BindCraft

BindCraft webserver overview

ProteinIQ runs an experimental Rosetta-free BindCraft variant for protein binder design. It retains AlphaFold2-guided backbone design, ProteinMPNN sequence redesign, binder-alone validation, and ranking by interface confidence. Restrained OpenMM minimization replaces Rosetta relaxation; FreeSASA and sc-rs provide separately named geometric measurements.

Rosetta/PyRosetta is not included due to licensing. Projects that need Rosetta relaxation, energies, or filtering should contact sales to discuss licensing and requirements. Accepted designs pass a different filter policy from the published BindCraft pipeline; equivalent experimental success rates have not been established.

FreeBindCraft is an alternative with staged OpenMM relaxation, FASPR side-chain repacking, and optional ipSAE ranking. The two tools record different methods and filter policies, so scores carry over between them only when those records agree.

Pricing

The standard compute quote reserves 5,400 credits for two hours on an A100 80 GB GPU at 45 credits per minute. Actual billable runtime is settled after the run and unused reserved credits are returned. The exact quote appears before submission. Full diagnostic artifacts do not change the standard compute reservation.

Target size, binder length, design count, and iteration settings can change runtime but do not change the standard compute reservation. The requested accepted-design count is a search goal, not a guaranteed yield.

Inputs

Preparing the target

Only the chains listed in Target chains are used, so that field defines the binding context the design sees, and a multi-chain site needs comma-separated IDs such as A,B. Read the residue numbering out of the file itself before writing a selector. Cropped or renumbered structures are the usual reason a hotspot lands on the wrong surface without any error.

The BindCraft authors recommend cutting an unnecessarily large target down to lower memory use and speed up sampling. Keep enough structure around the intended site to judge whether a binder could reach it, because an artificial cut surface is a poor basis for a binding claim. What the file omits matters just as much: a missing domain, partner chain, glycan, or membrane can block an interface that looks clean in isolation.

A hotspot steers the search toward a patch rather than pinning one interaction. Give Hotspot residues (optional) a small accessible patch instead of a single buried residue, leave it empty when the search should choose the site, and confirm in the Viewer afterwards that the accepted designs sit where they were asked to.

First run

The displayed defaults make a reasonable setup check: Design protocol Default, Minimum binder length 40, Maximum binder length 80, Number of designs 1, and Preset overrides off. They are defaults, not a validated optimum for a particular target.

Switching Design protocol to Peptide or Beta-sheet selects a different advanced preset but leaves both length fields exactly as entered, so a peptide run keeps the 40 to 80 range until those fields change. Check both after every protocol change.

The BindCraft repository ships a PD-L1 target example that makes a compact first submission: its PDL1.pdb sample as Target protein, A as Target chains, and A56 as the hotspot. The repository's own settings file asks for 65 to 150 residue binders and 100 final designs, which is a campaign rather than a setup check, so keep the hosted defaults for the first run. The exercise covers the inputs only; no acceptance rate or score is implied.

Settings

Defaults below are the displayed controls. With preset overrides off, the design protocol selects the advanced preset. Advanced options may inherit protocol-specific values; inspect advanced_settings.json for resolved values. Quality controls override matching thresholds even when a custom filter or no_filters preset is selected. Inspect scientific_protocol.json for effective filters.

Core settings

Protocol and model controls

Quality thresholds

Preset overrides

Runtime controls

Template options

ProteinMPNN options

Design bias and loss weights

Iterations and recycles

Contact and confidence weights

Full diagnostic artifacts

The Rosetta-free protocol

The protocol proteiniq-restrained-min-v1 prepares predicted complexes with PDBFixer 1.12, preserving protein residue identities and original target chain breaks. OpenMM 8.2.0 minimizes Amber99SB energy with harmonic restraints on the original heavy atoms. It adds missing atoms and hydrogens at pH 7, performs no molecular dynamics or sidechain repacking, and restores residue numbers and confidence B factors. Failed minimization stops execution rather than returning an unchanged structure as a relaxed result.

The original AF2 and ProteinMPNN stages retain their settings. When fixed-interface redesign is enabled, the relaxed trajectory determines which binder residues remain fixed. This means changing relaxation can change the resulting sequences as well as the final scores.

Filtering

The rosetta_free_v1 presets retain AF2 confidence, binder RMSD, secondary structure, contact counts, amino-acid composition and nonzero buried-area checks. The original confidence defaults remain pLDDT 0.8, pTM 0.55, i_pTM 0.5 and normalized i_pAE at most 0.35. Only models actually predicted participate in per-model checks and averages.

Rosetta energy, packing, and hydrogen-bond thresholds are disabled. Shape complementarity and surface hydrophobicity are reported without default acceptance thresholds because their new methods have not been calibrated against successful designs. Removing filters can increase the accepted count without improving experimental quality.

Custom filters must use the new metric names below. For example:

{
  "Average_ShapeComplementarity_sc_rs": {
    "threshold": 0.4,
    "higher": true
  }
}

This is an example of syntax, not a recommended cutoff. Unknown keys and non-null thresholds for unavailable Rosetta metrics are rejected. Old names such as Average_ShapeComplementarity and Average_Surface_Hydrophobicity are rejected rather than translated. Confidence sliders still set their corresponding confidence thresholds. An enabled filter fails when its required measurement is missing or nonfinite.

Outputs and understanding results

The Viewer shows selected designs and confidence scores. The Data tab and downloaded CSV files preserve the full per-model and average measurements. Ranked accepted structures are the outputs available for workflow reuse. Below-threshold candidates remain diagnostic files when full artifacts are enabled.

The Checks tab shows each evaluated candidate, stage, model, measured value, comparison and threshold. Filter it by candidate or show failures only. Disabled checks do not reject a candidate; an unavailable measurement fails an enabled final filter. Later checks may not run after an earlier failure. Trajectory screens use the native rounded confidence, initial complex-validation checks use the unrounded model value, and final filters use the values saved in the native design statistics.

The Residue confidence tab shows pLDDT on the 0-100 scale for every saved trajectory, complex prediction and binder-only prediction, including rejected predictions. Chain, residue number and insertion code follow the source PDB. These values come from prediction C-alpha B factors before minimization or file cleanup, without running another prediction. Historical jobs without these reports show an unavailable message.

Binder_Energy_Score, dG, dG/dSASA, PackStat, and interface hydrogen-bond and unsatisfied-hydrogen-bond metrics remain empty because Rosetta does not run. OpenMM potential energy is recorded only as a minimization diagnostic, never as a binding-energy estimate. Averages remain missing if a required model measurement is missing.

Choosing candidates

Rank order is one input to a decision. Open each candidate in the Viewer and check that the proposed interface would still be reachable in the complete biological target and that the binder does not clash with structure the design never saw. The Checks tab then gives the individual measurements behind a candidate, and the Residue confidence tab shows where a strong average hides one weak stretch at the interface.

Ranking uses interface pTM, while Average_Binder_RMSD reports the RMSD between the binder-alone prediction and the designed fold, so a candidate can rank well on one and fail the other. Prefer a shortlist of distinct sequences and poses over several near-identical designs with similar scores.

Expression and solubility, direct binding measurement, and off-target controls still have to be planned, and an intended inhibition or activation has to be tested separately from binding. The published BindCraft study reports experimental results for its own Rosetta-inclusive pipeline, which says nothing about the yield of this variant.

Downloads and provenance

Alongside structures and CSV files, each run preserves:

• scientific_protocol.json: method versions, effective filters, active prediction models and disabled model filters.
• postprocessing.jsonl: minimization outcomes, seed, energy change, coordinate displacement, residue mapping and interface-scoring status. It also contains the isolated and bound binder SASA and the fraction of isolated binder SASA buried.
• settings.json, advanced_settings.json, requested_filters.json: the resolved run configuration.
• bindcraft.log, failure_csv.csv, and trajectory/design statistics for diagnosing low yield or failures.
• candidate_checks.csv and residue_confidence.csv: complete downloadable versions of the two report tables, retained in compact results too.
• model_runtime.jsonl: which model roles were created or reused and their setup times.

No accepted designs can be a valid result. Inspect the recorded rejection reasons and confidence measurements before changing search settings or thresholds. High model confidence and favorable geometry require experimental validation.

Diagnosing a disappointing run

Relaxing every filter to force an accepted design changes the count without changing the designs. Find the stage that failed first, change one relevant choice such as the target patch or the binder length, and keep the earlier run and its resolved settings for comparison.

Historical Rosetta-based jobs retain their saved values and labels. Their energy and packing scores are not recomputed or presented as measurements from this variant. Older results must be interpreted using their saved scientific protocol.

BindCraft and FreeBindCraft

FreeBindCraft is a separate tool based on an independently maintained fork. Its staged relaxation, repacking and optional ipSAE workflow differ from this conservative restrained-minimization protocol. Results from the two tools should be compared using their recorded methods and filter policies.

For an existing backbone, use ProteinMPNN for sequence redesign. For existing sequences, use AlphaFold 2 for structure prediction.

Sources

• Pacesa M, et al. "One-shot design of functional protein binders with BindCraft." Nature (2025). doi:10.1038/s41586-025-09429-6
• BindCraft source code, settings, and target examples. martinpacesa, GitHub. github.com/martinpacesa/BindCraft
• Licensing PyRosetta. PyRosetta. pyrosetta.org/home/licensing-pyrosetta