| name | protein-binder-design |
| description | Design new proteins that bind a chosen surface, using BindCraft's AlphaFold2-guided hallucination or the RFdiffusion backbone plus ProteinMPNN sequence pipeline. Use this skill to specify a target epitope by hotspot residue, trim a receptor to the region worth designing against, set up a design campaign, and filter the output on the in-silico metrics that predict experimental success — interface predicted TM-score, predicted aligned error at the interface, buried surface area, and shape complementarity. Also trigger on BindCraft, RFdiffusion, ProteinMPNN, minibinder, hallucination, inverse folding, hotspot residue, epitope targeting, ipTM, or de novo binder. |
| license | MIT |
| allowed-tools | Read Write Edit Bash |
| compatibility | Requires Python 3.10+. The bundled scripts prepare target specifications and filter design metrics using only the standard library. Running a campaign needs BindCraft 1.5+ (MIT, from GitHub) with AlphaFold2 weights, or RFdiffusion plus ProteinMPNN, and an NVIDIA GPU — a single binder trajectory is tens of minutes. |
| metadata | {"version":"1.0","skill-author":"K-Dense Inc.","openclaw":{"emoji":"🧷","homepage":"https://github.com/martinpacesa/BindCraft"},"hermes":{"category":"research"}} |
De Novo Protein Binder Design
Designing a new protein that binds a chosen surface used to be a research project. BindCraft
reports 10–100% experimental success without high-throughput screening, and the pipeline is open
source. The hard part is no longer the algorithm — it is choosing where to bind, and knowing that
the metrics which select designs cannot tell you which one works.
Tools: BindCraft 1.5+ (Nature 2025, MIT), or
RFdiffusion + ProteinMPNN + AlphaFold2. Both need AlphaFold2 weights and an NVIDIA GPU; a single
trajectory is roughly half an hour. The bundled scripts prepare targets and filter output, and run
anywhere.
Read references/epitope-selection.md before anything else,
references/bindcraft-and-rfdiffusion.md to choose a
pipeline, and
references/filtering-and-validation.md before ordering —
that one is judgement, not syntax.
The three scripts
| Script | Answers |
|---|
binder_target_spec.py | Where should the binder bind, and is that site usable? |
design_manifest.py | Which pipeline, how many trajectories, what will it cost? |
binder_filter.py | Which designs survive, and which should I actually order? |
The epitope decides the campaign
Everything downstream is compute spent on this one choice, and a bad site produces designs that
fold beautifully and bind nothing — with no signal that the site was the problem.
python skills/protein-binder-design/scripts/binder_target_spec.py hotspots \
--pdb target.pdb --chain A --hotspots 45,47,52,89
# 4 hotspot residues, maximum separation 14.2 A
resseq resname neighbours exposure issue
45 TYR 16 surface
47 LEU 24 buried buried -- cannot be contacted
Three checks it applies: hotspots must be surface-exposed (a buried residue cannot be
contacted, and neither design tool will say so), there should be 3–6 of them, and they must sit
within ~25 Å — a wider spread is asking a single binder to do something impossible.
Then trim: designing against a 900-residue protein spends nearly all the compute on regions the
binder never touches. trim selects 100–200 residues around the epitope and warns outside that
band. Remove glycans and disorder first — neither is modelled, and both bias the interface toward
surface that is occluded in the real protein.
