ProteoCraft
proteocraftDesign de novo binders, TCR binders and CDR or VHH redesigns with ESMFold2.
ProteoCraft designs protein binders with frozen ESMFold2 models. It combines gradient-guided sequence proposals with a two-stage Monte Carlo tree search, and it checks candidate sequences with full complex predictions. The webserver runs version 1.0.0 of the open-source ProteoCraft ESMFold2 distribution (Wu et al., 2026) and returns the selected binder sequence, its predicted complex structure, the metrics of that prediction and the search records.
Four tasks share the same search:
De novo binder designs an entire binder from the target sequence and a binder length.TCR binder designs a separate binder against fixed TCR alpha and beta chains.Local/CDR redesign mutates the positions marked in a starting binder and keeps every other residue.VHH all-CDR redesign mutates all three CDRs of a starting VHH and keeps the framework.The target and every non-mutable binder position stay fixed in sequence. Coordinates are predicted anew in every run, so a fixed sequence does not mean fixed coordinates. The models stay frozen; only the binder sequence changes during the search.
Use ESMFold2 to fold a sequence you already have, ESMFold2 Binder Design for the Biohub binder recipe, or BindCraft for a structure-based design workflow.
ProteoCraft is billed by run time at 45 credits per minute. Before submission the quote shows an estimate of 2,700 credits, which is 60 minutes of run time. The job settles on its actual run time, and credits that were reserved but not used are refunded when the job ends. A single job runs for at most 12 hours.
Run time depends on the task, the target length and the binder length. The default search runs 110 rounds, 70 in Stage A and 40 in Stage B, and the estimate does not change with the settings on this page. Inputs with long targets or binders can exceed the memory of the GPU the job runs on, in which case the job fails with an out-of-memory error.
| Input | Accepted format | Notes |
|---|---|---|
Target | Raw sequence, FASTA text, .fasta, .fa or .txt | One protein chain per FASTA record, in order. Up to 61 chains. Uppercase one-letter amino acids; X marks an unknown residue. VHH all-CDR redesign takes the 20 standard amino acids only. |
Starting binder | Raw sequence, FASTA text, .fasta, .fa or .txt | Required for Local/CDR redesign and VHH all-CDR redesign. Optional for De novo binder and TCR binder, where the search starts from this sequence and every position stays mutable. One complete sequence of the 20 standard amino acids. |
Job name | Text | Optional label for finding the run later. |
The inputs are sequences. Structure files are not used, because ProteoCraft predicts the complex from sequence.
| Parameter | Type | Default | Description |
|---|---|---|---|
Task | enum | denovo | denovo, tcr, cdr_redesign or vhh_redesign. |
Binder length | integer | 120 | Length of the binder to design. At least 1. Used by denovo and tcr. Redesign tasks, and any run with a starting binder, take the length of that sequence. |
| Parameter | Type | Default | Description |
|---|---|---|---|
Chain roles | string | optional | One role per target chain in input order, target or scaffold. A target chain is the binding partner and a scaffold chain is fixed context. Empty means every chain is a target. At least one chain must be a target. |
Target hotspots | string | optional | Target residues the binder should contact, as 1-based positions or ranges such as 4,8-10. Positions count over all target chains joined in input order, so position 101 is the first residue of the second chain when the first chain has 100 residues. Hotspots must fall on target chains, not scaffold chains. VHH all-CDR redesign does not accept a position listed twice. |
| Parameter | Type | Default | Description |
|---|---|---|---|
Mutable positions | string | conditional | Local/CDR redesign only. One value per residue of the starting binder: 1 allows substitution and 0 keeps the amino acid. Write it as a string of 0 and 1, or as true and false separated by commas. The mask must match the binder length and contain at least one mutable position. For a 12-residue binder, 000111000000 selects positions 4 to 6. |
CDR1 positions | string | conditional | VHH all-CDR redesign only. 1-based positions of IMGT CDR1 in the starting VHH, such as 26-33. |
CDR2 positions | string | conditional | VHH all-CDR redesign only. 1-based positions of IMGT CDR2, such as 51-58. |
CDR3 positions | string | conditional | VHH all-CDR redesign only. 1-based positions of IMGT CDR3, such as 97-117. |
The CDR positions are sequence indices in the VHH you supply, not IMGT numbers. Number the VHH with an IMGT tool first, then convert each CDR to positions in your own sequence. The three regions must not overlap, and each must contain at least one position.
| Parameter | Type | Default | Description |
|---|---|---|---|
Seed | integer | 43 | Fixes the initial logits, the mutation sampling and the verifier diffusion. From 0 to 4294967295. |
Save every verified candidate structure | boolean | false | Writes the structure of every candidate checked by the full verifier. The final selected structure is always saved. |
The search budget, the number of roots, the soft-update counts and the verifier settings follow the bundled esmfold2_d profile and are not settings. The profile uses 8 initial roots, 70 Stage A rounds with 30 soft updates per proposal, 40 Stage B rounds with 20 soft updates per proposal, and a verifier of 4 loops with 50 sampling steps.
All files written by the run are returned and can be downloaded. The result page shows the selected structure in the viewer.
| Output | Description |
|---|---|
hallucination_result.txt | The selected Best Sequence, the Final Sequence, the resolved settings and the loss history of the search. |
esmfold2_artifacts/checkpoints/final/step_NNNN/step_NNNN_best.cif | The selected binder-target complex. NNNN is the completed search budget, which is 0110 for the default run. |
esmfold2_artifacts/checkpoints/final/step_NNNN/step_NNNN_best_meta.json | Metrics from the same prediction as the selected structure: sequence, total_loss, loss_dict and the ESMFold2 confidence summary. |
esmfold2_artifacts/run_summary.json | Summary of the search, including the best sequence, the best loss, the consumed budget and the candidates that the alternative selection rules would pick (stageb_selection_surfaces). |
esmfold2_artifacts/candidates.jsonl | Every discrete candidate checked by the full verifier, with its metrics. |
esmfold2_artifacts/trajectory.jsonl | The search trajectory. |
esmfold2_artifacts/checkpoints/ | Structures and metrics saved during the search. |
esmfold2_artifacts/handoff_checkpoints/ | Stage A state saved every 10 rounds and at the end of Stage A. |
optimization_report.json | Record of the kernel settings and counters of the run. |
run.log | Log of the run. |
VHH redesign runs add case_spec.json, command.txt, provenance.json and vhh_result.json. The last one records whether the framework residues of the best and final sequences match the starting VHH.
Best Sequence in hallucination_result.txt is the sequence ProteoCraft selects at the end of Stage B, and the selected structure in final/ is the prediction of that sequence. The selection is not simply the lowest loss. ProteoCraft ranks the verified candidates first by whether their chain-pair ipTM clears a quality floor, then by hotspot coverage and distance when hotspots are set, then by its tree reward and chain-pair ipTM, and only then by the lowest loss. Another candidate in the run can therefore have a lower total_loss than the selected one. esmfold2_artifacts/run_summary.json records, under stageb_selection_surfaces, the candidate that each of three rules would pick: the lowest total loss, the best composite reward and the hotspot-guarded quality rule that produced the selection.
The loss is the search objective, so lower is better when you compare candidates within one run. It combines the ESMFold2 confidence and geometry terms that steer the search. It is not a probability and cannot be compared across different targets, binder lengths or settings. The loss_dict in the metadata file splits the loss into its terms.
Final Sequence is written by ProteoCraft as the same sequence as Best Sequence.
The summary_confidence block in step_NNNN_best_meta.json holds the ESMFold2 confidence summary of the selected prediction, such as pLDDT, pTM, ipTM and chain-pair ipTM when ESMFold2 reports them. Read pLDDT as local confidence in the structure, and ipTM as confidence in the relative placement of the chains, which is the one that speaks to the binder-target interface. See pLDDT, ipTM and pTM and structure prediction confidence for how to interpret them.
These values describe a model prediction of the designed sequence. They do not measure binding, specificity or affinity, which need experimental testing. The sequence was also optimized against the same models that score it, so a high score is expected and is weaker evidence than a score from an independent model or protocol.
Design protein, peptide, and nanobody binders against protein or small-molecule targets.
Protein, peptide, nanobody and antibody binder design.
Design minibinders and antibody scFvs against a protein target.
Design VHH nanobody binders using structure-guided modeling.
Design antibody and nanobody binders with an end-to-end RosettaCommons pipeline
Design antibody sequences from structure with AI-powered inverse folding
Humanize antibodies with optimal V-gene matching.
Design antibody CDR loops from antibody-antigen complex structures using inverse folding.
Design antibody CDRs with Proteo-R1 reasoning, diffusion, and optional framework inpainting.
Design de novo protein binders for target surfaces using structure-guided sequence generation.