FlowDock

What is FlowDock?

FlowDock predicts the structure of a protein-ligand complex and estimates how tightly the two bind, starting from nothing more than a protein sequence and a ligand SMILES string. It is a generative model: instead of slotting a ligand into a fixed receptor, it builds the bound protein conformation and the docked ligand pose together, then reports a confidence score and an affinity estimate for each sampled complex.

That single trait separates it from most docking tools. Classic docking assumes a known receptor structure and treats the protein as rigid or nearly so. FlowDock generates the holo (bound) protein from an apo prior, so the backbone and side chains can move to accommodate the ligand. The cost of that flexibility is sampling: FlowDock draws several complexes per run and ranks them, rather than returning a single deterministic answer.

The method comes from Alex Morehead and Jianlin Cheng (ISMB 2025). It is most useful when no experimental holo structure exists, when binding may reshape the pocket, or when a fast affinity estimate alongside the pose is more valuable than a single high-resolution docking result.

How to use FlowDock online

ProteinIQ runs FlowDock on GPU infrastructure with no setup. Provide a protein as a sequence, FASTA, or receptor PDB, and a ligand as a SMILES string or SDF file. FlowDock samples several protein-ligand complexes and returns ranked PDB structures, ranked SDF ligand poses, per-pose plDDT confidence, and an affinity estimate for each, all downloadable. An optional template PDB can guide the structure prior.

Inputs

Input	Supported formats	Notes
Protein	Plain sequence, FASTA, or PDB	Up to 4096 residues. Separate sequence chains with |. A PDB is used directly as the receptor.
Ligand	SMILES or SDF	Up to 300 heavy atoms. Use | between SMILES strings for multi-ligand inputs.
Template	PDB (optional)	Used by the ESMFold structure prior when it matches the input sequence.

Two ProteinIQ tools pair naturally with the inputs here. If only a ligand name or PubChem CID is on hand, fetch the SMILES directly in the ligand card. If the protein is known only by sequence and a starting fold helps, predict one first with ESMFold and pass it as the template.

Settings

FlowDock's defaults reproduce the sampling configuration from the original sample.py, so the out-of-the-box run matches the published method. The settings below mostly trade speed against the number and diversity of sampled complexes.

Setting	Description
Number of samples	Complexes to generate (1-20, default 5). More samples explore more poses but take longer.
Sampling steps	Integration steps along the flow (10-80, default 40). More steps refine geometry at the cost of runtime.
Sampler	VDODE (default) or ODE. The variance-diminishing ODE sampler is the published default.
Sampler eta	Noise scale for the sampler (0.5-2.0, default 1.0). Lower values are more deterministic.
Start time	1.0 (default) starts from the full prior; auto lets FlowDock pick the start point.
Random seed	Optional positive integer recorded with the source run. Empty uses FlowDock's default stochastic behavior. This FlowDock GPU release can still produce different coordinates and scores across isolated executions with the same seed.
Use template	Applies the supplied template PDB to the structure prior when the sequence matches (default on).
Discard SDF coordinates	Ignores input ligand coordinates and samples poses from scratch (default on).
Detect covalent bonds	Enables FlowDock's source-native receptor-ligand covalent bond detection (default off).
Rank by confidence	Orders outputs by the confidence head and writes prot_rank*/lig_rank* files (default on). When disabled, FlowDock preserves source sample order in prot_0*/lig_0* files without assigning ranks.
Ranking score	Which plDDT drives ranking: Ligand (default), Protein, or the source's combined Protein + ligand setting. Ligand plDDT is the right choice when pose quality matters most. In this pinned release, the combined option is accepted only when Rank by confidence is off (so the ranking choice is ignored); enabling both makes FlowDock fail while writing its result table.

Outputs

Output	Description
prot_all.pdb	Multi-model PDB with every sampled protein structure.
lig_all.sdf	Multi-conformer SDF with every sampled ligand pose.
prot_rank*.pdb	Ranked protein structures (when confidence ranking is on).
lig_rank*.sdf	Ranked ligand poses (when confidence ranking is on).
prot_0*.pdb, lig_0*.sdf	Protein structures and ligand poses in source sample order when confidence ranking is off.
lig_ref.sdf	Reference ligand SDF, written when ligand coordinates are available.
*_auxiliary_estimation.csv	Native per-pose rank (when enabled), per-ligand plDDT values, and per-ligand affinity estimates in pK units.
run.log	Run settings, checkpoint identity, prediction count, and completion outcome.
flowdock-source.log	Detailed diagnostic output from the FlowDock run.

The Data tab preserves rank or sample index, every per-ligand plDDT and affinity value, and the paired protein and ligand filenames. The Viewer tab loads each ligand with its matching protein conformation. The Files tab retains the aggregate structures, individual predictions, native tables, and logs.

How FlowDock works

FlowDock treats docking as a transport problem. It learns a continuous flow that carries an unbound starting structure toward a bound protein-ligand complex, parameterized by conditional flow matching. At inference it integrates that flow over a fixed number of steps, moving atoms from the prior toward a plausible holo state.

The starting point comes from a structure prior. FlowDock uses an ESMFold-derived apo conformation of the protein, which is why a sequence is enough to start and why a matching template PDB can be supplied to steer the prior. The ligand is initialized from its molecular graph, with input coordinates discarded by default so the pose is sampled rather than copied.

Two auxiliary heads run on top of the generated geometry. A confidence head produces plDDT-style per-residue and per-ligand scores, and an affinity head predicts binding strength. Because generation is stochastic, FlowDock draws several complexes and uses the confidence head to rank them, so the top-ranked file is the model's best guess rather than the only one it considered.

Interpreting results

plDDT here mirrors the confidence metric from structure prediction: higher means the model is more certain about the local geometry it produced. Ligand plDDT speaks to pose reliability, protein plDDT to the backbone and pocket. Treat a high-plDDT top pose as a confident prediction and a run where even the best pose scores low as a signal that the pocket, the ligand, or the prior is hard for the model.

Affinity is reported in the source model's pK units (-log10 molar affinity), so higher values indicate stronger predicted binding. For multiple ligands, ProteinIQ preserves every per-ligand pK value from FlowDock's native CSV and separately shows the source's mean complex affinity. These are model estimates, not measurements: use them to rank candidates docked the same way, not as a substitute for an assay or a physics-based free energy.

Always inspect more than the top file. The ranked SDF and PDB outputs let alternative poses be compared, and a cluster of high-confidence poses that agree on the binding mode is more trustworthy than a single high score. For a physics-based cross-check of the top complex, an MM-PBSA free energy estimate on the ranked structure gives an independent read on binding.

When to use FlowDock vs alternatives

FlowDock fits the case where no holo receptor exists and the pocket may rearrange on binding. It generates the bound protein and ligand together and hands back an affinity estimate in the same run, which classical docking does not.

If a high-quality receptor structure is already in hand and the goal is fast, well-calibrated pose ranking, AutoDock Vina is the leaner choice, and GNINA adds CNN rescoring on the same rigid-receptor setup. For blind docking of a ligand into a known structure without a search box, DiffDock is a closer comparison, though it does not move the protein or score affinity. When the target is a broader complex, including multiple chains or cofactors, the all-atom co-folding models Chai-1, Protenix, and Boltz-2 predict the assembly and, in Boltz-2's case, affinity as well.