synthetic-sciences/openscience/bionemo-nims

NVIDIA BioNeMo NIMs (hosted)

The ten BioNeMo NIM adapters run against NVIDIA's hosted endpoints with the user's own NVIDIA API key. Every request goes through the scientific_capability tool; nothing here installs software or starts compute. The capability ids are boltz2, openfold2, openfold3, msa-search, diffdock, proteinmpnn, rfdiffusion, genmol, molmim and evo2.

Before the first request

• scientific_capability with action: "list" shows the catalog and each entry's maturity.

action: "doctor", id: "<id>" says whether the NVIDIA key is connected and the endpoint is reachable. Without a key, stop and tell the user: connect one under Customize → Connectors → NVIDIA API (an nvapi-… key from build.nvidia.com); requests are billed to that NVIDIA account under NVIDIA's terms. Do not look for the key in the environment, in compute targets, or in Modal secrets; the hosted route is separate from the self-hosted BioNeMo containers that protein-binder-design runs on Modal.

• action: "describe", id: "<id>" returns the hosted request_schema, the endpoint, the API

schema version and the terms. Build payload from that schema exactly; field names and enumerations differ between NIMs (a sequence list for Boltz-2, a PDB string plus ligand SMILES for DiffDock, an alignment for OpenFold2).

Each request

1. action: "plan", id, payload validates every cross-field requirement locally. Nothing is

sent and nothing is spent. Fix what it reports before going on.

1. action: "start", id, payload, name, purpose sends the request. The user sees an approval

card bound to this exact request: endpoint, schema version, payload size, request hash and what leaves the device (sequences, structures, ligands). The approval is one-time; a changed payload asks again. Say in purpose what the result is for.

1. action: "wait", job_id (or status) until it settles; long predictions poll NVIDIA's

status endpoint for you. action: "artifacts", job_id lists the delivered files with their SHA-256; read a structure or JSON from there, and save what the user keeps with artifact save_file.

1. Report the model, the API schema version from describe, the request hash and the

artifact hashes with the result. A predicted structure or pose is a prediction: give its confidence (pLDDT, iptm, DiffDock confidence) and never call it experimental validation.

Choosing a NIM

For a full binder campaign (RFdiffusion → ProteinMPNN → Boltz-2 refold → ranking), load protein-binder-design, which chains these the way the BioNeMo Agent Toolkit does.