rebuild missing side chains and short loops near the site with [PDBFixer](/app/pdb-fixer).
3. Decide on each hetero group: delete buffer molecules and the ligand you will redock, keep metal ions and cofactors the ligand touches, and decide on waters (below).
4. Add hydrogens and assign protonation states at the assay pH. [PROPKA](/app/propka) predicts residue pKa values and [PDB2PQR](/app/pdb2pqr) applies them; histidine tautomers and active-site aspartates and glutamates are the usual problems.
5. Write the format the docking program reads, such as a [PDBQT file](/guides/pdbqt-file) for Vina, and open it once to check that the pocket residues and their hydrogens are there.

Hydrogens looked optional in our test, and they are not. Vina recognizes hydrogen-bond donors only through their hydrogens. Without them, the receptor file had no donor atoms, the top pose was still correct at 1.27 Å, and its score fell from −12.99 to −11.54 kcal/mol. A correct pose with a quietly wrong score is the failure that a [docking score](/guides/vina-score) cannot reveal by itself, and it would distort any ranking of several ligands.

Waters are a judgment call. Roberts and Mancera found a statistically significant gain in docking accuracy when they kept waters that touched both ligand and protein in the crystal. Keeping the 30 crystal waters within 10 Å of imatinib left our pose at 0.77 Å and made the score 0.72 kcal/mol more favorable. Those waters were placed around imatinib itself, so they fit its pose by construction; for a new ligand, keep only waters that bridge protein and ligand in several structures, or predict them with [SuperWater](/app/superwater).

## Step 3: prepare the ligand

Ligand preparation decides what molecule the program docks. Start from a SMILES string or an SDF with explicit bond orders, not a PDB file, which has no bond orders. Set stereochemistry explicitly, enumerate the stereoisomers with the [stereoisomer generator](/app/stereoisomer-generator) when it is unknown, and remove salts and counterions.

Then choose the protonation and tautomer state at the assay pH with [Dimorphite-DL](/app/dimorphite-dl), build a 3D conformer and write the program's input format, for example with [SMILES to PDBQT](/app/smiles-to-pdbqt). The [SDF file](/guides/sdf-file) and [chemical file formats](/guides/chemical-file-formats) pages cover which formats keep which information.

Protonation is the step whose effect depends most on the scoring function. Szakács and colleagues measured imatinib's acid-base profile and found it mostly monocationic at blood pH. Docking the cation with the distal piperazine nitrogen protonated gave a top pose at 0.80 Å and a score of −12.95 kcal/mol, almost exactly the neutral result.

Vina has no electrostatic term, so the charge changes only which atoms count as donors. A force-field or [learned scoring function](/guides/docking-scoring-functions) can respond much more, and a wrong tautomer moves donors and acceptors in every program.

Each ligand in a [virtual screening](/use-cases/virtual-screening) library needs the same treatment, which is why library suppliers ship prepared 3D files and why the choice of [compound library](/guides/virtual-screening-compound-libraries) is partly a choice of preparation.

## Step 4: define the binding site and search box

The search box tells the program where the ligand may go, and in our test it was the cheapest way to get a wrong answer. When a ligand-bound structure exists, center the box on that ligand. When it does not, use mutagenesis, a homolog's ligand or a pocket finder such as [fpocket](/app/fpocket), and treat the pocket as a hypothesis.

Size the box to the ligand. Feinstein and Brylinski tested Vina on 3,659 complexes and found pose accuracy highest with a cubic box edge of 2.857 times the ligand's radius of gyration; the average RMSD fell from 4.9 Å with the default box to 4.0 Å. For imatinib, whose crystal pose has a heavy-atom radius of gyration of 6.69 Å, the rule gives 19.1 Å. We calculated the crystal pose's extent along its longest axis at 16.7 Å, so a box near the rule leaves little room.

An 18 Å cube, close to that rule, failed. The top pose was the molecule reversed end to end in the right pocket, 12.07 Å away, and the only near-correct pose ranked fifth at −7.75 kcal/mol: squeezed against the walls of the box, it scored 5.2 kcal/mol worse than the correct top pose in the 25 Å box.

A box centered on the hinge residue Met318, whose backbone imatinib hydrogen-bonds in the crystal, put the center 11.4 Å from the ligand's own center and cut off the back of the pocket. The top pose then had imatinib's crystal shape (0.9 Å after fitting) but sat 11 Å out of place, at a respectable −11.40 kcal/mol.

A box around the whole kinase domain, 60 × 51 × 71 Å, did no better: at the default exhaustiveness the top pose was 12.15 Å away and none of nine came closer than 11.9 Å. The Vina manual advises raising exhaustiveness for boxes larger than about 30 × 30 × 30 Å, or 27,000 Å³, and this one held about 214,000 Å³. [Blind docking](/use-cases/blind-docking) works better with tools built for it, such as [DiffDock](/app/diff-dock), or by docking into each predicted pocket in turn.

## Step 5: run the search and scoring

The docking program searches for low-scoring poses and ranks them with its scoring function. Two kinds of choice belong to this step: which program and scoring function to use, and how hard to search. The first is covered in [Vina vs smina vs GNINA vs DiffDock](/guides/vina-vs-smina-vs-gnina-vs-diffdock) and the [best free docking software](/guides/molecular-docking-software); the second is usually left at its default, and should not be.

Vina's search is stochastic. With the box, receptor and ligand fixed, we changed only the random seed. Three of five seeds (42, 2 and 4) found the crystal pose at 0.71 to 0.76 Å and a score near −12.99 kcal/mol. Seeds 1 and 3 returned the reversed pose at 12.1 Å and about −9.7 kcal/mol as their best answer, and none of their 18 poses came within 6.5 Å. At the default exhaustiveness of 8, a flexible ligand with 37 heavy atoms was found correctly in three runs out of five.

Raising exhaustiveness to 32 fixed one of the two failures. Seed 1 then found the crystal pose (0.78 Å, −13.04 kcal/mol) and seed 42 kept it, but seed 3 still returned the reversed pose at 12.08 Å and −9.80 kcal/mol. More search made a miss less likely, not impossible.

The useful signal was between runs, not inside one. The failed seeds scored about 3.2 to 3.3 kcal/mol worse than the successful ones, and every run that found −13 found the same pose. Run each new setup with three seeds, and treat a top pose that changes between them as a search that has not converged.

## Step 6: analyze the poses

Analysis turns poses into a hypothesis. Look at the top few poses, not only the first, and check that the score gap between them is larger than the program's noise. Read the score for what it is: a [docking score is not a binding affinity](/guides/docking-score-binding-energy-affinity), and Vina's correlation with measured affinity on CASF-2016 was R = 0.604, while it picked a near-native pose for 90.2% of complexes.

Then check the contacts. [PLIP](/app/plip) and [ProLIF](/app/prolif) list the hydrogen bonds, hydrophobic contacts and π-stacking of each pose, and the [protein-ligand interactions](/guides/protein-ligand-interactions) page explains how to read the diagrams. A pose that misses the hinge hydrogen bond of a kinase inhibitor, or the zinc of a metalloprotease inhibitor, is suspect whatever its score. Rescoring with a different function, such as the [GNINA CNN score](/guides/gnina-cnn-score), or with [MM/GBSA](/guides/mm-pbsa-vs-mm-gbsa) on a [short simulation](/guides/docking-vs-molecular-dynamics), is a second opinion on the same pose.

## Step 7: validate the setup

Validation asks whether the setup can find answers that are already known, before it is trusted with ones that are not. Bender and colleagues make the same case for large screens: run control calculations on known ligands and decoys and adjust the setup until it passes, before docking anything new.

[Redocking](/guides/redocking-docking-validation) is the first control: dock the crystal ligand back into its receptor and expect the top pose within 2 Å [RMSD](/guides/rmsd). Cross-docking into a structure solved with another ligand is the realistic test, and an enrichment run of known actives among decoys tests the ranking that a screen depends on.

[PoseBusters](/app/posebusters) checks that each pose is chemically and physically possible, a test that learned docking methods often fail. It does not check that the pose is right: in the cross-docking test of the redocking page, all 16 wrong imatinib poses passed every PoseBusters check.

Every row of Table 2 below but the AlphaFold one is a redocking test, the easiest case a setup will meet. Those failures came from settings, not from a hard target.

## One ligand, fifteen setups

Table 2 shows the full test. Each row changes one choice from the baseline and reports the top pose and the best of up to nine poses, all measured against the 1IEP crystal pose in the same frame without fitting.

  <caption><strong>Table 2. Imatinib redocked into c-Abl with one step changed at a time.</strong> Baseline: PDB 1IEP chain A, waters and hetero groups removed, hydrogens added with Open Babel 3.2.1, imatinib built from neutral SMILES (RDKit ETKDG, seed 42), Meeko 0.7.1, AutoDock Vina 1.2.7 with Vina scoring, exhaustiveness 8, seed 42, 25 Å cube on the crystal ligand, up to nine poses. RMSD is symmetry-corrected over 37 heavy atoms. Vina returns fewer than nine poses when the rest fall outside its 3 kcal/mol energy range. The AlphaFold model is AF-P00519-F1, superimposed on the 1IEP pocket Cα atoms.</caption>
  <thead>
    <tr>
      <th scope="col">Step</th>
      <th scope="col">Change from baseline</th>
      <th scope="col">Top score (kcal/mol)</th>
      <th scope="col">Top pose RMSD</th>
      <th scope="col">Best pose RMSD</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>None</td>
      <td>Baseline</td>
      <td>−12.99</td>
      <td>0.76 Å</td>
      <td>0.76 Å</td>
    </tr>
    <tr>
      <td>1. Receptor</td>
      <td>AlphaFold model instead of crystal</td>
      <td>−8.99</td>
      <td>10.34 Å</td>
      <td>10.34 Å</td>
    </tr>
    <tr>
      <td>2. Protein</td>
      <td>No receptor hydrogens</td>
      <td>−11.54</td>
      <td>1.27 Å</td>
      <td>1.27 Å</td>
    </tr>
    <tr>
      <td>2. Protein</td>
      <td>30 crystal waters kept</td>
      <td>−13.71</td>
      <td>0.77 Å</td>
      <td>0.77 Å</td>
    </tr>
    <tr>
      <td>3. Ligand</td>
      <td>Piperazine nitrogen protonated</td>
      <td>−12.95</td>
      <td>0.80 Å</td>
      <td>0.80 Å</td>
    </tr>
    <tr>
      <td>4. Box</td>
      <td>18 Å cube</td>
      <td>−9.71</td>
      <td>12.07 Å</td>
      <td>1.87 Å (rank 5)</td>
    </tr>
    <tr>
      <td>4. Box</td>
      <td>Centered on Met318, 11.4 Å off</td>
      <td>−11.40</td>
      <td>12.44 Å</td>
      <td>11.91 Å</td>
    </tr>
    <tr>
      <td>4. Box</td>
      <td>Whole kinase domain</td>
      <td>−9.62</td>
      <td>12.15 Å</td>
      <td>11.91 Å</td>
    </tr>
    <tr>
      <td>5. Search</td>
      <td>Seed 1</td>
      <td>−9.76</td>
      <td>12.09 Å</td>
      <td>6.67 Å</td>
    </tr>
    <tr>
      <td>5. Search</td>
      <td>Seed 2</td>
      <td>−12.99</td>
      <td>0.71 Å</td>
      <td>0.71 Å</td>
    </tr>
    <tr>
      <td>5. Search</td>
      <td>Seed 3</td>
      <td>−9.66</td>
      <td>12.08 Å</td>
      <td>6.56 Å</td>
    </tr>
    <tr>
      <td>5. Search</td>
      <td>Seed 4</td>
      <td>−12.97</td>
      <td>0.74 Å</td>
      <td>0.74 Å</td>
    </tr>
    <tr>
      <td>5. Search</td>
      <td>Exhaustiveness 32, seed 42</td>
      <td>−13.01</td>
      <td>0.80 Å</td>
      <td>0.80 Å</td>
    </tr>
    <tr>
      <td>5. Search</td>
      <td>Exhaustiveness 32, seed 1</td>
      <td>−13.04</td>
      <td>0.78 Å</td>
      <td>0.78 Å</td>
    </tr>
    <tr>
      <td>5. Search</td>
      <td>Exhaustiveness 32, seed 3</td>
      <td>−9.80</td>
      <td>12.08 Å</td>
      <td>6.56 Å</td>
    </tr>
  </tbody>

Read down the score column and the failures stand out only by comparison. Every wrong top pose scored between −8.99 and −11.40 kcal/mol, and every correct one between −11.54 and −13.71. Without a correct run beside it, −11.40 for the hinge-box pose looks like a strong result, and the reversed pose at −9.7 looks like an ordinary one.

The wrong answers were also not random. The two failed seeds and the 18 Å box returned the same reversed pose, within 0.26 Å of each other, and in seven of the eight correct runs a pose 12 to 13 Å from the crystal sat at rank 2 or 3, 1.9 to 2.9 kcal/mol behind the top one. The reversed pose is the pocket's runner-up, and a setup that weakens the search or cramps the box falls into it.

This is one ligand in one pocket, so it shows how each step can fail rather than how often it does. Imatinib is large and flexible, with 37 heavy atoms, and a small rigid fragment would be more forgiving of every setting here. The public Vina example below uses the same complex with ProteinIQ's own preparation and a box built from the crystal ligand.

![ProteinIQ AutoDock Vina single-ligand example of imatinib docked into c-Abl kinase 1IEP, with five ranked poses, their Vina affinities and RMSD lower bounds, and the energy terms of the top pose](/images/use-cases/screens/autodock-vina-cabl-imatinib-light.webp "**Figure 1. The same complex as a ProteinIQ job.** The public AutoDock Vina example (PDB 1IEP, imatinib, box built from the co-crystallized ligand, seed 42) returns five poses. Its top pose is 0.94 Å from the crystal pose. The RMSD columns compare poses with Vina's own top pose, not with the crystal ligand.")

## Common mistakes at each step

Each step has a usual mistake, and most of them leave a sign that a cheap check can catch (Table 3).

  <caption><strong>Table 3. Docking mistakes by step, their signs and the check that catches them.</strong> Signs and checks drawn from the imatinib test above and the cited benchmarks.</caption>
  <thead>
    <tr>
      <th scope="col">Step</th>
      <th scope="col">Common mistake</th>
      <th scope="col">Sign of trouble</th>
      <th scope="col">Cheap check</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Receptor</td>
      <td>Pocket in the wrong conformation, or a predicted model taken as final</td>
      <td>Scores several kcal/mol weaker than known binders; key contacts missing</td>
      <td>Dock a known ligand first; compare with other PDB entries of the target</td>
    </tr>
    <tr>
      <td>Protein</td>
      <td>No hydrogens, missing side chains, cofactor or metal removed</td>
      <td>Correct pose with a weaker score; donors or the metal contact absent</td>
      <td>Open the prepared file; count polar hydrogens in the pocket</td>
    </tr>
    <tr>
      <td>Ligand</td>
      <td>Wrong protonation, tautomer or stereoisomer; PDB input without bond orders</td>
      <td>Odd donor or acceptor pattern; poses of one compound disagree</td>
      <td>Check the SMILES at the assay pH; enumerate states</td>
    </tr>
    <tr>
      <td>Box</td>
      <td>Too small, off-center or the whole protein</td>
      <td>Top pose reversed or pushed against a box face</td>
      <td>Size the edge to about 2.9 times the ligand's radius of gyration plus margin; center on a known ligand</td>
    </tr>
    <tr>
      <td>Search</td>
      <td>One run at the default exhaustiveness</td>
      <td>Top pose changes between seeds</td>
      <td>Run three seeds; raise exhaustiveness until they agree</td>
    </tr>
    <tr>
      <td>Analysis</td>
      <td>Score read as affinity; only pose 1 inspected</td>
      <td>Ranking contradicts known activities</td>
      <td>Inspect contacts with PLIP; rescore with a second function</td>
    </tr>
    <tr>
      <td>Validation</td>
      <td>Skipped, or redocking only</td>
      <td>No known answer to compare with</td>
      <td>Redock, cross-dock and check poses with PoseBusters</td>
    </tr>
  </tbody>

The habit that covers most of the table is to dock a ligand with a known answer before docking one without. If the setup cannot reproduce imatinib in 1IEP, or your target's crystal ligand in its own structure, no score it gives a new compound means much.

## Molecular docking tools on ProteinIQ

The tools below follow the seven steps in order: fetch and prepare the receptor, prepare the ligand, find the pocket, dock, then analyze and check the poses. [Molecular docking](/use-cases/molecular-docking) and [structure-based virtual screening](/use-cases/structure-based-virtual-screening) show the same workflow for one compound and for a library, and [small-molecule discovery](/solutions/small-molecule) adds the ADMET triage and follow-up simulations around it.

  <caption><strong>Table 4. Tools for each docking step on ProteinIQ.</strong></caption>
  <thead>
    <tr>
      <th scope="col">Task</th>
      <th scope="col">Tool</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Download an experimental receptor structure</td>
      <td>[PDB Download](/app/pdb-download)</td>
    </tr>
    <tr>
      <td>Get a predicted structure when none exists</td>
      <td>[AlphaFold Database](/app/alphafold-database-download), [Boltz-2](/app/boltz-2)</td>
    </tr>
    <tr>
      <td>Rebuild missing atoms and add hydrogens</td>
      <td>[PDBFixer](/app/pdb-fixer)</td>
    </tr>
    <tr>
      <td>Predict residue protonation states</td>
      <td>[PROPKA](/app/propka), [PDB2PQR](/app/pdb2pqr)</td>
    </tr>
    <tr>
      <td>Predict ordered waters</td>
      <td>[SuperWater](/app/superwater)</td>
    </tr>
    <tr>
      <td>Set ligand protonation at a given pH</td>
      <td>[Dimorphite-DL](/app/dimorphite-dl)</td>
    </tr>
    <tr>
      <td>Repair a ligand file that fails preparation</td>
      <td>[Ligand Fixer](/app/ligand-fixer)</td>
    </tr>
    <tr>
      <td>Find candidate pockets</td>
      <td>[fpocket](/app/fpocket)</td>
    </tr>
    <tr>
      <td>Dock into a known site</td>
      <td>[AutoDock Vina](/app/autodock-vina), [smina](/app/smina), [GNINA](/app/gnina)</td>
    </tr>
    <tr>
      <td>Dock without a known site</td>
      <td>[DiffDock](/app/diff-dock)</td>
    </tr>
    <tr>
      <td>Dock with a flexible pocket</td>
      <td>[DynamicBind](/app/dynamic-bind)</td>
    </tr>
    <tr>
      <td>List the contacts of a pose</td>
      <td>[PLIP](/app/plip), [ProLIF](/app/prolif)</td>
    </tr>
    <tr>
      <td>Measure a pose against a reference ligand</td>
      <td>[RMSD Calculator](/app/rmsd-calculator)</td>
    </tr>
    <tr>
      <td>Check that poses are physically valid</td>
      <td>[PoseBusters](/app/posebusters)</td>
    </tr>
    <tr>
      <td>Rescore from simulation snapshots</td>
      <td>[gmx_MMPBSA](/app/gmx-mmpbsa)</td>
    </tr>
  </tbody>

## Frequently asked questions

### What are the steps of molecular docking?

Choose the receptor structure, prepare the protein, prepare the ligand, define the binding site and search box, run the search and scoring, analyze the poses and validate the setup. Only one of the seven is the docking program. In our imatinib test, changing the receptor, the box or the random seed moved the top pose from 0.76 Å to 10 to 12 Å from the crystal pose.

### What are the types of molecular docking?

Docking is grouped by what moves (rigid, flexible-ligand, flexible-receptor and ensemble docking), by what is known about the site (site-directed or blind docking) and by the partner (small molecule, peptide or protein). Learned methods such as DiffDock and co-folding models such as Boltz-2 form a newer group that predicts the pose without a classical search.

### How do you prepare a protein for docking?

Keep the chains that form the pocket, resolve alternate conformations, rebuild missing side chains, decide which waters, ions and cofactors to keep, add hydrogens with protonation states for the assay pH and write the program's input format. Skipping preparation steps systematically lowered virtual screening enrichment in a Glide study of 36 targets.

### Should I remove water molecules before docking?

Remove bulk water, but consider keeping waters that bridge the ligand and protein. Keeping such waters raised docking accuracy significantly in a study by Roberts and Mancera. Crystal waters from a ligand's own structure fit that ligand by construction, so for new ligands keep only waters conserved across structures.

### How big should the docking box be?

For Vina, Feinstein and Brylinski found pose accuracy highest with a cubic edge of 2.857 times the ligand's radius of gyration, tested on 3,659 complexes. Treat that as a minimum for long ligands. The rule gives 19.1 Å for imatinib, yet an 18 Å cube failed in our test and a 25 Å cube worked.

### Is AlphaFold good enough for docking?

Often not on its own. Against AlphaFold2 models, AutoDock-GPU placed the top pose within 2 Å for 17% of 2,474 complexes, against 41% for crystal structures. A confident model can still show the wrong pocket state, as the ABL1 model did for imatinib in our test.
