BioNeMo — the first scientific-computing adapter
speq/adapter-bionemo@0.1.0
A containerised biomolecular model served as a NIM microservice, reading a versioned reference database and producing a structure prediction with a per-residue confidence score.
What this page does not claim
An adapter describes what SPEQ can capture about one tool family. It is not an assessment of that tool, not a statement that the tool is fit for any use, and it carries no affiliation with, endorsement by, or approval from the vendor.
An adapter is thin on purpose
The architectural bet behind this layer is that the questions a QA unit must answer about an AI invocation are the same whatever the tool is, and that only the part which knows how to pin, invoke and read one tool family changes. This is that part, for one family. The engine that asks the questions contains no branch on tool family, and a test that reads the engine’s own source asserts it — so the bet is falsifiable rather than asserted.
A run of this family is seeded reproducible, which decides what can honestly be claimed about reproducing it. That is a property of the family, declared by the adapter, and the engine reports it rather than softening it.
THE ENDPOINT
- OpenFold3 structure prediction
- One allowlisted path, never a caller-supplied one — a request carrying SPEQ’s credential to an address a caller named is server-side request forgery with extra steps.
- Path documented at docs.nvidia.com/nim/bionemo/openfold3/latest (NIM 1.5.0) and probed 2026-09-03: 401 unauthenticated, against invented paths returning 404.
- Documented confidence fields: confidence_score, complex_plddt_score, complex_pde_score, ptm_score, iptm_score. A field the provider does not return is recorded as missing, never as zero — zero is a measurement at the bottom of a scale, and “not measured” is not a measurement.
WHAT THE OPERATOR MUST SUPPLY
Two questions cannot be answered by the tool at all. They are still asked: an unanswered question lands in the record’s gaps, with a consequence and a remediation path, rather than quietly disappearing from the dossier.
What regulated decision does this output inform?
Intended use is a property of the regulated decision the output feeds, which lives in the operator’s process and not in the tool.
What must a qualified person review and sign, and what may the agent never decide?
The approval boundary is an organisational decision. The tool can record that a review happened; it cannot decide who is qualified to perform it.
PILOT INPUT LIMITS
- Up to 512 residues per sequence, standard 20-letter alphabet only. The model’s own documented ceiling is far higher; this is deliberately tighter.
- Up to 2 molecules in one complex.
- Up to 2 independent structures per run.
- Up to 5 runs per account per day, claimed atomically before the call so a retry cannot spend twice.
- Public, synthetic or formally de-identified inputs only, enforced by the request schema at the door — a policy that depends on the sender behaving is not a boundary.
“WHAT DOES THAT CONFIDENCE NUMBER MEAN?”
The question this family actually gets asked, and the one a vendor datasheet does not answer. A model’s confidence is a statement the model makes about its own output; it is not a probability that the output is right, and it is generally not comparable between runs or between models. Each note below is prose a reviewer reads — never a threshold the engine applies on the tool’s behalf.
complex_plddt_score
The average per-residue confidence the model assigns to its own predicted local structure, across the whole complex.
WHAT IT DOES NOT MEANIt is not a probability that the fold is biologically correct and it is not experimental evidence. A high pLDDT on a wrong fold is a common and expected outcome, and the score is produced by the same model whose output it describes.
NOT COMPARABLE ACROSS RUNS OR MODELS
ptm_score
The model’s predicted TM-score for the complex as a whole — a global fold-similarity estimate.
WHAT IT DOES NOT MEANIt is not a measured TM-score against a solved structure. It is a prediction about a comparison nobody has made, so it cannot be treated as agreement with an experimental result.
NOT COMPARABLE ACROSS RUNS OR MODELS
iptm_score
The predicted TM-score restricted to the interfaces between chains in a complex.
WHAT IT DOES NOT MEANIt is not evidence that two chains actually interact. A predicted interface confidence says nothing about whether the complex forms in vivo, at concentration, or at all.
NOT COMPARABLE ACROSS RUNS OR MODELS
complex_pde_score
The average predicted distance error across the complex, in the model’s own units.
WHAT IT DOES NOT MEANIt is not a measured coordinate error and it is not a tolerance. Lower is better within one run; it does not convert into an accuracy claim about the structure.
NOT COMPARABLE ACROSS RUNS OR MODELS
confidence_score
The ranking score the model uses to order the diffusion samples it generated, best first.
WHAT IT DOES NOT MEANIt ranks samples against each other within a single run. It is not a quality bar, does not compare across runs or models, and a best-of-five ranking says nothing about whether the best of five is any good.
NOT COMPARABLE ACROSS RUNS OR MODELS
“WHAT INVALIDATES THIS QUALIFICATION?”
- Container image digest changes
- Model weights revision changes
- Reference database version changes
- Deployment moves between local and hosted
- GPU class changes where numerics differ
Declared by the adapter, evaluated by the engine, and written into the record’s own lapse conditions — so the answer to this question travels with the document rather than living in whoever remembers it.
WHAT IS PINNED, AND WHAT A CHANGE COSTS
container_image
CONTAINER DIGEST
A change here lapses the whole record.
model_weights
MODEL VERSION
A change here lapses the whole record.
reference_database
DATABASE VERSION
A change here re-opens only the questions its value supports.
gpu_class
HARDWARE CLASS
A change here re-opens only the questions its value supports.
endpoint_uri
ENDPOINT URI
A change here lapses the whole record.
FREQUENTLY ASKED
Does an adapter mean SPEQ has assessed this tool?
No. An adapter states what SPEQ can capture about a tool family — which versions to pin, what a returned metric may be read as, which questions the tool cannot answer about itself. It is a description of the governance surface, not a judgement about the tool, and it says nothing about whether the tool is suitable for any regulated use.
Is SPEQ affiliated with, endorsed by, or approved by the vendor?
No, in all three directions. The adapter was built against publicly documented interfaces, the endpoint path was verified by probing rather than assumed, and no vendor has reviewed or approved anything on this page. The vendor’s name appears because naming the tool accurately is the point of a scoped record.
Why are the input limits tighter than the model’s own?
Because the question this build answers is whether the governance path holds, not how large a complex the model can fold. A ceiling is also the simplest defence against cost abuse that does not depend on anyone remembering to check, and it is enforced by the request schema at the door rather than by asking callers to behave.
What happens to the scientific data?
SPEQ stores content addresses, the derived band and verdict, the observed metrics and what was missing. Not the input, not the predicted structure, not the provider’s response body. SPEQ is the system of record for the governance evidence and explicitly not for the customer’s scientific data, and keeping the payloads out of the table is what makes that a property of the system rather than a sentence in a policy.
What if the provider does not respond?
The run is recorded as attempted with no result, the verdict is forced to indeterminate, and the record says why. That is a governed outcome, not an error page: a quality system can file “it was attempted and produced no verdict”, and it cannot file a failed HTTP request.