PoseBusters

What is PoseBusters?

PoseBusters checks whether generated or docked ligand poses are chemically sane, geometrically plausible, and compatible with an optional protein binding site. The method was introduced by Buttenschoen, Morris, and Deane after showing that low RMSD alone can hide broken chemistry, impossible bond geometry, ligand self-clashes, and receptor overlaps in modern docking benchmarks.

The main question is not "did the pose score well?" It is "could this pose exist as a molecule in this pocket?" That makes PoseBusters useful after docking with GNINA, DiffDock, DynamicBind, SigmaDock, or other pose generation tools, especially when comparing many ranked conformations.

How to use PoseBusters online

Run PoseBusters online by uploading a predicted ligand pose, optionally adding a reference ligand and protein receptor, then choosing the validation mode. ProteinIQ returns a downloadable table of pass/fail checks for ligand chemistry, 3D geometry, receptor clashes, and reference-pose agreement when reference inputs are supplied.

Inputs

PoseBusters can run with just a ligand pose, but the most informative checks require the files that define the intended docking context.

Each predicted molecule file can contain several poses, and a job can include up to 20 predicted molecule files. PoseBusters reports one row per molecule or conformer unless Top poses to validate limits the run.

Predicted molecules are limited to 1,024 atoms and must have bond connectivity when they contain multiple atoms. Each uploaded file is limited to 50 MiB. Native loading failures appear in the report when PoseBusters can produce a diagnostic row.

Pairing different references and receptors

Choose CSV pairing table to pair predicted files with different reference ligands or receptors. Upload files into the three table file groups, then upload or paste a CSV with mol_pred and, when needed, mol_true and mol_cond columns. Values are 1-based upload positions within each group. They bind the pairing to the uploaded file order rather than editable display names.

mol_pred,mol_true,mol_cond
1,1,1
2,1,2

This pairs the first predicted file with the first reference and first receptor, then the second predicted file with that reference and the second receptor. Repeated rows repeat the native evaluation and remain in the report. The same predicted-file limit of 20 applies.

Settings

Custom configurations

Choose Custom configuration (JSON) in Validation mode, then paste a complete JSON configuration. The dictionary replaces the built-in configuration; it is not merged with a preset. JSON represents the same dictionary used by PoseBusters' YAML configurations. For example, this configuration runs only the native loading check:

{
  "modules": [
    {
      "name": "Loading",
      "function": "loading",
      "chosen_binary_test_output": ["mol_pred_loaded"]
    }
  ]
}

Use the PoseBusters 0.6.5 configurations as references for module parameters, selected outputs and output names. A configuration's top_n takes precedence over the Top poses setting, including null to evaluate all poses. Native parallel SDF processing applies that limit separately to each chunk.

Pricing retains a minimum of 5 credits, with 5 credits per evaluated pose in standard or Custom mode and 2 in fast modes. Table pricing counts repeated evaluations and custom configuration limits, including native parallel chunks.

Validation modes

Auto-select from inputs is usually the safest choice. Manual modes are useful when the files are available but the intended evaluation task is narrower than the full input set.

Fast modes run the corresponding check set with speed-oriented settings. They are useful for triage on large SDF files, followed by the full mode on the best candidates.

Results

The results table starts with file and molecule identifiers, then reports Boolean check columns. True means the pose passed that check. False marks a specific failure worth inspecting before using the pose for ranking, rescoring, molecular dynamics, or free-energy calculations.

Full report adds diagnostic measurements that explain failures. For example, number_clashes helps separate one local contact from a badly collapsed conformer, while energy_ratio highlights conformations that pass simple bond checks but remain energetically strained.

CSV reports and the ordered native_table in run metadata retain duplicate columns. Named JSON rows contain one value per column name; row identifiers take precedence if a custom check reuses file, molecule, or position. Run metadata also records the resolved native configuration and software versions.

Interpreting PoseBusters results

A good docking result should pass both physical plausibility checks and task-specific accuracy checks. In redocking, a pose within 2 Angstrom RMSD but with failed stereochemistry or volume overlap should not be counted as a clean success. The pose may be close to the crystal ligand by coordinate RMSD while still representing the wrong stereoisomer or colliding with protein atoms.

Failures have different practical meanings:

• File loading failures: Check file format first. SDF with explicit 3D coordinates is the least ambiguous ligand format for multi-pose validation.
• Sanitization or InChI failures: Repair the ligand before interpreting docking quality. Ligand Fixer can help standardize small-molecule files before rerunning validation.
• Bond length, angle, or internal clash failures: Treat the pose as physically invalid unless a known unusual chemistry explains the geometry. These failures are common when generative models output coordinates without enough chemical constraints.
• Ring flatness and double-bond flatness failures: Inspect conjugated systems manually. Aromatic and double-bond systems should not be arbitrarily twisted to fit a pocket.
• Protein overlap failures: The ligand is colliding with receptor atoms, cofactors, metals, or waters. Receptor preparation and protonation may be involved, but severe volume overlap usually means the pose is unusable.
• Reference identity failures: In redocking benchmarks, RMSD should be interpreted only after confirming formula, bond pattern, and stereochemistry are unchanged.

Passing PoseBusters does not prove that a pose binds strongly. It proves that the pose clears a set of chemistry and geometry sanity checks. Binding affinity, ranking, induced fit, waters, protonation states, and assay relevance remain separate questions.

How PoseBusters works

PoseBusters runs a suite of RDKit-based tests against the predicted ligand and any optional context structures. The checks combine molecule graph validation, 3D geometry validation, receptor-contact analysis, and optional reference comparison.

The paper emphasized a practical benchmarking problem: many AI docking methods were evaluated mainly by ligand RMSD, but some low-RMSD poses contained nonphysical chemistry. PoseBusters adds a second axis to pose evaluation. A method should recover native-like binding modes and generate structures that pass basic chemistry and steric criteria.

The check set changes with the validation mode:

• Molecule checks validate that the ligand can be parsed, sanitized, converted through InChI, kept connected, and represented without radicals or impossible geometry.
• Reference checks compare the predicted and true ligand graphs, formulas, bond patterns, double-bond stereochemistry, tetrahedral chirality, and reference-pose RMSD.
• Receptor checks measure whether the ligand is reasonably positioned with respect to the protein and whether volumes overlap with protein atoms, cofactors, inorganic cofactors, or waters.

PoseBusters is stricter than a visual inspection because it catches subtle graph and geometry problems that can be hard to see in a 3D viewer. It is also narrower than a physics simulation: it does not relax the pose, compute binding free energy, model receptor flexibility, or decide whether the ligand is a true binder.

When to use PoseBusters vs alternatives

For ligand-pose workflows, a common sequence is docking first, PoseBusters second, and deeper analysis third. For example, candidate poses from GNINA or DiffDock can be filtered for chemistry and receptor clashes before expensive molecular dynamics or free-energy calculations with GROMACS or OpenFE.

Practical caveats

PoseBusters results depend on the structure files supplied. Missing hydrogens, unusual protonation, metal coordination, alternate locations, covalent ligands, or incomplete receptor preparation can create failures that need chemical review rather than automatic rejection.

For benchmark reporting, keep the mode consistent across methods. Comparing one docking method with Redock checks and another with Molecule only checks mixes different definitions of success. Report both the physical pass rate and the pose-recovery metric when a reference ligand is available.