BindCraft : tips and guidance

These tips are taken directly from the BindCraft wiki page and BindCraft paper supplementary info, but reformatted to be more concise and readable, with a few extra tips:

Installation

  • System Requirements: Runs exclusively on Linux systems with conda or mamba.

  • CUDA & JAX: Specify your CUDA version during installation to avoid version conflicts, especially with JAX.

  • Dependencies: Uses ColabDesign, ProteinMPNN, and PyRosetta. Commercial users need a PyRosetta license.

  • Environment Path: If the conda environment fails to activate, its path can be manually specified in the run script.

  • Execution: Designed for HPC clusters with SLURM but can be run locally via bash or python scripts.

  • Apptainer containers: are also available (eg ) - these can simplify installation when the provided install_bindcraft.sh script fails

Hardware Considerations

  • GPU: A CUDA-compatible NVIDIA GPU is mandatory. Recommended models include L40, V100, A100, or H100.

    • An H100 GPU is roughly 4 times faster than an A100.
    • eg, on H100: 900 residue trajectory == 2-3 hours, 250 residue trajectory == 5 minutes

  • Google Colab: Can be run on Colab, but it is approximately 10 times slower than a local installation with the same hardware.

  • GPU Memory: This is the primary bottleneck. A 32 Gb card handles a complex of ~550 residues, while an 80 Gb card manages ~950 residues.

  • Parallelization: The core process cannot be split across multiple GPUs, but you can run multiple separate jobs outputting to the same folder to speed up sampling.

  • Storage: A few terabytes are recommended. The model weights alone are 5.3 Gb. You can save space by disabling plot and animation outputs.

  • CPU & RAM: A single CPU core is usually sufficient. At least 40 Gb of RAM is recommended to avoid out-of-memory errors from model compilation or PyRosetta features.

Getting Started

  • Input: Requires a target protein structure (PDB), a binder size range, and the desired number of designs.

  • Binder Size:

    • Globular binders: Optimal between 60-180 amino acids (max reliable size is 250 AAs).

    • Peptide binders: 8-25 amino acids, requiring special peptide settings and filters. Peptide design success rates are typically 5-10 times lower.

  • Hotspots: You can specify “hotspot” residues for the binder to target. If none are provided, the pipeline finds an optimal site on its own.

  • Monitoring & Early Stopping:

    • Check the Accepted folder during the run to manually remove unsuitable designs.

    • The script may terminate prematurely if design success rates are too low, saving computational resources.

    • To stop early and rank current designs, kill the jobs, set the design count lower than what’s in the Accepted folder, and rerun.

  • Run Time: Can vary from a few hours to several weeks on a single GPU, depending on target difficulty and complex size.

    • “Easy” targets might only require ~100 trajectories, difficult ones might require 1000 - 10,000.
    • It’s possible for difficult targets to yield no successful designs.

Under the Hood

  • Core Method: Uses backpropagation through the AlphaFold2 multimer network in MSA-free (single sequence) mode.

  • Sequence Generation: Optimizes the sequence in a “relaxed” (soft) space before converting it to a standard one-hot encoded sequence.

  • Self-Consistency Checks:

    1. Model Swapping: Randomly swaps between the five trained AlphaFold2 multimer models during design to ensure robustness and avoid overfitting.

    2. Sequence Optimization: Uses ProteinMPNN to refine the binder’s core and surface while preserving the designed interface. This is done because purely hallucinated sequences can be difficult to purify experimentally.

      • By default, “soluble” MPNN weights are used, which usually result in a negatively charged binder surface. “Original” weights can be used for a more neutral surface.

    3. Monomer Reprediction: Uses the AlphaFold2 monomer model for the final prediction. Since this model was not trained on complexes, it serves as a highly stringent test of a well-defined interface.

    4. Binder Folding Check: The binder is also repredicted alone (without the target) to assess its structural change (RMSD) upon binding.

  • Final Analysis: PyRosetta is used to calculate interface scores and provide additional biophysical metrics for filtering (eg Rosetta dG, surface hydrophobicity, shape complementarity, number of interface residues, H-bonds and unsaturated H-bonds).

  • Output Structure: Designs are collected in sequential folders (Trajectory → MPNN → Accepted) with corresponding metrics in separate CSV files.

Target Preparation & Hotspot Selection

  • Input Structure:

    • Accepts experimental (NMR, CryoEM)
    • or predicted (AlphaFold2, Boltz-2 etc) PDBs.
      • Using an experimental structure as a template for a new AlphaFold2 or Boltz-2 prediction can help fill in missing loops.
    • Even small variations of the ‘same’ target (experimental vs. predicted vs. alternative trimming) can change in silico success rates significantly.

  • Trimming:

    • Trim the target to only essential domains to save memory and time. Remove flexible ends and large flexible loops if these are not involved in the binding interface.

    • Trim at realistic points like domain boundaries or hinge residues (e.g., Gly, Pro), which can be identified by high B-factors or low pLDDT scores.

    • “Unrealistic” trims (e.g., splitting GPCR helices) are supported if they preserve the local structural context and don’t expose core hydrophobic residues.

  • Hotspot Selection:

    • Define hotspots by residue number and chain ID (e.g., A23,A27-50,B45).

    • Targeting a radial patch of surface residues on secondary structures is recommended, especially sites containing hydrophobic residues (F, Y, W, I, L, M).

    • Single residues also work in well defined binding sites. You can also leave hotspots undefined and let the model choose.

  • Off-Target Binding: your hotspots may be ignored if the choice of target site is suboptimal or if there is a significantly better binding site nearby. To encourage your preferred binding site you can:

    • Mutate the unwanted binding site residues to lysines.

    • Trim away the off-target region.

    • Use the _hardtarget advanced setting.

    • Pre-block the off-target binding site with another binder ! First generate a design that binds to the off-target site, and use this new complex as a structural input for a new run.

Advanced Settings

  • Default Settings (default_): Extensively tested and recommended for the highest chance of success.

  • _hardtarget:

    • Sets "predict_initial_guess": true.

    • Biases the prediction by providing binder atom positions as a starting point. Can help rescue designs that fail after the MPNN step. (As of late 2025 this setting has not been systematically tested and is likely to reduce success rates).

  • _flexible:

    • Sets "rm_template_seq_design/predict": true.

    • Masks the target’s amino acid sequence, which allows for greater backbone flexibility and domain movements during design and reprediction.

  • _mpnn:

    • Sets "mpnn_fix_interface": false.

    • Allows ProteinMPNN to redesign and optimize the interface residues, rather than keeping the AlphaFold2-designed interface fixed.

  • _betasheet:

    • Enforces beta-sheet structures by penalizing helicity ("weights_helicity": -2.0).

  • _peptide:

    • Uses a 3stage algorithm instead of the default 4stage.

    • Changes multiple loss weights to favour peptide-like properties (e.g., increases helicity weight, reduces contact weights).

    • Does not use the radius of gyration (rg_loss).

  • "predict_bigbang": true setting:

    • For complexes > 600 residues, this setting facilitates more efficient prediction of large complexes in the final stage

Filtering

  • Default Filters: The default filters are robust and generally should not be changed, except when designing peptides.

  • Key Metrics: The most important metrics for filtering are pLDDT, i_pTM, and i_pAE.

    • The i_pTM score is a good binary predictor of binding, but values between 0.6 and 0.8 are a “grey zone” where predictions can be incorrect.

  • Affinity: None of the computational metrics are predictive of binding affinity.

  • Early Rejection: Trajectories are terminated early if they have:

    • CA clashes > 0 or Interface clashes > 25.

    • pLDDT < 0.7.

    • A “floating” binder with < 3 contacts to the target.

  • Binder RMSD Filter: This filter can be relaxed or disabled if more conformational flexibility or an “induced fit” binder is desired.

Default filters

Designs that satisfy these criteria are kept as ‘Accepted’:

  • AlphaFold2 Metrics
    • Average_pLDDT > 0.8
    • Average_pTM > 0.55
    • Average_i_pTM > 0.5
    • Average_i_pAE < 0.35
  • Rosetta Metrics
    • Average_Binder_Energy_Score < 0
    • Average_Surface_Hydrophobicity < 0.35
    • Average_ShapeComplementarity > 0.6
    • Average_dG < 0
    • Average_dSASA > 1
    • Average_n_InterfaceResidues > 7
    • Average_n_InterfaceHbonds > 3
    • Average_n_InterfaceUnsatHbonds < 4
    • Average_InterfaceAAs: K < 3
    • Average_InterfaceAAs: M < 3
  • Structural Metrics
    • Average_Binder_Loop% < 90
    • Average_Hotspot_RMSD < 6
    • Average_Binder_pLDDT > 0.8
    • Average_Binder_RMSD < 3.5

Design Selection

  • Quantity: Generate at least 100 designs that pass filters to ensure good sampling of interface diversity.

  • Ranking: Final designs in the Accepted/Ranked folder are ranked by i_pTM. This is for prioritization only; all designs that pass filters are high-quality candidates.

  • Sub-selection:

    • Prioritize sampling diverse interfaces over similar designs from the same trajectory.

    • Perform a final visual check in PyMOL or ChimeraX.

    • Consider repredicting top candidates in the context of the full biological complex and check that the target interface is free of post-translational modifications.

  • Experimental Success:

    • Screening ~10 designs is often sufficient to find a nanomolar binder.

    • Screening 50-100 designs is recommended when aiming for picomolar affinity.

    • Generated binders are typically well-behaved biochemically (thermostable, easy to purify).

Troubleshooting

  • Mismatched Atoms Error: Your input PDB likely has partial residues or atoms left over from trimming. Clean the file before rerunning.

  • Binder Ignores Hotspots: This is an intended feature to avoid forcing binders into poor sites. See the “Target Preparation” section for strategies to enforce binding to a specific site.