openfold2-nim

Files (13)

• BENCHMARK.md
• SKILL.md
• skillspector-baseline.yml · config/
• config.yml · evals/
• evals.json · evals/
• trigger_evals.json · evals/
• api.md · references/
• examples.md · references/
• parameters.md · references/
• science.md · references/
• validation.md · references/
• skill-card.md
• skill.oms.sig

SKILL.md (first lines)

# OpenFold2 NIM

Predict a single protein-chain structure from an amino-acid sequence, with
optional A3M multiple sequence alignments and mmCIF templates. Use this
guide for basic hosted/local NIM use; load supplemental files only when
the task needs deeper context:

- `references/api.md`: exact endpoints, schemas, Docker flags, response fields.
- `references/science.md`: model scope, strengths, limitations, and handoffs.
- `references/parameters.md`: MSA, template, model-selection, and relax effects.
- `references/validation.md`: artifact and scientific sanity checks.
- `references/examples.md`: compact hosted/local payload patterns.

## Choose Mode

Ask only when context is unclear:

> Hosted NVIDIA API or local Docker NIM?

- Hosted URL: `https://health.api.nvidia.com/v1/biology/openfold/openfold2/predict-structure-from-msa-and-template`
- Local URL: `http://localhost:8000/biology/openfold/openfold2/predict-structure-from-msa-and-template`
- Local readiness: `http://localhost:8000/v1/health/ready`

Mode difference: hosted and local use the same prediction path except local
does not include `/v1/`. Hosted requests use `Authorization: Bearer
$NGC_API_KEY`; local inference requests use no auth header after readiness.

## Auth And Environment

Do not print API keys. Confirm they exist with shell tests, not echoes.

Hosted needs `NGC_API_KEY` in the request header. Supported local Docker
startup uses `NGC_API_KEY`, or `NVIDIA_API_KEY` as a fallback, plus
`LOCAL_NIM_CACHE`. A repo-root `.env` file may be sourced as a local override.

## Local Docker

Use the official OpenFold2 NIM image and mount `LOCAL_NIM_CACHE` at
`/opt/nim/.cache`. Current docs recommend at least 80 GB disk, 64 GB system
RAM, 8 CPU cores, and one supported GPU; the container is roughly 55 GB and
first startup downloads about 10 GB of model parameters.

For the exact startup preflight (`.env` sourcing, `NGC_API_KEY`/`NVIDIA_API_KEY`
handling, `docker login`, and the `docker run` for
`nvcr.io/nim/openfold/openfold2:latest`), copy the command block in
[`references/api.md`](references/api.md) under **Local Docker** verbatim — do
not drop `.env`, `NGC_API_KEY`, `LOCAL_NIM_CACHE`, or the no-auth local request.

Readiness check:

```bash
until curl -sf http://localhost:8000/v1/health/ready; do sleep 5; done
```

## Request Pattern

Use Python `requests`; curl escaping is fragile for A3M/mmCIF text. The
`sequence` field is required. `input_id`, `alignments`, `selected_models`,
`relax_prediction`, `use_templates`, and `explicit_templates` are optional.

```python
import os
import requests

hosted = True
url = (
    "https://health.api.nvidia.com/v1/biology/openfold/openfold2/predict-structure-from-msa-and-template"
    if hosted
    else "http://localhost:8000/biology/openfold/openfold2/predict-structure-from-msa-and-template"
)
headers = {"Content-Type": "application/json"}
if hosted:
    headers["Authorization"] = f"Bearer {os.getenv('NGC_API_KEY')}"

seq = "MTEYKLVVVGAGGVGKSALTIQLIQNHFVDEYDPT"
payload = {
    "sequence": seq,
    "input_id": "kras_fragment",
    "selected_models": [1],
    "relax_prediction": False,
    "alignments": {
        "uniref90": {
            "a3m": {
                "alignment": f">query\n{seq}",
                "format": "a3m",
            }
        }
    },
}

response = requests.post(url, headers=headers, json=payload, timeout=300)
response.raise_for_status()
result = response.json()
```

Payload gotchas:

- OpenFold2 is monomer-only. For protein-ligand, protein-DNA/RNA, or
  multi-chain complexes, use OpenFold3 or Boltz2 instead.
- `sequence` must use valid amino-acid IUPAC symbols.
- Hosted API docs list sequence length 1-1000; local docs say current NIM
  supports sequences up to 2048 residues on supported hardware.
- A3M alignments go under `alignments` by database name, then `a3m` with
  `alignment` and `format`. When the user needs to create or deepen an MSA,
  hand off to `msa-search-nim` / MSA Search and map its A3M output into this
  `alignments` shape.
- Starting with OpenFold2 2.0.0, use `explicit_templates` with mmCIF content;
  do not write new HHR-template examples.
- `selected_models` chooses AlphaFold2/OpenFold parameter sets 1-5. Select one
  or two models for smoke tests; use all five for stronger production runs.

## Save And Interpret Output

The response includes one prediction per selected model, ordered by confidence.
Save every returned structure-like text field and the full JSON response so
field-shape differences are auditable. Production answers should explicitly
write `.pdb` or `.cif` artifacts, preserve the response JSON, and print any
confidence/ranking fields the service returns.

From NVIDIA/skills, Apache-2.0 AND CC-BY-4.0.

More from NVIDIA/skills