AI molecular docking

What is AI molecular docking?

AI molecular docking is a computational method for predicting how a small molecule binds to a protein using models trained on known protein–ligand structures. Given a protein and a ligand, the model generates candidate three-dimensional binding poses and ranks them with a learned score.

Classical docking programs such as AutoDock Vina search many ligand placements and score each one with a hand-built physics or empirical function. AI docking replaces part or all of that loop with a neural network. A diffusion model, for example, starts from a random ligand placement and learns how to move it step by step toward a plausible pose.

Most AI docking models are built to find the right pose, not to measure binding strength. Their confidence values estimate how likely a pose is to be correct. Some models also report a predicted affinity, but that number is a learned estimate, not a measurement, and each model uses its own scale.

A top-ranked pose is still a hypothesis. Learned models can put a ligand in the right pocket with distorted bonds, steric clashes or inverted stereocenters, errors that physics-based programs rarely make. Checking poses with PoseBusters before interpreting them is part of the method, not an optional extra.

When to use AI molecular docking

AI docking is most useful when the usual inputs for classical docking are missing or uncertain:

• There is no experimental structure of your ligand bound to the target.
• You do not know where the ligand binds, a natural fit for blind docking.
• The pocket may change shape on binding, or you only have an apo or predicted structure.
• You want an independent check on a result from a physics-based docking program.

Strengths and limitations

The practical gain is less setup and more ways to be wrong independently. Global models find a pocket without a hand-drawn docking box, and flexible-receptor models are designed to adjust apo or AlphaFold-style structures toward a bound conformation. Because diffusion, flow matching and CNN scoring fail in different ways, running several of them on the same inputs shows quickly which poses depend on one model's assumptions.

The main weakness is physical validity. High-confidence poses can still contain clashes, strained torsions or broken stereochemistry. Accuracy also drops for protein families, pockets and chemistries that are rare in the training data, such as metal sites, covalent ligands and large macrocycles. Published benchmarks can overstate performance when test complexes resemble training complexes, and a confident pose says little about how tightly the ligand actually binds.

How AI molecular docking works

AI docking models differ in three choices: how they generate a pose, whether they need a defined binding site, and whether the protein is allowed to move. These choices decide which model fits your question more than headline benchmark numbers do.

Model	How it generates poses	Binding site	Protein
DiffDock-L	Diffusion over ligand position, orientation and torsions	Whole protein	Rigid
SurfDock	Diffusion guided by protein surface shape and electrostatics	Defined pocket	Rigid
SigmaDock	Fragment-based diffusion	Defined pocket	Rigid
DynamicBind	Diffusion that also moves the protein	Whole protein	Flexible
FlowDock	Flow matching from unbound to bound complex	Whole protein	Flexible
GNINA	Classical search with CNN scoring	Search box	Rigid by default

Diffusion models

A diffusion model starts with the ligand placed randomly and removes that noise over many small steps, adjusting its position, orientation and rotatable bonds until it settles into a pose. DiffDock-L searches the entire protein this way. SurfDock adds a description of the pocket surface, and SigmaDock diffuses rigid fragments of the ligand, then ranks the assembled poses with Vinardo.

Flexible-receptor models

DynamicBind and FlowDock generate the ligand pose and a new protein conformation together. This lets them open pockets or shift loops that a rigid receptor would block, which matters when the input is an apo or predicted structure. The cost is a second thing to validate: the predicted protein motion must be plausible, not just the ligand pose.

Hybrid models

Hybrid methods keep a conventional pose search and use a neural network to score or rescore the result. GNINA, for example, samples poses much like Vina and ranks them with a convolutional network trained on protein–ligand complexes. Because the search is classical, hybrid poses tend to have sound geometry, but they still depend on a correctly placed search box.

Most generative models are stochastic, so two runs with the same inputs can return different poses. Sample several poses per model and treat the top rank as a starting point for review.

Applications of AI molecular docking

The most common use is predicting a pose for a ligand that has never been crystallized with its target, either to study the interactions it might make or to give a medicinal chemist a starting model. Global models extend this to pocket discovery: run on a protein with no annotated site, they suggest where a ligand could bind, including cryptic pockets that only open in the bound state.

AI docking also works well with predicted structures. Rigid docking into an AlphaFold or homology model often fails because side chains sit in their apo positions, and flexible-receptor models can partly correct for that. Researchers also run learned and physics-based methods side by side on the same inputs, the idea behind consensus docking, and redock ligands from known complexes to check how often each model recovers the experimental pose on their target class before trusting it on new compounds.

How to do AI molecular docking online

1. Prepare the protein. Start from an experimental or predicted structure, repair missing atoms with PDBFixer, and decide which chains, cofactors, metals and waters belong in the binding site.
2. Prepare the ligand. Provide an SDF with correct bonds, stereochemistry, protonation and tautomer. SurfDock uses the ligand's coordinates to define the pocket, so place it in the site first or leave SurfDock out.
3. Open the AI docking workflow. The ProteinIQ workflow sends the same protein and ligand to DiffDock-L, DynamicBind, FlowDock and SurfDock, so differences between results come from the models rather than the inputs.
4. Choose the sampling depth. Each model generates several poses per run. More samples improve coverage at the cost of runtime; record the settings so the run can be repeated.
5. Check every pose with PoseBusters. Remove poses that fail bond geometry, stereochemistry, planarity or clash checks, whatever their confidence.
6. Compare the surviving poses. Look for poses from different models that occupy the same pocket with similar contacts. ProLIF interaction fingerprints make that comparison concrete instead of visual.

How to interpret AI molecular docking results

Check the structure before the score. A confident pose with impossible geometry should be rejected, while a physically sound pose with lower confidence can still be a useful hypothesis.

Each model reports a different kind of number, and they cannot be compared with each other directly:

Model	Reported score	What it means
DiffDock-L	Confidence	Likelihood that a pose lies within about 2 Å of the true binding mode. Ranks poses within one run.
DynamicBind	lDDT and pKd	lDDT is confidence in the predicted complex, including protein motion. pKd is a learned affinity estimate.
FlowDock	plDDT and affinity	Ligand plDDT ranks sampled complexes. The affinity is a model output, not a measured constant.
SurfDock	SurfScore	Relative ranking signal for pose quality, not an energy.
SigmaDock, GNINA	Vinardo or Vina score, CNN score	Empirical energy in kcal/mol, plus a 0 to 1 CNN pose confidence.

Do not average these values or rank poses across models by raw score. Compare models by whether their valid poses agree on the pocket, orientation and key interactions.

AI changes how poses are generated, not what counts as evidence. Stronger conclusions combine a valid pose with known pocket biology, structure–activity data, mutagenesis, simulation or experiment. Keep the inputs, settings, sampled poses and selection rationale together so the result can be reproduced.

How the AI molecular docking workflow works

Compare DiffDock-L, DynamicBind, FlowDock, and SurfDock using the same protein and SDF ligand inputs.

1. Check input compatibility. Upload a complete protein PDB and a standardized ligand SDF with explicit stereochemistry, bonding, and charge. Check each selected model’s accepted formats.
2. Review model domains. Consider whether the target class, ligand chemistry, metals, cofactors, covalent mechanism, or unusual residues fall outside typical training examples.
3. Run learned pose generation. Generate poses with DiffDock-L, DynamicBind, FlowDock, and SurfDock from the same inputs and preserve each model’s native structures and confidence.
4. Compare confidence and geometry. Use PoseBusters and direct structural inspection to identify clashes, strained conformations, implausible bond geometry, and disconnected interactions.
5. Validate selected poses. Review agreement across learned models and, where possible, a classical baseline, known ligands, mutational evidence, or experimental structures before prioritizing a pose.

Inputs and outputs

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

Inputs

Structural inputs

PDBSDFSMILES

A protein PDB and standardized ligand SDF compatible with the selected learned models.

Method context

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

Outputs

Docked structures

PDBPDBQTSDF

Predicted complexes, model-specific confidence, ranked ligand poses, and downloadable structures.

Review evidence

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