BindCraft

What is BindCraft?

BindCraft designs de novo protein binders against a target structure by combining AlphaFold2-guided backbone hallucination, ProteinMPNN sequence redesign, and PyRosetta interface scoring. It is built for cases where the goal is not to rank existing binders, but to generate entirely new mini-proteins that recognize a chosen surface patch.

The method is strongest on structured protein surfaces with clear geometric constraints. It is less reliable on highly flexible, disordered, or poorly resolved targets, where the hallucinated binder can optimize against an uncertain surface.

How to use BindCraft online

BindCraft on ProteinIQ accepts a target PDB structure plus chain selection, runs the native multi-stage binder-design pipeline in the cloud, and returns ranked binder-target complexes, full accepted-design metrics, and the remaining native CSV and artifact files for download in one job.

Inputs

Settings

Outputs

How does BindCraft work?

AlphaFold2-guided hallucination

BindCraft uses AlphaFold2-Multimer as an optimization target rather than only as a predictor. The sequence is updated iteratively so that the predicted complex improves interface confidence, binder compactness, and other structural objectives.

Hotspot residues change the loss function, not the final scoring rule. A hotspot list can steer the search toward a preferred epitope, but the model can still drift toward a nearby surface if that geometry produces a more favorable predicted interface.

Multi-stage optimization

The 4stage protocol moves through progressively more discrete sequence representations. Early stages optimize softer sequence distributions, later stages enforce one-hot amino acid identities and mutation steps. The shorter 3stage and 2stage presets reduce compute cost by dropping some of that optimization path.

This matters most on difficult targets. Easy targets often produce acceptable binders under more than one preset, while hard interfaces benefit from the broader exploration provided by the four-stage search.

ProteinMPNN redesign and binder-alone validation

After a promising backbone is found, ProteinMPNN redesigns the binder sequence around that geometry. native BindCraft then re-predicts the binder-target complex and also predicts the binder alone, which is why the accepted-design table includes both interface metrics and binder-only validation metrics such as Average_Binder_pLDDT, Average_Binder_pTM, Average_Binder_pAE, and Average_Binder_RMSD.

Filtering and relaxation

native filtering is applied at multiple points. Early AlphaFold checks screen out weak trajectories, final acceptance combines the Average_* and model-indexed thresholds, and PyRosetta relaxation adds interface energetics and packing metrics such as Average_dG, Average_PackStat, Average_ShapeComplementarity, and Average_dSASA.

Understanding the results

The viewer shows a compact subset of fields for fast triage. The Data tab now contains the full accepted-design row from final_design_stats.csv, which is the best place to inspect why one design outranks another.

A run can also finish with no accepted designs. This is still a valid native BindCraft outcome when the reserved credit runtime budget ends before any trajectory passes the active filters. In that case, inspect failure_csv.csv and trajectory_stats.csv, then reserve more credits, relax filters, or lower pLDDT, pTM, and i_pTM thresholds for exploratory runs.

Metrics that usually matter first

Practical interpretation

High-confidence designs usually combine strong Average_i_pTM, acceptable Average_pLDDT, and favorable interface energetics. A design with excellent Average_dG but weak Average_i_pTM often reflects a Rosetta-favored interface built on an uncertain AlphaFold geometry. A design with strong interface confidence but poor binder-alone metrics may bind only in the model because the isolated binder is unstable.

The native rank in final_design_stats.csv should be treated as the starting point, not the final answer. For experimental follow-up, the most useful shortlist often balances:

• Interface confidence from Average_i_pTM
• Structural confidence from Average_pLDDT
• Packing and energetics from Average_ShapeComplementarity, Average_PackStat, and Average_dG
• Binder-alone stability from Average_Binder_pLDDT, Average_Binder_pTM, and Average_Binder_RMSD

When to use BindCraft vs alternatives

BindCraft is for de novo binder generation. It is not the right choice when a sequence already exists and only structure prediction, affinity ranking, or docking is needed.