AlphaFold for docking: when the predicted structure is usable

AlphaFold has unblocked targets that had no crystal structure for decades. It has also generated a steady stream of plausible-looking models that fail silently when you run a docking screen against them. The distinction matters for any program trying to use AlphaFold to replace, rather than supplement, experimental structure. Here is what the validation literature actually shows, and the pre-flight checks that flag a model as risky before you commit a screen to it.

What AlphaFold gets right (and how well)

AlphaFold2 (Jumper et al., Nature 2021) produces backbone coordinates that are, on average, within experimental resolution for well-folded globular domains. The original CASP14 results put median backbone RMSD around 1 Å for domains where the model was confident (per-residue pLDDT > 90). AlphaFold3 (Abramson et al., Nature 2024) extends the same architecture to protein-ligand and protein-nucleic-acid complexes, with reported success on PoseBusters v1 above the level of conventional docking pipelines.

For a docking user the practical translation: if your binding site sits inside a high-pLDDT region (call it > 85 across every pocket residue), the backbone you’re docking against is most likely a faithful representation of the underlying fold. Most kinase and GPCR cores fall into this regime. Many active-site loops, post-translational-modification sites, and intrinsically flexible regions do not.

Where it falls down — the failure modes that matter for docking

• Side-chain rotamers in the pocket. Karelina, Noh and Dror (eLife 2023) docked against AlphaFold2 models for a panel of GPCRs and kinases. Backbone was generally fine; pocket side chains were often in the wrong rotamer relative to the holo (ligand-bound) crystal. The result was that pose recovery was significantly worse than against crystal structures, even when the backbone RMSD was < 1.5 Å. Side-chain prediction is a different problem than backbone prediction, and a 90+ pLDDT does not promise the rotamer is right.
• Apo-like conformations. AlphaFold tends to produce one conformational state, and it is trained on a database dominated by apo and holo crystal structures with no preference signal between them. For targets where the pocket opens on ligand binding (induced fit), the predicted structure may show a closed pocket that no inhibitor can reach. Díaz-Rovira et al. (J Chem Inf Model 2023) reported degraded virtual-screening performance against AlphaFold structures relative to holo crystals on a DUD-E benchmark, with closed pockets a frequent cause.
• Disordered loops near the active site. Low-pLDDT loops can be physically real (the protein is disordered there) or a confidence artifact. In either case, docking against a poorly-modeled loop next to the pocket pollutes the pose with spurious clashes or false contacts. The mitigation is usually a short MD relaxation of the loop, or — better — choosing a template that doesn’t have a disordered loop adjacent to the pocket.
• Cofactor and metal-binding sites. AlphaFold2 predicts unligated protein. If your target needs Mg²⁺, Zn²⁺, heme, NADP, or another cofactor for catalytic competence, the prediction will not place it. AlphaFold3 closes part of this gap by co-predicting common cofactors, but the published accuracy is uneven across cofactor classes.

The pre-flight checks worth running

Before committing a virtual screen to an AlphaFold structure:

• Pocket-residue pLDDT scan. Identify the pocket (CASTp, fpocket, or a known ligand-contact set if a homolog structure exists) and average the per-residue pLDDT across just the pocket-lining residues. If that average is below ~80, treat the screen as exploratory only.
• Sidechain sanity against any known holo template.If any homolog of your target has been crystallized with a ligand, superpose and compare pocket side-chain orientations. A handful of wrong rotamers is normal; a wholesale rearrangement is the signal that the apo prediction has the wrong pocket shape.
• Pocket volume check. Compute the pocket volume on the predicted structure and on the closest holo homolog. If the AlphaFold pocket is > 30% smaller, the structure is most likely in an apo-like closed conformation. Either model an induced-fit step or look for an alternative structure.
• Self-dock a known binder. If even one ligand is known for the target or a close homolog, dock it against the AlphaFold structure and check whether the top pose recapitulates the published binding mode. If it doesn’t, your screen on unknown chemistry will not either.

When AlphaFold is the right call anyway

Despite the failure modes, the alternative on a target with no crystal structure is almost always worse: homology modeling from a distant template, a one-off cryo-EM map at low resolution, or nothing at all. AlphaFold gets you into the screening regime earlier than the experimental pipeline ever could. The discipline is to treat the predicted structure as a starting hypothesis rather than as ground truth — a tool that points to a hit list that still needs orthogonal confirmation (biophysics, mutagenesis, or eventually a co-crystal) before it advances.

Try it yourself

Liganx uses experimental crystal structures by default for every target in its catalog and falls back to AlphaFold only for targets that lack one. When an AlphaFold structure is used, the target card surfaces the average pLDDT across pocket residues so you can see at a glance whether the model is in the high-confidence regime. Open Studio and pick any target — for the kinases that have co-crystals with approved inhibitors, dock the known binder first and confirm the pose looks right before you screen unknown chemistry. That single self-docking check catches most of the silent failures the validation papers warn about.

Liganx brings molecular docking online into the browser, so you can run a self-docking check and screen against a predicted or experimental structure without standing up a local pipeline. Molecular docking is only as good as the structure underneath it, and the target card tells you which kind you are working with.

Primary sources

• Jumper J, et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583-589 (2021). doi:10.1038/s41586-021-03819-2
• Abramson J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493-500 (2024). doi:10.1038/s41586-024-07487-w
• Karelina M, Noh JJ, Dror RO. How accurately can one predict drug binding modes using AlphaFold models? eLife 12, RP89386 (2023). doi:10.7554/eLife.89386
• Díaz-Rovira AM, et al. Are deep learning structural models sufficient for virtual screening? J Chem Inf Model 63, 1668-1674 (2023). doi:10.1021/acs.jcim.2c01270