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. 2. `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. 3. `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`. 4. 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 | Need | Capability | | --- | --- | | Protein or complex structure from sequence, ligands and ions allowed | `boltz2` | | Monomer structure from an MSA (`msa-search` first) | `openfold2` | | Complex structure with templates | `openfold3` | | Protein–ligand poses for a given pocket | `diffdock` | | Sequences for a fixed backbone | `proteinmpnn` | | New backbones or binders against a target | `rfdiffusion` | | Small molecules from a fragment or property target | `genmol`, `molmim` | | DNA sequence generation and scoring | `evo2` | 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.