Blind docking

What is blind docking?

Blind docking is a molecular docking method for searching a protein for ligand binding poses without specifying a binding pocket in advance. The search covers the receptor surface or a set of candidate cavities, so it tests both where the ligand might bind and how it might sit there.

In local docking, the search region comes from a known ligand or other site evidence. Blind docking removes that constraint. A classical engine can search a box around the whole receptor, while a site-agnostic learned model can generate poses without a box. Both depend on the supplied protein conformation.

The output is a collection of candidate sites and poses, not a map of occupied sites in a cell. A large search space gives more opportunities to find a favorable score at an irrelevant surface patch. It also spreads a finite sampling budget over more locations, so a plausible site may be missed.

When to use blind docking

Use a broad search when the binding location itself is part of the question:

• A target has no reliable annotated ligand site.
• A compound may bind outside a known active site.
• Local docking repeatedly gives implausible poses at the assumed pocket.
• You want candidate sites to test against mutagenesis or structural data.

Strengths and limitations

Blind docking does not force the ligand into a preselected site. Agreement between independent searches can identify pockets worth closer inspection, and fpocket can provide a separate description of cavity geometry.

Whole-protein searches are harder to sample than local ones. Surface grooves, crystal contacts, missing cofactors and an incorrect oligomeric state can all produce attractive but irrelevant poses. A cavity detector can miss transient sites, while a learned model may favor pocket patterns represented in its training data.

How blind docking works

The main approaches differ in how they cover the receptor. None determines biological occupancy from a pose alone.

Approach	Search region	Main limitation
Whole-protein classical search	One box enclosing the receptor	Sampling becomes thin as the box grows
Cavity-guided local searches	Separate boxes around detected pockets	Misses sites absent from the cavity list
Site-agnostic learned docking	No user-defined box	Depends on model training and pose geometry

Whole-protein and learned search

AutoDock Vina can search the whole protein using its whole-protein setting. Its Vina score is an approximate energy in kcal/mol, not a probability that the selected site is correct. DiffDock-L samples ligand poses across the receptor and ranks them by learned pose confidence. The two methods offer different search assumptions, which makes disagreement informative.

Cavity-guided docking first identifies pockets, then runs a local search for each. This can spend more effort on each site, but the initial pocket selection becomes a filter on what can be found. Keep the cavity results separate from docking evidence.

Applications of blind docking

Blind docking is useful for a target with little site information or for testing whether a ligand might occupy an alternative pocket. For an enzyme, a predicted site away from the catalytic cleft can guide residue selection for an allostery experiment. The docked pose gives a geometric proposal for that test, not evidence of an allosteric effect.

It can also challenge an assumed binding site. If a compound repeatedly occupies another cavity under different protein preparations, compare both sites against conservation, accessibility, known ligands and biochemical data. Searching one ligand across many unrelated proteins is a different task, inverse virtual screening.

How to do blind docking online

1. Prepare the full receptor. Use the relevant biological assembly, inspect missing loops and retain cofactors or metals that shape the surface. PDBFixer can repair missing atoms, but review what its repair changed.
2. Set the ligand chemistry. Supply a defined stereoisomer and plausible protonation and tautomer states. Run uncertain states separately.
3. Inspect candidate cavities. Run pocket detection as context, then note any experimentally supported site. Do not limit the blind search to the top detected pocket.
4. Run the blind docking workflow. Give the same receptor and ligand to whole-protein AutoDock Vina and DiffDock-L. The workflow keeps their poses and native scores separate.
5. Group poses by location. Compare which regions recur across methods and ligand states, then inspect contacts and accessibility in the biological assembly.
6. Reject invalid poses. Use PoseBusters with the receptor to check ligand geometry and clashes before prioritizing a site.

How to interpret blind docking results

Interpret the predicted site before the within-method pose rank. A recurring pocket is more interesting if it is accessible, chemically sensible and supported by independent structural or functional evidence. DiffDock-L confidence concerns pose accuracy; Vina energy ranks its own solutions. Neither is the probability that a site is occupied.

Record the protein state, ligand state, search settings and all candidate sites, including failures. Prioritize sites for competition experiments, mutagenesis or direct structural study. Blind docking predicts locations to test; it does not establish binding, affinity or biological activity.

How the blind docking workflow works

Compare whole-protein AutoDock Vina with site-agnostic DiffDock-L from the same protein and ligand.

1. Prepare the whole protein. Use the biologically relevant protein assembly where possible. Repair missing atoms, inspect unresolved loops, retain important cofactors, and remove irrelevant crystallization components.
2. Check candidate cavities. Provide the intended stereoisomer, tautomer, protonation state, and charge. Run distinct chemical states separately because blind searching does not correct ligand chemistry.
3. Standardize the ligand. Run fpocket to obtain independent pocket locations and descriptors. Use these results as context or to define separate local searches, not as confirmed binding sites.
4. Search globally. Compare AutoDock Vina in whole-protein search mode with DiffDock-L from the same receptor and ligand. Increase search effort deliberately when using a very large classical search box.
5. Cluster sites and poses. Cluster poses by spatial location, inspect geometry and recurring contacts, compare predictions across methods, and prioritize sites supported by conservation, mutagenesis, or experimental pocket evidence.

Inputs and outputs

Check formats before running, then inspect and download the result from every workflow step.

Inputs

Structural inputs

PDBSDFSMILES

A complete prepared protein PDB and a chemically valid ligand.

Method context

Binding-site evidence, restraints, receptor-state provenance, known ligands, or reference complexes when available.

Outputs

Docked structures

PDBPDBQTSDF

Candidate binding locations, ranked poses, confidence or docking scores, and pocket context.

Review evidence

Method-native rankings, confidence, logs, interaction context, failures, and files for reproducible follow-up.