BindCraft is an automated pipeline for designing de novo protein binders that bind to user-specified target proteins with nanomolar affinity. The tool achieves 10-100% experimental success rates across diverse targets including cell-surface receptors, allergens, de novo designed proteins, and multi-domain enzymes like CRISPR-Cas9.
Unlike traditional methods that require high-throughput screening or experimental optimization, BindCraft generates functional binders in a single computational run. The pipeline has demonstrated therapeutic potential by reducing IgE binding to allergens, modulating Cas9 gene editing activity, and neutralizing bacterial enterotoxins.
BindCraft combines AlphaFold2 backpropagation for structure hallucination, ProteinMPNN for sequence optimization, and PyRosetta for structural refinement. For downstream validation of designed binders, you can use molecular docking tools like DiffDock or AutoDock Vina.
BindCraft leverages AlphaFold2-Multimer as an energy function by optimizing structure prediction confidence metrics through gradient descent. The algorithm iteratively updates sequence embeddings to maximize the interface predicted TM-score (i_pTM), which measures predicted alignment quality between binder and target interfaces.
The optimization uses a composite loss function that balances multiple design goals: interface prediction confidence, binder compactness (measured by radius of gyration), preferred secondary structure content (helicity), and residue count. When you specify hotspot residues, the loss function includes a contact term to focus binding on those positions, though AlphaFold2 may select alternative sites if they offer better binding geometry.
The design process proceeds through multiple stages of increasing sequence specificity. The 4-Stage algorithm (recommended) performs gradient-based optimization on continuous sequence logits, then on the softmax matrix, followed by one-hot encoding without and with randomly sampled mutations. The 3-Stage and 2-Stage variants reduce computational time by skipping intermediate optimization steps, though they may produce lower quality results.
At each stage, AlphaFold2-Multimer predicts the complex structure and computes confidence metrics. The gradient of the loss function with respect to sequence embeddings guides updates toward sequences that form tighter, more confident binding interfaces.
After backbone hallucination, ProteinMPNN redesigns the binder sequence while keeping the backbone geometry fixed. The tool optimizes surface and core residues separately to improve protein stability and solubility while preserving the binding interface. We recommend keeping ProteinMPNN enabled for better sequence quality and reduced aggregation risk.
Final designs undergo validation with AlphaFold2-Monomer to ensure the binder folds correctly without the target present. Only candidates passing quality thresholds (pLDDT, pTM, i_pTM) proceed to PyRosetta relaxation, which refines side-chain conformations and computes interface binding energy ().
BindCraft accepts target structures in PDB format, either uploaded directly or fetched from the RCSB PDB database. The target should have resolved 3D coordinates for the region you want to target.
You must specify which chain(s) from the target structure to design binders against. For multi-chain targets, use comma-separated chain IDs (e.g., A,B).
Hotspot residues are optional but can improve success rates when you know which surface patch mediates biological function. Format hotspot specifications as chain-residue pairs: A10,A15,A20 targets residues 10, 15, and 20 on chain A.
Leave this field empty to let AlphaFold2 automatically select the binding site. The algorithm may override your hotspot selection if it finds a better binding surface nearby, since the contact loss is balanced against other design objectives.
We recommend targeting secondary structure elements (helices, strands) over flexible loops. Ideal hotspot patches contain hydrophobic residues with few rotameric states: phenylalanine, tyrosine, tryptophan, isoleucine, leucine, or methionine.
Number of designs controls how many final binder candidates to generate. More designs increase the likelihood of finding a successful binder but extend runtime proportionally. For most targets, 5-10 designs provide a good balance.
Minimum and maximum binder length set the allowed size range in residues. The pipeline randomly assigns lengths between these bounds for each design attempt. Shorter binders ( residues) fold faster and are easier to express, while longer binders ( residues) can accommodate larger binding interfaces.
Binder length affects the search space significantly. Very short binders ( residues) may lack sufficient surface area to form stable interfaces, while very long binders ( residues) take longer to design and may suffer from folding or solubility issues.
pLDDT measures per-residue confidence in the predicted structure, normalized to . Values above indicate high backbone confidence, while values below suggest unreliable geometry. The threshold filters out designs where AlphaFold2 is uncertain about local structure.
pTM is an integrated measure of how well AlphaFold2 predicted the overall complex structure, computed as the predicted TM-score between the model and the hypothetical true structure. This metric captures global fold quality.
Thresholds around are typical for binder design, as this indicates reasonable overall structural confidence. Higher thresholds may exclude valid designs where the binder itself folds well but the complex orientation has uncertainty.
i_pTM specifically measures predicted alignment quality between binder and target interfaces. This is the most critical metric for binder design, as it directly reflects binding interface geometry confidence.
Values above represent confident high-quality predictions. Values between and fall in a grey zone where the interface may be viable but requires experimental validation. Below typically indicates a failed design.
We recommend starting with and increasing the threshold if you get too many low-quality candidates. Lowering the threshold may help for challenging targets where few designs pass stricter filters.
After PyRosetta relaxation, BindCraft computes interface binding energy () in Rosetta Energy Units (REU). More negative values indicate stronger predicted binding affinity. This metric complements the AlphaFold2 confidence scores by estimating thermodynamic stability.
The 4-Stage algorithm provides the best quality by thoroughly exploring sequence space through multiple optimization regimes. Use 3-Stage when you need faster results without significant quality loss, or 2-Stage for rapid prototyping when speed is critical.
We recommend keeping multimer mode enabled for better interface prediction. The multimer model was specifically trained on protein complexes and provides more accurate i_pTM scores. Only disable this for very small targets where the monomer model might suffice.
Disabling ProteinMPNN saves computation time but typically produces sequences with lower stability and higher aggregation risk. Keep this enabled unless you have specific reasons to use the raw AlphaFold2-hallucinated sequences.
BindCraft returns designs ranked by quality metrics, with each entry showing i_pTM, pLDDT, interface , and downloadable PDB files. The 3D viewer displays the binder (typically colored distinctly) complexed with the target structure.
Focus on designs with i_pTM , pLDDT , and negative interface values. These metrics together indicate confident structure prediction, high local quality, and favorable binding energetics.
Designs with high i_pTM but less negative may still be worth testing experimentally, as Rosetta scoring doesn't always correlate perfectly with experimental binding affinity. Conversely, designs with excellent but mediocre i_pTM should be viewed with caution.
Target selection: BindCraft works best on targets with well-defined binding surfaces and available high-resolution structures. Flexible or disordered regions make poor targets because AlphaFold2 struggles to hallucinate binders to undefined geometry.
Start with defaults: The recommended settings (4-stage algorithm, multimer model, ProteinMPNN enabled) provide a good starting point for most targets. Adjust quality thresholds based on your success rate—if all designs pass easily, increase thresholds; if none pass, decrease them.
Iterate on failures: If BindCraft fails to generate passing designs, try specifying different hotspot residues or adjusting binder length ranges. Some targets may require multiple runs with varied parameters to find successful binding modes.
BindCraft requires several hours of computation per target, making it unsuitable for high-throughput screening of many targets. The credit cost reflects this substantial computational investment.
The tool designs monomer binders—it does not currently support designing dimeric or oligomeric binders. For targets requiring multivalent binding, you would need to design monomers separately and engineer dimerization domains.
Success rates vary by target difficulty. Globular proteins with defined binding pockets typically achieve higher success rates than intrinsically disordered proteins or membrane proteins with flexible regions.
After designing binders with BindCraft, validate them using molecular docking: DiffDock for AI-powered pose prediction, AutoDock Vina for physics-based scoring, or GNINA for deep learning-enhanced docking.
For alternative protein design approaches, explore RFdiffusion and RFdiffusion2 for diffusion-based backbone generation, or BoltzGen for generative structure design. Use LigandMPNN to design sequences that bind small molecules instead of proteins.
Visualize designed structures with PDB Viewer, or predict binder structures independently using AlphaFold 2, Chai-1, or Boltz-2 to validate folding without the target.
