# AlphaFold docking vs co-folding: which to use for a ligand

> Docking into an AlphaFold model works when the model shows the pocket in the state your ligand binds, and fails when it does not, however high its pLDDT. Co-folding with AlphaFold 3 or Boltz-2 removes that choice but leans on training data. See what each route predicts, an ABL1 kinase test with two inhibitors, and how to read Boltz-2 binding affinity.

## Should you dock into an AlphaFold model or co-fold the ligand?

Dock into a predicted structure when you have reason to think the model shows the pocket in the state your ligand binds. Co-fold the ligand when you do not, and read the result as a hypothesis that is only as good as the training data behind it. Neither route measures binding affinity.

One AlphaFold Database model of the kinase ABL1 shows how a confident model can serve one ligand and fail another. The 32 residues around imatinib average a pLDDT of 90.9, and the model's DFG phenylalanine, Phe382, sits within 0.4 Å of where the dasatinib complex has it, so no atom of the model lies within 2.5 Å of crystal dasatinib.

Imatinib binds a different state of the same kinase, and against it Phe382 is 8.5 Å out of place and 20 atoms of the model lie within 2.5 Å of the ligand. In our Vina runs the model failed imatinib, as the overlap predicts, but the program also failed in the crystal receptors, so the overlap count was the more reliable signal.

Co-folding models such as AlphaFold 3, [Boltz-2](/app/boltz-2), [Chai-1](/app/chai-1) and [OpenFold3](/app/openfold-3) avoid the choice of receptor, because they predict the protein and the ligand together. They pay for it elsewhere: their accuracy follows how close the complex is to something in the Protein Data Bank, and their affinity output is a ranking score, not a measurement. What decides the route is whether you know the pocket state, how novel the chemistry is, and whether you need a pose, a hit list or a ranking.

## What docking into an AlphaFold model assumes

A predicted structure gives [docking](/guides/molecular-docking-steps) one conformation chosen without knowledge of your ligand. AlphaFold2 predicts a protein as though nothing were bound to it, so pocket side chains and loops sit where they are most common in the training structures, whatever your molecule needs.

Benchmarks of the gap agree. Holcomb and colleagues redocked 2,474 PDBbind complexes with [AutoDock-GPU](/app/autodock-gpu): the top pose was within 2 Å for 41% of complexes against the crystal receptor and 17% against the AlphaFold2 model, so 83 of 100 predictions missed. The model's confidence score did not predict docking success, although it did track how well the Cα atoms matched the experiment.

Karelina and colleagues found the same on GPCRs with Glide. The median pocket error was 1.3 Å for AlphaFold2 models against 3.3 Å for traditional homology models, yet pose accuracy was 15% against 9%, not significantly different, and 44% on experimental structures solved with other ligands. The backbone is nearly right and the side chains decide.

Screening is more forgiving than pose prediction. Zhang and colleagues ran Glide on 27 DUD-E targets where the model could be refined. The early enrichment of known actives (EF1%) averaged 13.0 on the AlphaFold2 models, 11.4 on apo crystal structures and 24.2 on holo structures, so the model recovered about half of what a ligand-bound structure gave. Refining the model around a known ligand with induced-fit docking raised it to 18.9, which closes about half of the remaining gap.

Prospective tests point the same way. Lyu and colleagues docked large libraries against unrefined AlphaFold2 models of the σ2 and 5-HT2A receptors and tested hundreds of new molecules; hit rates and affinities were high and similar to docking into the experimental structures. Retrospective pose benchmarks and prospective screens can both be true, because a screen only needs the model to admit some ligands of some chemotype, while a pose benchmark needs it to admit one specific ligand.

Two habits follow. Check the pocket against any crystal structure of the target before docking, since pLDDT does not report which conformational state the model chose. Where several states exist, dock into more than one, as [ensemble docking](/use-cases/ensemble-docking) does. The [structure prediction confidence guide](/guides/structure-prediction-confidence) shows why [pLDDT](/guides/plddt) and [PAE](/guides/pae) cannot tell you this.

## What co-folding does differently

A co-folding model takes the protein sequence and the ligand as input and predicts the complex in one pass. There is no search box and no receptor structure to choose, and the pocket takes the shape the model associates with that ligand. AlphaFold 3 brought the approach to small molecules, and its authors report that it greatly outperformed classical docking tools such as Vina on the PoseBusters benchmark of 428 complexes released in 2021 or later.

That benchmark is a favorable case, and the cost shows up when the complex is unlike anything in training. Škrinjar and colleagues built Runs N' Poses, 2,600 high-resolution protein-ligand systems released after the 30 September 2021 training cutoff, and sorted them by similarity to training complexes. AlphaFold 3, Boltz-1, Chai-1 and Protenix produced an accurate pocket in more than 90% of predictions, yet the authors conclude that the methods "largely memorize ligand poses from their training data". Finding the site is not the hard part. Placing a new ligand in it is.

Masters and colleagues tested whether the models respond to the chemistry. They replaced the 11 residues around ATP in CDK2 with glycine, removing every side chain that could hold the ligand, and all four co-folding models still predicted the same ATP binding mode. On the 285 CASF-2016 complexes, more than half of AlphaFold 3's correct predictions kept their pose after the binding site was disrupted. A pose that survives the loss of its contacts was not computed from them.

Head-to-head studies do not agree on a winner, and they depend on what the target looks like. On the 1,312 cross-docking entries of PoseX, the top pose was within 2 Å for 68.6% with AlphaFold 3, 77.0% with the learned docking model [SurfDock](/app/surfdock) and 54.1% with GNINA. On 489 evaluable Mac1 complexes, AlphaFold 3 and Chai-1 predicted about 70% of ligands within 2 Å, Boltz-2 53% and the docking program DOCK3.7 40%.

In a prospective σ2 screen, docking gave 27 hits among 108 molecules (25%) and AlphaFold 3 gave 15 among 114 (13%), according to a preprint. In these studies co-folding led on poses for well-covered chemistry, and docking led on the one prospective screen.

Co-folded poses need the checks that [learned methods](/use-cases/ai-molecular-docking) need. A pose can sit in the right place with distorted bonds or clashes, which the [PoseBusters tool](/app/posebusters) tests. AlphaFold 3 is also not open to everyone: its weights carry non-commercial terms, set out in the [AlphaFold license guide](/guides/alphafold-license-commercial-use), so the open examples here are Boltz-2 and Chai-1.

## One AlphaFold model, two kinase inhibitors

We docked imatinib and dasatinib into three receptors of human ABL1: the crystal structure of the ligand's own complex, the crystal structure of the other ligand's complex, and the AlphaFold Database model. Imatinib binds an inactive conformation of the kinase and dasatinib an active-like one, so each ligand meets one receptor it was crystallized in, one it was not and the model. The settings are in the caption of Table 1.

  <caption><strong>Table 1. Imatinib and dasatinib docked into three ABL1 receptors.</strong> AutoDock Vina 1.2.7, Vina scoring, exhaustiveness 8, nine poses, three seeds (42, 1, 2), 25 Å cube centered on the crystal ligand, ligand built from SMILES, receptors protonated with [Open Babel](/app/open-babel). Each receptor was superposed on the ligand's own crystal complex over the Cα atoms of the residues within 6 Å of that ligand (32 for imatinib, 27 for dasatinib). RMSD is symmetry-corrected over heavy atoms with no fitting. Structures: 1IEP (imatinib, 2.1 Å), 2GQG (dasatinib, 2.4 Å), chain A of each, and AF-P00519-F1.</caption>
  <thead>
    <tr>
      <th scope="col">Ligand</th>
      <th scope="col">Receptor</th>
      <th scope="col">Site Cα RMSD</th>
      <th scope="col">Receptor atoms within 2.5 Å of the crystal ligand</th>
      <th scope="col">Top Vina score, three seeds</th>
      <th scope="col">Nearest of nine poses to the crystal pose</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Imatinib</td>
      <td>1IEP, own complex</td>
      <td>0 Å</td>
      <td>0</td>
      <td>−9.4 to −10.6 kcal/mol</td>
      <td>6.5 to 9.8 Å</td>
    </tr>
    <tr>
      <td>Imatinib</td>
      <td>2GQG, dasatinib complex</td>
      <td>1.42 Å</td>
      <td>18</td>
      <td>−8.7 to −9.8 kcal/mol</td>
      <td>7.8 to 8.0 Å</td>
    </tr>
    <tr>
      <td>Imatinib</td>
      <td>AlphaFold model</td>
      <td>1.96 Å</td>
      <td>20</td>
      <td>−9.0 to −9.1 kcal/mol</td>
      <td>9.5 to 10.3 Å</td>
    </tr>
    <tr>
      <td>Dasatinib</td>
      <td>2GQG, own complex</td>
      <td>0 Å</td>
      <td>0</td>
      <td>−10.9 kcal/mol</td>
      <td>1.5 to 2.0 Å</td>
    </tr>
    <tr>
      <td>Dasatinib</td>
      <td>1IEP, imatinib complex</td>
      <td>1.69 Å</td>
      <td>1</td>
      <td>−7.7 to −8.4 kcal/mol</td>
      <td>1.9 to 3.5 Å</td>
    </tr>
    <tr>
      <td>Dasatinib</td>
      <td>AlphaFold model</td>
      <td>1.50 Å</td>
      <td>0</td>
      <td>−8.2 to −9.0 kcal/mol</td>
      <td>5.2 to 9.7 Å</td>
    </tr>
  </tbody>

The pocket overlap in Table 1 needs no docking and is the clearest result. The model and the dasatinib complex hold Phe382 in the same place, within 0.4 Å, so nothing of the model lies within 2.5 Å of dasatinib. For imatinib, 20 atoms of the model from five residues (Ile360, Phe382, Leu384, Ser385 and Arg386) sit inside that distance, and so do 18 atoms of the dasatinib structure. The model is not a bad structure; it shows the state dasatinib binds.

Docking agrees with the overlap for imatinib and then shows a problem with the test. In the two receptors that crowd imatinib, no pose of the nine came within 7.8 Å of the crystal pose in any run. But imatinib failed in its own crystal receptor too, with the nearest pose 6.5 to 9.8 Å away in the three seeds above. Two further seeds on 1IEP gave one top pose at 0.79 Å and one miss, so a correct answer exists and Vina found it in one of five runs.

The dasatinib rows say two other things. Into its own structure the top pose was 5.1 Å from the crystal pose in all three seeds, although a pose within 1.5 to 2.0 Å was among the nine. Into the model, which has no atom near the ligand, the nearest pose was 5.2 to 9.7 Å away and the best score was 1.9 to 2.7 kcal/mol weaker than in the crystal receptor; exhaustiveness 32 on one seed gave a top score of −9.19 kcal/mol and a nearest pose 4.8 Å away. The top-ranked pose was within 2 Å in none of the 18 runs, so no receptor won on that measure.

Two lessons follow. A clean overlap check is necessary and not sufficient, because the model has no clash with dasatinib and still did not reproduce its pose. And a receptor comparison means little when the control fails: [redock the crystal ligand](/guides/redocking-docking-validation) first, and compare receptors only on a ligand the setup can redock. Here the cheap overlap count predicted the imatinib failure, and the docking runs alone could not have separated the receptors.

Co-folding sidesteps this model's state, since the ligand conditions the pocket, but ABL1 is a poor place to test it. The imatinib and dasatinib complexes were deposited in 2002 and 2006, long before the 30 September 2021 cutoff that the Runs N' Poses benchmark uses for training data. A co-folded pose that matches them would show recall as easily as prediction, and the result on a new ABL1 inhibitor is the one worth running.

## What Boltz-2 binding affinity means

Boltz-2 adds an affinity head to co-folding, which returns two numbers for one protein and one small molecule. The binary probability is meant for separating binders from decoys at the hit-discovery stage. The affinity value is a predicted log10(IC50), with IC50 in micromolar, meant for comparing active molecules during hit-to-lead work, and the documentation says it should not be used on inactives.

The authors describe it as a measure of binding strength that supports ranking and "can be approximately interpreted as an IC50-like value", trained on a mixture of Ki, Kd, IC50, AC50, EC50 and XC50 measurements.

Converting it into the reader's units shows what kind of number it is. The public Boltz-2 example of KRAS G12C with a covalent inhibitor returns an affinity of −0.401, which is an IC50 of 10<sup>−0.401</sup> µM, about 0.40 µM. The repository gives the conversion to an energy as (6 − y) × 1.364 kcal/mol, which makes this 8.7 kcal/mol. One log unit is 1.364 kcal/mol and a tenfold change in IC50, so an error of one unit moves a compound from 0.4 µM to 4 µM.

How often the ranking is right depends on the target. On the 4-target FEP+ subset the authors report an average Pearson correlation of 0.66 with experiment, but on eight internal blinded assays the correlation exceeded 0.55 for three and was limited for the other five.

In the independent Mac1 study, the predicted pIC50 correlated at r = 0.6 with measurements for 202 compounds, against 0.31 for AlphaFold 3 confidence and −0.23 for DOCK3.7. The preprint comparing AlphaFold 3 with docking on σ2 reports that using Boltz-2 to filter the 222 tested compounds raised the hit rate from 19% to 30% at a binding probability above 0.8, and to 44% at a predicted log10(IC50) of −1.

These are respectable numbers for a ranking tool, and not a replacement for a measurement. They sit near the figures in the [docking score guide](/guides/docking-score-binding-energy-affinity), where a scoring function that picks the right pose for 90.2% of CASF-2016 complexes correlates with measured affinity at R = 0.604, but the sets differ and the two cannot be compared directly.

The limits that matter in practice are stated by the Boltz authors: affinity is unreliable if the model misses the pocket, the module does not handle cofactors, ions, water or multimeric partners, and ligands much larger than 56 atoms are outside its training range. Calibrate on known compounds from your own series before ranking new ones.

## Choosing a route

Table 2 turns the evidence into a starting point. Whatever the route, a second route that fails differently is the cheapest check, which is the idea behind [consensus docking](/use-cases/consensus-docking).

  <caption><strong>Table 2. Which route to start with.</strong> Starting points drawn from the studies above; none is a rule.</caption>
  <thead>
    <tr>
      <th scope="col">Your situation</th>
      <th scope="col">Start with</th>
      <th scope="col">Add</th>
      <th scope="col">Check</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>An experimental structure with a ligand like yours</td>
      <td>Dock into it</td>
      <td>A co-folded pose as an independent second opinion</td>
      <td>Redock the crystal ligand, then PoseBusters</td>
    </tr>
    <tr>
      <td>Only an AlphaFold model, and a known ligand of the same target</td>
      <td>Superimpose that complex on the model and count overlapping atoms; dock if there are none</td>
      <td>Co-fold the ligand to see whether the pocket changes</td>
      <td>Pocket residues against every crystal state of the target</td>
    </tr>
    <tr>
      <td>No structure and no known site</td>
      <td>Co-fold, or blind docking into the model</td>
      <td>A pocket finder on the model</td>
      <td>Whether several seeds agree on one site</td>
    </tr>
    <tr>
      <td>A new chemotype with no close relatives in the PDB</td>
      <td>Docking into several receptor conformations</td>
      <td>A co-folded pose as a hypothesis</td>
      <td>How similar your complex is to PDB entries from before 2021</td>
    </tr>
    <tr>
      <td>A large library to screen</td>
      <td>Docking, for cost and hit rate</td>
      <td>Co-folding or Boltz-2 binding probability on the top hits</td>
      <td>PoseBusters, then contacts against the known pharmacology</td>
    </tr>
    <tr>
      <td>Ranking a congeneric series</td>
      <td>Boltz-2 affinity, calibrated on measured members</td>
      <td>A physics-based method if the series matters</td>
      <td>Correlation on the compounds you already know</td>
    </tr>
  </tbody>

Record the structure, the model version and the seed with every pose, since both routes are stochastic. A co-folded pose that disagrees with a docked one is information about which assumption is shaky, not a vote to be settled by confidence scores.

## Docking and co-folding tools on ProteinIQ

Fetch or predict the receptor first, dock or co-fold second, and validate last (Table 3).

  <caption><strong>Table 3. Tools for docking into predicted structures and co-folding ligands.</strong></caption>
  <thead>
    <tr>
      <th scope="col">Task</th>
      <th scope="col">Tool</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Get a predicted structure of a known protein</td>
      <td>[AlphaFold Database download](/app/alphafold-database-download)</td>
    </tr>
    <tr>
      <td>Predict a structure from a sequence</td>
      <td>[AlphaFold2](/app/alphafold-2)</td>
    </tr>
    <tr>
      <td>Download experimental structures to compare pocket states</td>
      <td>[PDB download](/app/pdb-download)</td>
    </tr>
    <tr>
      <td>Find candidate pockets on a model</td>
      <td>[fpocket](/app/fpocket)</td>
    </tr>
    <tr>
      <td>Dock into a chosen receptor and box</td>
      <td>[AutoDock Vina](/app/autodock-vina), [GNINA](/app/gnina)</td>
    </tr>
    <tr>
      <td>Dock without a defined box</td>
      <td>[DiffDock](/app/diff-dock)</td>
    </tr>
    <tr>
      <td>Co-fold a protein with a ligand</td>
      <td>[Boltz-2](/app/boltz-2), [Chai-1](/app/chai-1), [OpenFold3](/app/openfold-3), [Protenix](/app/protenix)</td>
    </tr>
    <tr>
      <td>Estimate binding strength for related ligands</td>
      <td>[Boltz-2](/app/boltz-2)</td>
    </tr>
    <tr>
      <td>Check chemistry, geometry and clashes of a pose</td>
      <td>[PoseBusters](/app/posebusters)</td>
    </tr>
    <tr>
      <td>List the contacts a pose makes</td>
      <td>[PLIP](/app/plip)</td>
    </tr>
  </tbody>

## Frequently asked questions

### Can you dock into an AlphaFold structure?

Yes, but pose accuracy drops. In 2,474 PDBbind complexes, AutoDock-GPU placed the top pose within 2 Å for 17% of AlphaFold2 models against 41% of crystal receptors. Screening holds up better: unrefined AlphaFold2 models gave hit rates similar to experimental structures for the σ2 and 5-HT2A receptors.

### Is AlphaFold 3 better than docking?

For the pose of a single ligand on well-covered chemistry, often yes: AlphaFold 3 put 68.6% of PoseX cross-docking ligands within 2 Å against 54.1% for GNINA. In a prospective σ2 screen, docking gave a 25% hit rate against 13% for AlphaFold 3, reported in a preprint. Neither result removes the need to check the pose.

### What is co-folding?

Co-folding predicts a protein and its ligand, nucleic acid or partner chains together from sequence and chemistry alone, with no receptor structure or search box. AlphaFold 3, Boltz-2, Chai-1 and OpenFold3 work this way, and their ligand accuracy falls as the complex becomes less similar to the training data.

### Does co-folding understand the physics of binding?

Evidence says only partly. When Masters and colleagues replaced the 11 residues around ATP in CDK2 with glycine, all four models still predicted the same ATP binding mode. Treat a co-folded pose as a hypothesis, and check it against known pharmacology or a docked pose.

### How accurate is Boltz-2 binding affinity?

It depends on the target. The Boltz-2 authors report an average Pearson correlation of 0.66 on a 4-target FEP+ subset and limited performance on five of eight internal blinded assays; an independent Mac1 study found r = 0.6 for 202 compounds. One log unit of predicted IC50 is 1.364 kcal/mol, so use it to rank related compounds, not to quote an IC50.

### Should I dock into the AlphaFold model or the co-folded structure?

Dock into whichever shows the pocket in the state your ligand binds, and check that by superimposing a crystal complex. A co-folded structure already has a ligand in it, so redocking that ligand will look successful whether or not the pocket is right.
