PD-L1 · BindCraft2
Smoke campaign
COMPUTATIONAL · NO WET-LAB DATA
PD-L1 · BindCraft2
Smoke campaign
COMPUTATIONAL · NO WET-LAB DATA
A bounded BindCraft2 campaign produced one 60-residue candidate that passed its computational filters. Explore the predicted complex, compare both refolds, and see what the evidence supports.
PD-L1 is grey (chain A). The designed binder is violet (chain B). Orange marks the requested hotspots, A54, A56, A66 and A115. Drag to rotate; switch to the rejected refold for comparison.
The accepted candidate merits structural review. Its 12 reported all-atom clashes require inspection; passing the upstream filters does not establish binding or affinity. No finite all-atom-clash cutoff was active in this run.
The run stopped after one accepted design, within its two-trajectory cap. One sequence passed; the other failed the interface-confidence filter.
The accepted refold has bound binder pLDDT 0.91 and unbound binder pLDDT 0.96, on the source 0-1 scale. These are model confidence scores.
Upstream reports a hotspot-contact fraction of 1.0 and 14 binder interface residues. Binder RMSD is 0.28 Å under the upstream calculation.
This integration smoke run does not estimate a campaign hit rate. Binding, selectivity, expression, stability and affinity have not been measured.
Both versions use AlphaFold2 gradients to optimise a binder sequence, ProteinMPNN to redesign it, and AlphaFold re-prediction to evaluate candidates. Held-out AF2 validation was already part of BindCraft 1. BC2 expands the objectives and molecular formats that can be expressed in a campaign.
| Aspect | BindCraft 1 documented workflow | BindCraft 2 documented workflow |
|---|---|---|
| Design brief | A binder against a supplied target structure, with optional hotspots. | Multiple desired targets, explicit off-target avoidance, and coldspots that should remain clear. |
| Molecular format | Focused on de novo protein binders. | Named presets for de novo binders, linear and cyclic peptides, scaffolded antibodies, repeat proteins and assemblies. |
| Target input | A target structure and selected chains. | Structures plus FASTA input for co-folding binders with disordered regions or motifs. |
| Ranking | ipTM recommended for prioritising accepted designs. | i_pDAE is the default ranking score; other confidence and geometry measurements remain part of review. |
A useful design brief can ask for recognition across two species while discouraging interaction with a specified off-target. BC2 lets desired and avoided targets participate in the shared sequence objective. Coldspots add a different constraint: keep a region of the intended target clear. These controls make the biological question more explicit, but only the supplied off-targets are assessed. Computational avoidance does not establish experimental specificity.
At Crane, our team uses Linus internally to support design work. The customer receives the design work and its supporting evidence. Choosing relevant target states, interpreting failures and deciding which candidates deserve an experiment remain central to that work.
One predicted complex produces several confidence scores. Each asks a different question. Their meaning comes from the calculation behind them, rather than the number of decimal places.
| Score | The question it addresses | BC2-02 | How to read it |
|---|---|---|---|
| pLDDT | How confident is AF2 in the binder's local structure? | 0.91 | Higher is more confident. BC2 reports 0-1 here; structure colouring uses 0-100. |
| i_pTM | How confident is AF2 in the relative arrangement of the interacting chains? | 0.74 | Higher is more confident. This is an interface-confidence score. |
| i_pAE | How uncertain is the relative placement across the interface? | 0.21 | Lower is better. BC2 divides the mean interface PAE by 31 Å: this value corresponds to approximately 6.5 Å. |
| i_pDAE | How confident is the predicted interface when the calculation focuses on nearby cross-chain residues? | 0.18 | Higher is better. BC2 uses this distance-masked interface-confidence score for its default ranking. |
In the implementation used here, BC2 selects cross-chain residue pairs whose Cα atoms are within 8 Å. It converts their predicted aligned errors into TM-like confidence values, with a distance scale based on the number of nearby partners. For each residue, it averages over those partners, then keeps the highest residue-centred score across both sides of the interface.
That makes i_pDAE a focused reading of interface confidence. Because the calculation keeps the best residue-centred summary, the scalar does not describe every part of the interface equally. It also reuses AF2 uncertainty and the predicted geometry; it does not add an independent predictor's vote. The exact calculation is in the pinned scoring code.
i_pDAE 0.18 does not mean an 18% chance of binding. Its scale cannot be interpreted using i_pTM or ipSAE cutoffs. In this run, 0.18 ranks ahead of the other candidate's 0.11; two candidates provide no experimental calibration of either score.
For BC2-02, confident local structure, the preferred i_pDAE ranking and contact with the requested hotspots support further review. The reported interchain clashes remain a reason to inspect the pose. We would then seek additional prediction evidence and assess candidate diversity and developability before choosing designs for physical testing.
This is the judgment behind the shortlist: connect each score to the question it can answer, inspect the structure, and identify the experiment needed to resolve what remains uncertain.
Confidence in a cropped complex leaves a practical question open: is the same surface accessible in the biological setting? This PD-L1 run used a 115-residue fragment. Before nominating a sequence, we would place the predicted pose in full-length target context and inspect glycan positions, membrane orientation, neighbouring domains and the intended assay geometry.
These are proposed review steps, not completed analyses of this candidate. A plausible interface can still be obstructed in an assay or on a cell. The upstream scoring guidance makes the same distinction between prediction confidence and biological accessibility.
A controlled comparison would match the target structures, epitope, binder lengths and compute budget, repeat independent seeds, and evaluate both outputs with common criteria. GPU time and prediction calls should be recorded alongside candidate counts. Different default filters can change how many designs are accepted without changing experimental success.
BC2 provides a benchmark profile that fixes the campaign seed and disables both autotuning and the separate fallback ladder called desperation. That ladder can relax design and validation conditions after prolonged failure, so it matters when interpreting acceptance. A multi-seed benchmark would explicitly vary the seed while retaining fixed comparison settings. These controls are documented in the BC2 settings reference.
Our smoke test demonstrates a working integration that returns inspectable candidates and scores. It does not establish higher hit rates, better affinity or faster design than BC1. Those claims require a matched benchmark and, for binding performance, physical measurements.
AF2 gradients, ProteinMPNN redesign and held-out AF2 prediction.
Inspect clashes, expand independent trajectories and obtain an additional predictor check.
After nomination: synthesis, expression and binding measurements with specificity controls.
PD-L1 chain A (115 residues, numbered 18-132), hotspots A54/A56/A66/A115, binder length 60, seed 42. Two ProteinMPNN draws per successful trajectory; one accepted-design goal.
| Candidate | Upstream outcome | i_pDAE | i_pTM | pLDDT | All-atom clashes | Crane review |
|---|---|---|---|---|---|---|
| BC2-02 | Passed | 0.18 | 0.74 | 0.91 | 12 | review |
| BC2-01 | Rejected: i_pTM | 0.11 | 0.70* | 0.90 | 14 | reject |
SRDEARLEEMVERFREALARGDLHFARFWLWAMRFILSRSGDPAVRARFEPVLRELEGLL
This is the designed amino-acid sequence, without an expression tag, linker or cloning construct.
i_pDAE is the upstream distance-masked interface-confidence ranking score. It is not ipSAE. i_pAE is normalized here; 0.21 must not be read as 0.21 Å. Report pLDDT uses 0-1; structure B-factor fields use 0-100.
Design used five AF2 multimer-v3 parameter sets; refolding used model_1_ptm and model_2_ptm. Held-out checkpoints reduce direct optimization bias, but this is not independent cross-family validation.
Complex chains A/B map to target/binder. Binder residues retain upstream numbering 0-59. Six returned mmCIF structures passed parser and finite-coordinate checks. Report views use the two refold complexes.
The report packages a deployed workflow smoke test. It is not an experimental validation or a customer-specific design campaign. No biological controls or cross-predictor consensus were run.
Runtime: A100-80GB. Modal call fc-01M33Q722T3QN0HME0ER7S04BZ.
BindCraft2 source: 18a9042fbe9a5373a5b7d98fe82335127c2fd70d.
Accepted complex SHA-256: e05a9c4da56685d293859a4deabfb051bd85ab3d93f71bd2846f3859620047ce.