NVIDIA protein model outputs

NvidiaNIM_proteinmpnn designs sequences from a backbone; NvidiaNIM_esmfold predicts a monomer structure from a sequence. They require NVIDIA_API_KEY. The output format is selected by the tool configuration, so an HTTP text/plain response does not automatically mean PDB coordinates.

Reading results

Check status before saving or evaluating a prediction. HTTP 200 can contain an inner model failure; ToolUniverse returns those responses with status="error" and retains the native JSON in data. An empty or malformed PDB JSON envelope is also an error, rather than a successful structure.

ProteinMPNN JSON responses retain the native fields (including mfasta and any scores) under data. Plain Multi-FASTA responses expose the complete text as both data["mfasta"] and sequences, with format="mfasta".

ESMFold returns PDB text as data and structure, with format="pdb". When a JSON response contains multiple PDB strings, structures retains all of them in their original order; structure remains the first one for existing callers. A valid single-structure result omits structures.

# result is the complete response from tu.run_one_function(...).
if result.get("status") != "success":
    raise RuntimeError(result.get("error", "Prediction failed"))

if result.get("format") == "mfasta":
    fasta = result["data"]["mfasta"]
elif result.get("format") == "pdb":
    pdbs = result.get("structures", [result["structure"]])

In a multi-model workflow, retain each original response and all returned samples alongside model names, versions, inputs and seeds. Count failed calls separately from completed predictions. Structural confidence scores alone do not establish binding affinity or pH selectivity.

NVIDIA documents ProteinMPNN’s output as Multi-FASTA in its ProteinMPNN overview.