{
"created": "<string>",
"tokens_used": 123,
"target_sequence": "<string>",
"binder_format": "<string>",
"seed": 123,
"potential_sequence_of_concern": false,
"warning_messages": [
"<string>"
],
"hotspots": [
{}
],
"target_chain_ids": [
"<string>"
],
"binder_sequence": "<string>",
"self_scored": {
"ptm": 123,
"interface_ptm": 123,
"plddt_mean": 123,
"hotspot_min_distances_angstrom": [
123
]
},
"self_scored_by_replica": [
{
"model": "<string>",
"step": 123,
"ptm": 123,
"interface_ptm": 123,
"plddt_mean": 123
}
],
"generation_metadata": {
"recipe_version": 123,
"fold_model": "<string>",
"fold_replicas": [
"<string>"
],
"fold": {},
"schedule": {},
"steps_run": 0,
"stopped_reason": "completed"
},
"accounting": {
"gradient_steps": 0,
"gpu_seconds": 0
}
}Design result
Fields a completed design_binder item returns.
include_responses: true.
Read a task returns one record in data.response and a signed PDB download in data.artifacts.structure.
List responses never include those artifact URLs.
Interpretation
self_scored is the design models’ assessment of the complex they optimized.
It comes from the models the trajectory spent every step optimizing against, so it is inflated.
Rank a population with it, then judge the winners with independent folds of your own.
self_scored_by_replica preserves per-model disagreement that an average can hide.
hotspot_min_distances_angstrom follows the requested hotspot order.
A null distance is missing evidence, not a passing contact.
Check the echoed hotspot residue letters against your intended crop.
PDB numbering and chain labels can differ from request addresses, so retain the mapping you submitted.
generation_metadata records the recipe version, models, schedule, steps run, and stop reason.
It describes what ran; it does not let the caller select a historical recipe.
accounting reports gradient steps and GPU seconds, while the task’s tokens_used reports billed usage when available.ISO formatted date time
Number of tokens consumed in the request
Indicate that the returned sequence may be of potential concerns.
Warning messages returned by FastAPI.
The requested hotspots with the residue each one landed on, as {chain, position, residue}. Echoed in the same form the caller asked in so a wrong number surfaces as a wrong amino acid. The recipe's internal residue-only count is deliberately not returned: it coincides with the plddt slot index for a two-chain target and diverges for three, so a caller indexing with it would be silently wrong.
Caller-provided target chain IDs in target order, or the generated A, B, C IDs for target_sequence input. The PDB artifact uses positional chain labels, which can differ from caller-provided IDs.
The designed binder. Absent unless stopped_reason is 'completed' -- an interrupted trajectory may not have annealed to a discrete sequence.
The design model's own view of its final complex.
Free -- these come from the last step, which already computed them. But they come from the model the trajectory spent every step optimizing against, so they are inflated. Rank a population with them; judge the winners with independent critic folds.
hotspot_min_distances_angstrom is the weakest of them for a ranking, because
the epitope loss steers toward those exact residues. Most designs report
themselves on-epitope, so the distances separate a population far less than
interface_ptm does.
Show child attributes
Show child attributes
Per-replica view of the final complex, in the same order as generation_metadata.fold_replicas. self_scored above averages these; this is each of them, so a divergence stays visible.
Show child attributes
Show child attributes
What actually ran, so a result can be reproduced or explained.
Show child attributes
Show child attributes
Show child attributes
Show child attributes