Protein Binder Design

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

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.

ipTM is not pTM

The most consequential confusion in reading design output. pTM scores the whole complex and is dominated by a large well-folded target; ipTM scores the interface. A design can have excellent pTM and no interface at all.

python skills/protein-binder-design/scripts/binder_filter.py filter --csv metrics.csv --all

design  iptm  ipae  plddt  dsasa  shape_complementarity  unsat_hbonds  passes  failures
d1      0.88  7.2   88     1450   0.62                   2             true
d2      0.61  14    72     800    0.48                   7             false   iptm<0.8|ipae>10|plddt<80|...

These are a conjunction. Published success rates come from designs passing all of them together; sorting by ΔSASA alone gets you large, loosely packed interfaces.

The metrics cannot pick the winner

They come from the same model family that generated the designs — AlphaFold2 hallucination optimises until AlphaFold2 is confident, then AlphaFold2 confidence judges the result. That circularity is not fatal (the correlation with experimental success is the empirical finding the field rests on) but it means the metrics are self-consistency measures, not affinity predictions, and among survivors they cannot rank.

So the plate is the experiment:

python skills/protein-binder-design/scripts/binder_filter.py diverse --csv metrics.csv --n 24

Ranking by ipTM alone returns near-identical designs — a top-24 list can be one solution tested 24 times. diverse enforces a pairwise identity ceiling. Order at least 20.

Cost is trajectories per ordered design

python skills/protein-binder-design/scripts/design_manifest.py plan --want 24 --pass-rate 0.03 --gpus 4

# 801 trajectories to expect 24 filter survivors at a 3% pass rate
# 400.5 GPU-hours -> 4.17 days on 4 GPU(s)

The filter pass rate is the hidden cost, and it is target-dependent. Set the trajectory count from it, not from the number of binders you want.

BindCraft or RFdiffusion

BindCraft co-folds binder and target at every iteration, so target flexibility is accounted for and no known binding site is needed. RFdiffusion generates against a fixed target, so induced fit is invisible to it — but it gives explicit control of binder length, fold, and secondary structure. On a flexible epitope, prefer BindCraft.

Four things to expect

1. Most designs fail, and that is normal. A 10% success rate means nine of ten do not bind.
2. Affinity confirms binding, not the model. A binder can bind well through an interface entirely different from the designed one. Only a structure tells you.
3. Specificity is not predictable. Designs frequently bind close paralogues; counter-screen early, before optimisation.
4. Nothing passing the filters usually means the epitope, not the trajectory count. More compute against a bad site produces more confident failures.

Composing with the rest of the bundle

• uniprot-rcsb → before: an experimental structure beats an AlphaFold model, which biases toward closed apo states with unreliable surface side chains.
• binding-site-analysis → before: the hydrophobic-patch logic transfers, though protein-protein interfaces are flatter than small-molecule pockets and score lower.
• esm → alongside: sequence-level sanity checks on the designs.
• immunogenicity → after, not optional: a de novo binder is entirely non-germline.
• glycoengineering → after: check the designs for introduced N-glycosylation sequons.
• adaptyv → after: BLI/SPR on the plate you designed.
• tamarind → instead: runs both pipelines in the cloud when there is no local GPU.

Reporting results honestly

Give every filter and threshold and say they were applied as a conjunction. Report trajectories run and survivors — the pass rate is the informative number. State that the metrics come from the same model family that produced the designs. Report ordered, expressed, and bound as three separate counts, because that chain is what the campaign actually delivered. Never quote ipTM as a predicted affinity.