BindCraft 2: A Broader Brief, a Closer Look at the Scores

Before ordering a designed binder, a team needs to know what each score supports. Does the sequence fold? Is the predicted interface convincing? Can the molecule reach that surface in the assay?

We integrated BindCraft2 into the Crane Protein Intelligence Platform, which our team uses internally to deliver protein design work. In a deliberately small PD-L1 smoke test, one trajectory produced two ProteinMPNN candidates and one passed the configured computational filters. The accepted 60-residue design, BC2-02, had interface confidence i_pTM 0.74 and distance-masked interface confidence i_pDAE 0.18.

That is enough to inspect the result and verify the workflow. There is no experimental binding, affinity, selectivity or developability measurement for this candidate. One trajectory cannot establish a campaign hit rate. The practical decision is further review before any synthesis shortlist.

BindCraft is a pipeline. AlphaFold2 guides sequence optimisation toward a predicted complex, ProteinMPNN proposes sequences for the resulting backbone, and AlphaFold2 re-predicts the candidates. Validation with different AF2 models was already part of BindCraft 1; it is a useful check within the same model family, not an independent prediction from another model family. The main change to explain is the range of questions the documented workflow lets a designer ask.

Comparison of the documented workflows, not a claim about every custom extension or fork. Sources: BindCraft 1 reference and BindCraft2 design guide. The broader formats and objectives were not tested in this PD-L1 run.

Both numbers use a 0-1 scale, but they summarise different information. They are not probabilities of binding. An i_pDAE of 0.18 does not mean an 18% chance of success, and an ipTM cutoff cannot be transferred to i_pDAE.

In the pinned BC2 implementation, i_pDAE selects cross-chain Cα contacts within 8 Å and converts their predicted aligned errors into TM-like confidence terms. It averages those terms around each residue, with a distance scale set by that residue's contact count, then takes the maximum residue-anchored summary across both chain directions. It is not an average over the entire interface, and it is not ipSAE.

These scores reuse the structure predictor's outputs. Agreement between them is not equivalent to agreement between independent experiments. Confidence values in the CSV are averaged over the validation models used; geometric measurements come from the first model. A candidate rejected early may have only a partial validation ensemble.

Definitions: BC2 measurements and the i_pDAE implementation. The interactive report includes the rejected candidate, BC2-01, which failed the i_pTM filter; its rounded score of 0.70 conceals the difference used by that filter.

BC2-02 had 14 contacting binder residues and a 0.28 Å aligned displacement between its free and bound predictions. Those observations support a compact structural hypothesis. They do not establish binding.

The same result contains 12 interchain atom pairs within 2.5 Å. BC2 calls these all-atom clashes. It reports zero backbone clashes, a narrower count of interchain Cα pairs within 2.5 Å. The all-atom limit was not a mandatory rejection threshold in this run, so computational acceptance did not remove this concern. Atomic review and any justified relaxation must precede another assessment.

A confident fold still needs an accessible interface and a testable biological hypothesis. For PD-L1, that means checking the design in the full-length target context, including glycans, the membrane and the proposed assay geometry. A cropped target makes a calculation smaller; omitted biology can still block the approach. The upstream design guide makes this review part of candidate selection.

The broader BC2 brief is useful when a project needs recognition across chosen orthologues while avoiding a related protein. Desired target states and explicit detargeting states let the designer represent both aims; coldspots mark surfaces that should remain clear. These objectives describe the intended experiment. A favourable prediction against the selected states still needs binding and counter-screening measurements.

Our team uses that brief to choose the design format, inspect alternative poses and propose a test panel. Customers receive design work and its supporting evidence. The PD-L1 example here tests a single-target integration, not those broader specificity capabilities.

This smoke test does not establish that BindCraft2 is better than BindCraft 1 in speed, affinity or experimental hit rate. A fair comparison would hold the target, epitope, binder lengths, compute budget and evaluation criteria fixed, repeat across seeds and assess both outputs with the same downstream checks. Counts passing different filters are not comparable hit rates.

BC2's benchmark profile fixes its campaign seed and switches off autotuning and the "desperation" ladder. That ladder can ease the design task and its validation after prolonged failure. Checking the recorded settings matters when comparing candidates, even within one version. The profile helps control a comparison; it does not supply the comparison itself.

The report contains both candidates, their predicted complexes, score definitions and review decisions. The standalone explorer supports side-by-side inspection.

• Open the interactive report
• Open the structure explorer

Run and score definitions use BindCraft2 revision 18a9042. BindCraft 1 documentation is pinned to 7713aa0. The report records the computational result, including its unresolved clashes; it is not an experimentally validated binder package.