Binding Pocket Prediction on AlphaFold Models vs Experimental Structures

Binding pocket prediction works on AlphaFold and ESMFold models, but the answer is only as reliable as the residues that line the pocket. A ligand-free predictor such as P2Rank (Krivák & Hoksza, J. Cheminform. 2018) reads geometry and chemistry from the structure you give it. On an experimental (PDB) structure, that geometry was measured. On a predicted structure it was inferred. In a community assessment of AlphaFold2 applications, pockets built from high-confidence residues (mean pLDDT > 90) overlapped known binding sites as well as pockets from experimental structures did, while low-confidence pockets generally did not overlap known sites (Akdel et al., Nat. Struct. Mol. Biol. 2022). So the practical rule is: use a predicted structure to locate a pocket, check the pLDDT of every pocket-lining residue, and prefer an experimental structure whenever one exists for the docking step.

This article picks up where ESMFold vs AlphaFold2 leaves off. That piece covers how to get a structure. This one covers what happens when you ask that structure where a drug could bind.

What P2Rank Does (and What It Does Not)#

P2Rank is a template-free, machine-learning binding site predictor. It places points on the protein's solvent-accessible surface, predicts the "ligandability" of the local chemical neighbourhood around each point, and clusters high-scoring points into ranked pockets (Jendele et al., Nucleic Acids Res. 2019). It needs only 3D coordinates: no bound ligand, no homologous template and no sequence database. The original paper reports that it outperforms Fpocket, SiteHound, MetaPocket 2.0 and DeepSite, and needs under one second per protein (Krivák & Hoksza, 2018).

What P2Rank returns is a ranked list of candidate sites, each with a score and a calibrated probability, plus the residues that line it. It does not return:

• Binding affinity. A high-probability pocket says nothing about how tightly any particular molecule binds there.
• A ligand pose. That is a docking question.
• A guarantee that the pocket exists in solution. It scores the conformation you submitted, and nothing else.

Benchmark accuracy is measured with the DCA criterion: a prediction counts as a hit when the centre of a predicted site lies within 4 Å of any ligand atom. In the PrankWeb 3 evaluation, the current default model identified the ligand site among its top-n predictions (where n is the number of ligands in the structure) in 71.6% of cases on COACH420 and 72.7% on HOLO4K (Jakubec et al., Nucleic Acids Res. 2022). Both of those datasets are experimental structures. Neither number is a measured accuracy on AlphaFold or ESMFold models.

Why a Predicted Structure Changes the Pocket#

AlphaFold2 (Jumper et al., Nature 2021) and ESMFold (Lin et al., Science 2023) are judged mainly on how well they place the backbone. A pocket predictor reads something finer: which side chains point into a cavity, how deep that cavity is, and which surface atoms are exposed. Four effects follow from that gap.

1. Side-chain placement#

A backbone that is right to within an ångström can still carry side chains that narrow, widen or block a pocket. In a docking benchmark of AlphaFold models against 22 targets, the authors found that "even on very accurate models, small side-chain variations impact the performance" (Scardino et al., iScience 2023). Their per-target notes read like a pocket-prediction checklist: side chains "slightly pulled inside the pocket, narrowing the binding site", and in one target a pocket blocked by a disordered N-terminal loop.

2. Low-confidence (low-pLDDT) regions#

pLDDT is the per-residue confidence score from both models. The .bloom* ESMFold results guide uses these bands:

Low-confidence loops and tails are where predicted structures differ most from reality, and they can manufacture surface concavities that look like pockets. Akdel et al. found that low-quality AlphaFold2 pockets (mean pLDDT ≤ 90) were larger on average, included particularly large pockets, and generally did not overlap known sites. They suggest that the mean pLDDT of a pocket's residues can be used as an extra criterion for pocket selection (Akdel et al., 2022).

3. Apo versus holo conformations#

An experimental structure solved with a ligand (holo) shows the pocket in its bound shape. One solved without a ligand (apo) may show the same pocket partly closed. A predicted structure is neither by construction. For GPCRs, AlphaFold2 binding pockets came out at a median all-atom RMSD of 1.3 Å against experimental structures, compared with 3.3 Å for traditional template-based models. That error is nearly as small as the difference between two experimental structures of the same receptor with different ligands bound (Karelina et al., eLife 2023). A pocket that is close to holo is enough to find the site. As the docking section below shows, it is often not close enough to place a ligand.

4. Cryptic pockets#

Some binding sites are absent from the ground-state structure and open only on ligand binding or through protein motion. A predictor that scores one static conformation can't see a site that isn't open in it. That holds for P2Rank on any input, and the P2Rank docs list it as a known limitation. Cryptic pockets need a different method: PocketMiner, a graph neural network trained on molecular dynamics, identified experimentally confirmed cryptic pockets with a ROC-AUC of 0.87 on a curated set of 39. Applied across the human proteome, it suggests that over half of proteins thought to lack pockets probably contain cryptic ones (Meller et al., Nat. Commun. 2023).

P2Rank's Two Models: default vs alphafold#

P2Rank ships two prediction profiles, and .bloom* runs exactly those two (P2Rank inputs). The difference is narrower than the name suggests, and it matters for the reason that follows.

Crystallographic structures store atomic B-factors (a measure of positional uncertainty) in the B-factor column of the PDB file, and the default model uses that column as a feature. AlphaFold models put the per-residue pLDDT in the same column, and so do ESMFold models: the ESMFold results guide notes that structure.pdb carries pLDDT in the B-factor field. A model that reads that column as a B-factor is reading confidence scores as if they were thermal motion.

The PrankWeb 3 team therefore trained B-factor-free models for AlphaFold input (Jakubec et al., 2022). The P2Rank 2.4 alphafold profile is described upstream as the same as the default profile except that it doesn't use B-factor as a feature, and it is recommended for AlphaFold, NMR and cryo-EM structures.

Success rates are from PrankWeb 3, Table 2 (Jakubec et al., 2022), using the 4 Å DCA criterion and the top-n predictions. Note that the two models score about the same on experimental structures. The alphafold model is not a correction for an inaccurate prediction. It stops the model from misreading pLDDT as a B-factor. If the predicted side chains or loops are wrong, both models will score the wrong shape.

Predicted vs Experimental Structures for Pocket Prediction#

How to Read a Pocket Score#

P2Rank writes a raw score and a 0–1 probability for each pocket. On .bloom*, the pocket_score in the result table is that calibrated probability, and pockets are ranked by it (P2Rank outputs). Read it as "how much does this surface region look like the ligand-binding sites P2Rank was trained on". It is not:

• Druggability. .bloom* adds a druggability label (druggable, intermediate, difficult, undruggable). That label is a heuristic computed from the pocket's P2Rank probability and residue composition. It isn't a P2Rank output or a measured druggability, and the outputs guide notes that fragments, covalent binders and PPI inhibitors can bind pockets labelled difficult.
• Affinity. Two pockets with the same probability can bind the same ligand with very different strength, or not at all.
• Biological relevance. The troubleshooting guide says it directly: P2Rank ranks by probability, not by biological importance, so a known active site may rank second or third.

On a predicted structure, add one more check before you trust a score: look up the pLDDT of every residue that lines the pocket. A high-probability pocket formed by low-confidence residues is a hypothesis about a loop, not a binding site.

What Happens When You Dock Into a Predicted Pocket#

Finding the right pocket and placing a ligand in it correctly are different tasks, and predicted structures do much better at the first. Three independent benchmarks agree:

• Redocking across PDBbind. On 2,474 complexes, AutoDock-GPU placed the ligand within 2 Å RMSD in 41% of cases against the crystal structure and 17% against the AlphaFold2 model. The model's own quality measure was "not a good predictor of docking performance". Removing low-confidence regions and making side chains flexible improved the results (Holcomb et al., Protein Sci. 2022).
• GPCRs. Even though AlphaFold2 binding pockets were close to experimental ones, docking to them gave correct poses for 15% of ligands. That was not significantly better than traditional homology models (9%), and far below docking to experimental structures solved with a different ligand (44%) (Karelina et al., 2023).
• Virtual screening. Across 22 targets and four docking programs, high-throughput docking on unrefined AlphaFold models was consistently worse than on the matching PDB structures (Scardino et al., 2023).

The lesson for a pipeline is that a predicted structure is good evidence for

where a site is, and weak evidence for how a ligand sits in it.

When to Use Each#

Use an experimental structure when:#

• One exists for your target. Search the PDB first (PDB on .bloom*); a deposited structure is the highest-confidence input.
• You are going to dock. Pose prediction is where predicted structures lose the most (see above). A holo structure, even with a different ligand, is the strongest starting point.
• You need to compare variants or orthologs. Measured side chains make pocket differences interpretable.

Use a predicted structure when:#

• No experimental structure exists, or the deposited ones cover only part of the protein. Fetch the AlphaFold DB model if the protein is catalogued (Varadi et al., Nucleic Acids Res. 2023), or predict it with ESMFold if the sequence is novel, engineered or mutated.
• You are triaging targets. Asking "does this protein have a plausible small-molecule site at all?" across many proteins is where predicted structures pay off.
• You will check pLDDT. Restrict conclusions to pockets lined by confident residues (pLDDT > 70 as a minimum, > 90 where it matters).

Limitations#

• One conformation in, one answer out. P2Rank scores the structure as submitted. Cryptic and ligand-induced pockets are out of reach unless you supply a conformation in which they are open.
• Published accuracy is on experimental structures. The DCA success rates above come from crystal-structure benchmarks. On a predicted structure, your own validation (does the top pocket match a known site of a homolog?) is the only accuracy estimate you have.
• No conservation features on .bloom*. PrankWeb also offers models that add evolutionary conservation (Jakubec et al., 2022). The .bloom* tool runs only default and alphafold.
• Druggability labels are heuristic. Treat them as a sorting aid, not a verdict.
• Disordered regions have no pockets. A region below pLDDT 50 should be trimmed or ignored, not interpreted.

How to Run It on .bloom*#

On .bloom*, the docking arc is pocket → pose → rescore: P2Rank (p2rank_predict), then DiffDock (diffdock_dock), then MM-GBSA (mmgbsa_calculate). Put a fold step in front of it when you have no structure. Ask for the whole chain in one request and Ivy plans the run and chains the tools (how orchestration works), or run each step on its own:

1. Get the structure. Give P2Rank a PDB ID or a UniProt accession in the same field. For a UniProt accession, the tool picks an experimental structure first and falls back to the AlphaFold DB model, and it sets the P2Rank model to match the source (default for experimental, alphafold for AlphaFold DB). You can also bring a structure from a previous job: download it from the PDB, fetch it from AlphaFold DB, or predict it with ESMFold.
2. Pick the model for structures you supply. When you pass a structure in directly, the model is yours to set. Use alphafold for an ESMFold or AlphaFold model and default for a crystal structure. max_pockets (1–50, default 10) caps the list.
3. Read the pocket table. Each pocket comes with its rank, probability, lining residues and a druggability label, plus a result card with a 3D view, a PyMOL script, and CSV/JSON downloads (P2Rank outputs).
4. Check pLDDT under the pocket. For a predicted structure, compare the pocket's residue list with the per-residue pLDDT. Treat pockets lined by residues below 70 as unconfirmed. If a low-confidence tail sits over the site, trim it and run again (troubleshooting).
5. Dock, then judge the pose against the pocket. DiffDock on .bloom* runs blind (whole-protein) docking: it takes no box or residue list, so the P2Rank pocket doesn't steer the search. Use the pocket to judge whether the returned poses land where you expect (DiffDock inputs).
6. Rescore the poses with MM-GBSA. It gives a physics-based energy estimate per pose, which DiffDock's confidence ranking doesn't. Read it as a ranking signal, not a measured affinity.

If you have a predicted structure and an experimental one for the same target, run P2Rank on both. Agreement on the top pocket is useful evidence. Disagreement usually points back to the low-confidence region or the apo/holo difference discussed above.

Key Takeaways#

• Pocket predictors such as P2Rank run fine on AlphaFold and ESMFold models. Pockets lined by high-confidence residues (mean pLDDT > 90) match known sites about as well as pockets from experimental structures; low-confidence pockets generally do not (Akdel et al., 2022).
• Use P2Rank's alphafold model on predicted structures. It ignores the B-factor column, which in predicted models holds pLDDT instead of a B-factor. It doesn't make an inaccurate model accurate.
• A pocket probability ranks candidate sites. It is not druggability, affinity or biological importance.
• Predicted structures are much better at locating a site than at supporting pose prediction: 17% vs 41% redocking success against crystal structures in one large benchmark (Holcomb et al., 2022).
• Prefer an experimental structure for docking whenever one exists. Use predicted structures for coverage and target triage.

Next step: once you have a pocket, choosing the docking method is the next decision. See DiffDock vs GNINA vs AutoDock Vina.

Run P2Rank on your target, with pLDDT checked under every pocket, on the .bloom* platform.

References#

1. Krivák, R. & Hoksza, D. "P2Rank: machine learning based tool for rapid and accurate prediction of ligand binding sites from protein structure." Journal of Cheminformatics 10, 39 (2018). doi:10.1186/s13321-018-0285-8
2. Jendele, L., Krivák, R., Škoda, P., Novotný, M. & Hoksza, D. "PrankWeb: a web server for ligand binding site prediction and visualization." Nucleic Acids Research 47, W345–W349 (2019). doi:10.1093/nar/gkz424
3. Jakubec, D., Škoda, P., Krivák, R., Novotný, M. & Hoksza, D. "PrankWeb 3: accelerated ligand-binding site predictions for experimental and modelled protein structures." Nucleic Acids Research 50, W593–W597 (2022). doi:10.1093/nar/gkac389
4. Jumper, J., Evans, R., Pritzel, A. et al. "Highly accurate protein structure prediction with AlphaFold." Nature 596, 583–589 (2021). doi:10.1038/s41586-021-03819-2
5. Lin, Z., Akin, H., Rao, R. et al. "Evolutionary-scale prediction of atomic-level protein structure with a language model." Science 379, 1123–1130 (2023). doi:10.1126/science.ade2574
6. Varadi, M., Bertoni, D., Magana, P. et al. "AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences." Nucleic Acids Research 52, D368–D375 (2023). doi:10.1093/nar/gkad1011
7. Akdel, M., Pires, D.E.V., Porta Pardo, E. et al. "A structural biology community assessment of AlphaFold2 applications." Nature Structural & Molecular Biology 29, 1056–1067 (2022). doi:10.1038/s41594-022-00849-w
8. Scardino, V., Di Filippo, J.I. & Cavasotto, C.N. "How good are AlphaFold models for docking-based virtual screening?" iScience 26, 105920 (2023). doi:10.1016/j.isci.2022.105920
9. Karelina, M., Noh, J.J. & Dror, R.O. "How accurately can one predict drug binding modes using AlphaFold models?" eLife 12, RP89386 (2023). doi:10.7554/eLife.89386
10. Holcomb, M., Chang, Y.-T., Goodsell, D.S. & Forli, S. "Evaluation of AlphaFold2 structures as docking targets." Protein Science 32, e4530 (2022). doi:10.1002/pro.4530
11. Meller, A., Ward, M., Borowsky, J. et al. "Predicting locations of cryptic pockets from single protein structures using the PocketMiner graph neural network." Nature Communications 14, 1177 (2023). doi:10.1038/s41467-023-36699-3