BindCraft

BindCraft is a pipeline for de novo protein binder design from the Laboratory of Protein Design and Immunoengineering at EPFL, led by Martin Pacesa in Bruno Correia's group and published in Nature in 2025. Its premise is that a binder can be designed once and work — no library, no display selection, no affinity maturation. Earlier pipelines produced hit rates low enough that high-throughput screening was an obligatory second stage; BindCraft reports experimental success rates from 10% to 100% across targets, averaging 46.3%, from ordering roughly ten designs per target.

The method is hallucination, not generation from a learned prior. BindCraft backpropagates through the frozen weights of AlphaFold 2 Multimer, treating the structure predictor's interface confidence as an objective and descending on the binder sequence until the predictor is convinced the complex forms. Nothing is retrained: the same fixed composition of AlphaFold 2, ProteinMPNN, and PyRosetta was applied to every target in the paper — cell-surface receptors, allergens, de novo designed proteins, and multi-domain nucleases alike.

Because it needs no target-specific training and no known binding site, BindCraft has become a common base layer for later work. Two catalog entries build directly on it: BindEnergyCraft replaces its interface-confidence objective with a statistical energy derived from the same frozen predictor, and FoldCraft adds conditioning on a chosen binder fold. It competes with diffusion approaches such as RFdiffusion, which generate backbones before sequences rather than optimizing sequence through a predictor.

#Key Features

• One-shot design: Designs are ordered and tested directly rather than screened from a library, with the authors recommending ten designs from the top twenty ranked by interface pTM.
• No binding site required: Trajectories can be run without specifying hotspot residues, letting the optimization locate an interface on the target rather than being handed one.
• Composition of frozen models: AlphaFold 2 Multimer supplies the design gradient, a soluble variant of ProteinMPNN redesigns non-interface positions, and PyRosetta supplies physical scoring — none of the three is fine-tuned for the target.
• Explicit interface filters: Nine criteria gate every design, including pLDDT above 0.8, interface pTM above 0.5, interface predicted aligned error below 0.35, shape complementarity above 0.60, at least three interface hydrogen bonds, and surface hydrophobicity below 35%.
• Open and runnable: The MIT-licensed code ships with a Google Colab notebook, so a design campaign needs a CUDA GPU rather than a bespoke cluster.

#Technical Details

A trajectory begins with sequence optimization through AlphaFold 2 Multimer via ColabDesign, running roughly 140 iterations of stochastic gradient descent that pass through continuous, probabilistic, straight-through, and discrete sequence representations in turn. Successful trajectories are handed to the soluble ProteinMPNN variant, which redesigns the binder core and surface while fixing residues within 4 Å of the target and emitting 20 candidate sequences per trajectory. Each is repredicted with AlphaFold 2 monomer using three recycles and two template-based models, then relaxed with PyRosetta FastRelax over 200 iterations and scored by InterfaceAnalyzer before the nine filters are applied. Reaching 100 filter-passing designs typically requires sampling 300 to 3,000 trajectories. Experimental validation spanned twelve targets: SpCas9 (6 of 6 designs binding, roughly 300 nM by surface plasmon resonance), CLDN1 (6 of 7), PD-L1 (7 of 9), BBF-14 (6 of 11, best 20.9 nM), PD-1 (13 of 53), CD45 (4 of 16, best 14.7 nM), the allergens Der f7 (4 of 10, best 12.8 nM), Der f21 (4 of 7) and Bet v1 (2 of 7), IFNAR2 (3 of 9), the centriole protein SAS-6 (micromolar affinity), and the prokaryotic Argonaute CbAgo (2 of 12), whose best binder reached a 5 nM dissociation constant by biolayer interferometry and cut the nuclease's catalytic rate up to 80-fold.

#Applications

BindCraft generates therapeutic candidates against cell-surface receptors, neutralizers for allergens, affinity reagents for structural and cell biology, and modulators of gene-editing enzymes — the SpCas9 and CbAgo binders act as protein-based off-switches for CRISPR and Argonaute systems. Because it requires no structural knowledge of the binding site and no screening infrastructure, it is accessible to groups without display or high-throughput selection capability.

#Impact

BindCraft is among the most widely adopted binder-design tools in the field, with over 1,100 stars on GitHub and continued maintenance, and it has become the reference baseline newer binder pipelines report against and the loop several of them modify rather than replace. Its limitations are practical rather than conceptual. Success rates remain strongly target-dependent, so experimental testing of a small panel is still required. Runs need a CUDA GPU with 32 GB of memory recommended, and while BindCraft's own code is MIT-licensed, the PyRosetta dependency requires a separate license for commercial use. The pipeline also inherits AlphaFold 2's blind spots: it optimizes what the predictor believes about an interface, a proxy for binding rather than a measurement of it.