Protein Design · Ovchinnikov Lab / Baker Lab
BindCraft
An automated pipeline designing target-specific protein binders via AlphaFold2 co-folding hallucination and MPNN.
What BindCraft is
BindCraft is an automated computational pipeline for de novo protein binder design, developed by researchers associated with the Ovchinnikov Lab and the Baker Lab. The framework tackles the problem of generating synthetic binders directly against a chosen target protein by co-folding the target and candidate sequences through AlphaFold2 backpropagation.
Traditional binder discovery often demands testing hundreds of experimental candidates. BindCraft combines backpropagation-guided sequence hallucination with AlphaFold2, sequence design through MPNN, and structural refinement and scoring via PyRosetta. This workflow achieves reported experimental hit rates that require approximately ten ordered screens to identify a functional binder.
The software provides an end-to-end process from input coordinate files to ranked, filtered candidate complexes. It evaluates designed interfaces using structural metrics such as interface predicted alignment error (ipAE), interface predicted template modelling score (ipTM), and interface shape complementarity, allowing researchers to select candidates directly for experimental validation.
How it works
BindCraft generates binders by co-folding the target structure and candidate sequences using AlphaFold2 backpropagation. The trajectory updates binder representations across iterative stages—supporting 2-stage, 3-stage, standard 4-stage, greedy, and Markov chain Monte Carlo (MCMC) design algorithms. In the default 4-stage protocol, design moves from continuous sequence logits, to softmax distributions, to one-hot discrete amino acid representations, followed by position-specific scoring matrix (PSSM) semi-greedy iterations. The loss function guides structural formation using weighted metrics including predicted local distance difference test (pLDDT), intra- and inter-chain predicted alignment errors (pAE), contact distance constraints, interface pTM (ipTM), helicity propensity, and radius of gyration.
Following trajectory generation, sequence optimisation is handled by ProteinMPNN (using either original or soluble weights), which redesigns candidate binder sequences for the generated backbones while optionally fixing the interface residues. The resulting binder–target complexes and isolated binder monomers are repredicted and cross-validated across AlphaFold2 models (such as AF2-ptm and AF2-multimer). PyRosetta relaxes the structural complexes and computes biophysical metrics—including buried unsatisfied hydrogen bonds, interface delta solvent accessible surface area (dSASA), interface binding energy (dG), and shape complementarity—to filter out unviable candidates.
Key facts
- Domain
- protein-design
- Tasks
- binder_design
- Modalities
- protein_sequence, protein_structure
- Core algorithms
- AlphaFold2 hallucination, ProteinMPNN, PyRosetta
- Software version
- 1.1
- GPU memory recommendation
- 32 GB to 40 GB
- Weight requirements
- AlphaFold2 weights (~5.3 GB)
What you can use it for
- Target-directed de novo binder design
- Generate novel protein binders against specific functional epitope hotspots on a target protein PDB structure.
- Unbiased interface binder generation
- Design binders across entire protein chains or allow AlphaFold2 to automatically identify suitable binding patches when no hotspot residues are provided.
- Scaffold and motif grafting preparation
- Minimise the distance between the N- and C-termini of designed binders using specialised terminal distance loss functions to facilitate grafting workflows.
- Multiparametric in silico candidate filtering
- Filter and rank hundreds of candidate binders according to Rosetta energy scores, interface clashes, hydrogen bonding satisfaction, and AF2 confidence scores prior to ordering.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Not runnable yet
This model isn't available to run right now.Run this model
BindCraft
Strengths and limits
Good at
- High reported experimental efficiency, yielding functional binders in approximately 10 ordered screens rather than hundreds.
- Flexible target site options that permit precise hotspot residue selection, multi-residue patches, or fully unconstrained site selection.
- End-to-end automation incorporating AlphaFold2 trajectory hallucination, MPNN sequence redesign, PyRosetta structural relaxation, and automated metric-based filtering.
- Extensive control over structural propensity and geometric loss terms, including helix propensity, radius of gyration, and contact distance cutoffs.
Watch out for
- High GPU memory demand, requiring at least 32 GB to 40 GB of VRAM for larger target and binder complexes.
- Requires substantial compute trajectories, often needing a few hundred to a few thousand runs to yield designs that pass strict filters for difficult targets.
- Reduced performance on hydrophilic interfaces, as AlphaFold2 is notably less effective at designing hydrophilic contacts than hydrophobic interfaces.
- Susceptible to deformed or 'squashed' structural trajectories during AF2 multimer hallucination due to sensitivity to sequence inputs, requiring trajectory rejection.
- ipTM ranking serves as a binary indicator of binding capability rather than a reliable predictor of binding affinity.
Inputs and outputs
You provide
Target structure (PDB)
Hotspot residuesoptional
Optional epitope constraint on the target chain.
Binder lengthoptional
You get back
All-atom coordinates in PDB format, viewable and downloadable.
Ranked binders with interface pAE, ipTM and shape complementarity.
Licence
Commercial use permitted · MIT
The MIT licence allows use in commercial research and products.
Questions
Can I use BindCraft commercially?
Commercial use permitted. The MIT licence allows use in commercial research and products.
How do I run BindCraft without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 2× A100, typically in 15min, and a run costs about 6,300 credits. Visitors without an account get 1,000 credits a month.
What inputs are required to run BindCraft?
BindCraft requires a target protein structure in PDB format, the target chains to interact with, and a desired binder length or range of lengths. Specific hotspot residues can also be provided to constrain the design site, or set to null to allow AlphaFold2 to discover a binding location.
How many designs should be generated before selecting candidates for wet-lab testing?
It is recommended to generate at least 100 final designs that pass all computational filters, and subsequently order the top 5 to 20 ranked candidates for experimental characterisation.
Does a higher ipTM score imply higher binding affinity?
No. While interface pTM (ipTM) has been shown to be a good binary predictor of whether a design will bind, it is not a reliable predictor of binding affinity. To find high-affinity binders, testing a larger number of candidates is recommended.
How can GPU memory usage and runtimes be reduced?
Trimming the input target PDB structure to the smallest functional size possible significantly lowers GPU memory requirements and accelerates binder generation. Selecting lighter design algorithms, such as greedy or MCMC modes, also reduces memory usage compared to default multi-stage backpropagation.
How to cite it
Written from BindCraft's own documentation (code). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.