SigmaDock

What is SigmaDock?

SigmaDock is a molecular docking method that predicts how small molecule ligands bind to protein targets. Developed at the University of Oxford and presented at ICLR 2026, it achieves approximately 80% Top-1 success rate (RMSD <2 Å and PB-valid) on the PoseBusters benchmark, making it the first deep learning method to surpass classical physics-based docking on this metric.

Rather than treating ligand flexibility through torsional angles, SigmaDock decomposes ligands into rigid-body fragments at their rotatable bonds and learns SE(3) transformations for each fragment. This design improves both the physical validity of generated poses and the model's ability to generalize to proteins not seen during training.

How to use SigmaDock online

ProteinIQ hosts SigmaDock on GPU infrastructure, so no software installation or local GPU is needed. Submit a protein structure and a ligand, and the model generates multiple ranked binding poses within minutes.

Inputs

Input	Description
Protein	PDB file or RCSB PDB ID. Must contain protein atoms; maximum 1000 residues.
Ligand	SMILES string, SDF or MOL file, or PubChem CID. The docked query must be organic. MOL2 is available only as the paired SDF fallback.

SigmaDock expects a standard protein receptor structure that native parsing tools can read cleanly. Receptors with malformed coordinates, unusual fibril assemblies, or heavily modified/non-standard residues may be rejected before docking rather than producing unreliable poses.

Settings

Docking parameters

Setting	Description
Number of poses	How many candidate binding poses to generate (1-40, default 10). More poses increase diversity and the chance of finding the correct binding mode, at the cost of computation time.
Diffusion steps	Number of SE(3) denoising steps in the diffusion process (10-50, default 30). Higher values refine poses more thoroughly but run slower.
Scoring method	Function used to rank generated poses. Vinardo (default) is recommended for most applications; Vina is available as an alternative.
Noise scale	Controls the diversity-fidelity tradeoff (0-1, default 0). At 0, sampling is deterministic and maximizes pose fidelity. Increasing the value adds stochastic noise at each diffusion step, producing more diverse but potentially less accurate poses.

Binding site

SigmaDock requires a known binding pocket. It is not a blind docking method. By default, the binding site is automatically determined from the co-crystallized ligand in the PDB file. For structures without a bound ligand (e.g., AlphaFold predictions or apo crystal structures), the pocket center can be specified manually.

Setting	Description
Pocket center X/Y/Z	Cartesian coordinates of the binding pocket center in Ångströms. Leave blank to use automatic detection.

Results

Each run produces a set of ranked poses as SDF files, displayed in a 3D interactive viewer alongside the input protein. A spreadsheet summarizes the key metrics for each pose.

Column	Description
Rank	Pose ranking by score (1 = best)
Score	Vinardo or Vina binding affinity in kcal/mol; more negative indicates stronger predicted binding
CNN Score	GNINA neural network pose score (0-1) when CNN scoring is enabled. SigmaDock disables CNN scoring in its Vina/Vinardo protocol; reported placeholders are not binding confidence
CNN Affinity	GNINA predicted affinity in pK units when CNN scoring is enabled. SigmaDock's Vina/Vinardo protocol disables CNN scoring, so this field is not a neural affinity estimate for these runs
PB Valid	PoseBusters average validity score (0-1); summarizes 28 checks for bond geometry, steric clashes, and stereochemistry . higher values indicate more physically realistic poses
RMSD	Symmetry-corrected RMSD in Å from the prepared input ligand. For SMILES or translated ligands, this is not an experimental docking accuracy measure
File	SDF file for the docked pose

Interpreting scores

The primary Score column reflects the chosen scoring function (Vinardo or Vina). More negative values rank ahead of less negative values. These calculated scores do not establish a measured dissociation constant or concentration range.

The PB validity average summarizes the native checks. Inspect the complete per-check results and the pose geometry when a check fails. The downloadable native-results archive retains predictions, resolved settings, scores, validation checks and failure records for each seed.

Advanced sampling and cross-docking

Advanced settings include the starting seed, Euler or Heun solver, noise schedule, pocket cutoff and ligand fragment preparation. Defaults match the released sampling configuration. Heun evaluates the model twice per step and can take longer. Keep the number of poses and diffusion steps within the displayed limits.

Turn scoring or PoseBusters checks off to use the source's disabled postprocessing modes. Disabled checks have no resulting metrics. Choose redock or redock-fast for enabled PoseBusters validation.

For cross-docking, upload a reference ligand SDF alongside the query. Its coordinates define the native pocket and center of mass. Reference SDFs may contain metals; the query being docked must still meet SigmaDock’s organic-ligand requirements. Without a reference ligand, ProteinIQ retains its existing manual-center, bound-ligand and protein-centroid placement order.

An optional MOL2 file is a paired fallback for an SDF query that the native parser cannot sanitize. Both files are staged with the same basename. This is not a standalone MOL2 docking input. Without manual coordinates, native fallback coordinates are retained. With manual coordinates, ProteinIQ uses the native parser to select the query and applies its existing rigid pocket placement. For a selected MOL2, only coordinate fields change. Original uploaded files and separately labeled derived files remain downloadable. An optional reference SDF keeps its native reference role; manual placement of that reference is labeled separately.

How SigmaDock works

Molecular docking requires sampling a very large conformational space efficiently. Classical methods like Vina grid-search this space using empirical force fields; diffusion models like DiffDock operate in torsional angle space. SigmaDock takes a different route.

Fragment decomposition

A ligand is broken at its rotatable bonds into rigid fragments. Each fragment maintains its fixed internal geometry (bond lengths, angles) throughout the docking process. The docking problem is then reframed as predicting an SE(3) transformation (translation + rotation in 3D space) for each fragment rather than a set of torsional angles.

SE(3) diffusion

The model uses Riemannian diffusion on the SE(3) manifold. During training, fragment poses are progressively randomized via Brownian motion on the translation and rotation spaces. The neural network learns to reverse this process . denoising noisy fragment configurations back into plausible bound poses. At inference, docking starts from randomized fragment poses and iteratively applies the learned denoising steps.

Architecture

SigmaDock extends EquiformerV2, a graph transformer with SO(3)-equivariance, with a hierarchical graph topology using virtual nodes and edges to capture multi-scale protein-ligand interactions. The prediction head is SO(3)-equivariant to avoid local coordinate frame ambiguities. Soft geometric constraints via triangulation distance conditioning enforce chemical validity across fragment junctions throughout the diffusion trajectory.

Inductive biases over scale

SigmaDock intentionally uses a relatively small training set (~19,000 examples) but achieves AlphaFold3-comparable accuracy through well-designed inductive biases: equivariance, fragment rigidity, and geometric constraints. This contrasts with purely data-driven approaches that rely on much larger training corpora.

Performance

On the PoseBusters benchmark with a temporal train/test split:

Method	Top-1 (RMSD <2 Å & PB-valid)
SigmaDock	~80%
AlphaFold3	comparable
DiffDock-L	12-31%
AutoDock Vina	lower
GOLD	lower

SigmaDock also maintains strong performance on proteins with low sequence similarity to its training set, a common failure mode for deep learning docking methods.

Limitations

• Not a blind docking method. A binding site must be known. The tool auto-detects it from co-crystallized ligands in the PDB, or coordinates can be supplied manually.
• Protein backbone is held rigid; conformational changes upon ligand binding (induced fit) are not modeled.
• Receptor parsing is strict. Standard protein PDBs work best; malformed exports, fibrils, and structures with modified/non-standard residues may need cleanup before submission.
• Docked queries containing metal ions are not supported. A reference SDF may contain metals because it defines pocket coordinates rather than sampled query topology.
• Proteins over 1000 residues must be trimmed to the binding domain before submission.