Skip to main content
All blog

Blog · August 19, 2026 · 2 min read

FreeBindCraft: open-source binder design on CogniChem

FreeBindCraft runs BindCraft's de novo binder design loop with open late-stage scoring. CogniChem hosts it as a GPU job against a target PDB, with miniprotein or peptide modes and workflow chaining.

By CogniChem Team

Designing a protein that binds a chosen target is now a computational experiment rather than a purely empirical search. The difficulty is not inventing a fold in isolation. It is closing a loop from a target structure to many ranked complexes, each generated, sequenced, relaxed, and scored at the interface. That loop has to hold at campaign scale, where a single target may produce dozens or hundreds of candidates before any construct is ordered.

BindCraft, introduced by Pacesa, Nickel, Schellhaas, and colleagues and maintained at martinpacesa/BindCraft, is the method that made this loop practical. AlphaFold2, used through ColabDesign, proposes binder geometry against the target. ProteinMPNN then assigns sequences. A later stage relaxes the complex and asks whether the interface is geometrically and chemically plausible. The original later stage used PyRosetta, which is closed-source. FreeBindCraft, an MIT-licensed fork of BindCraft 1.52 by Aaron Ring, keeps the same design sequence and performs that later stage with open tools, so the campaign can run from a PDB to ranked complexes.

The later stage has two jobs. First it removes clashes and rebuilds side chains so the complex is a fair input to interface measurements (OpenMM for relaxation, FASPR for packing). Second it scores the interface: geometric fit with sc-rs, solvent-accessible and buried surface with FreeSASA and Biopython, and, if requested, ipSAE from Dunbrack et al. (2025), which reads AlphaFold's predicted alignment error at contacting residues. The authors set out how those measurements were chosen, and how they compare with Rosetta-derived values, in their technical overview and README. BindCraft remains the design method; FreeBindCraft is the open path through the scoring stage.

On CogniChem, FreeBindCraft runs as a GPU job. You provide a target PDB, choose a structured miniprotein or a short peptide, and select a GPU tier. Those two classes use different filter and advanced-settings profiles because a short peptide and a folded miniprotein do not present the same backbone flexibility or interface geometry. Ranked structures are the job result. The same job type is a node in a workflow, so a predicted or prepared structure can feed design, and design can feed docking or simulation, without a separate file transfer.

The practical consequence is that a binder campaign can be iterated on the target: change the structure or the design class, recover ranked complexes, and pass them into the next calculation. Cite Pacesa et al. for BindCraft and Ring for the open scoring path. The job is on the toolkit; create an account to submit one.

Sources