ProteinMPNN NIM

<!-- nv-carps: dummy edit to trigger NIM skill validation. -->

Design protein sequences for a supplied backbone PDB. Use this guide for first-pass hosted/local usage; load supplemental files only when needed:

• references/api.md: exact endpoints, schemas, Docker flags, response fields.
• references/science.md: inverse-folding uses, limits, and validation.
• references/parameters.md: design controls, fixed positions, sampling.
• references/validation.md: FASTA, score, and structure checks.
• references/examples.md: compact hosted/local request patterns.

Choose Mode

Honor the user's explicit mode; otherwise use the configured runtime. NIM_API_MODE=local selects the local service at PROTEINMPNN_NIM_URL; the URL defaults to http://localhost:8000 for a NIM running in the same host or container. Ask only when neither the environment nor the user's request makes the mode clear:

Hosted NVIDIA API or local Docker NIM?

• Hosted: https://health.api.nvidia.com/v1/biology/ipd/proteinmpnn/predict
• Local: append /biology/ipd/proteinmpnn/predict to PROTEINMPNN_NIM_URL (default base URL: http://localhost:8000).

Local inference paths do not include /v1/. Hosted requests use Authorization: Bearer $NGC_API_KEY. Supported local Docker startup uses NGC_API_KEY (or NVIDIA_API_KEY via the preflight) for registry login, entitlement checks, and first-run model downloads; pass it into the container with -e NGC_API_KEY. Local inference requests use no auth header after readiness. Warm-cache key-free startup varies by image/version and should not be assumed.

Data Transfer and Authorization

Before a hosted request, tell the user that the entire PDB file and design parameters will be uploaded to NVIDIA's hosted API at the endpoint above. Proceed if the user has explicitly requested hosted processing of that PDB or already approved the transfer; otherwise ask for confirmation before submitting. For confidential structures, recommend a local NIM in the user's approved environment. A configured local URL may point to another machine; use only the configured or user-selected destination. Do not switch from local to hosted processing without the user's authorization.

The client reads NIM_API_MODE, PROTEINMPNN_NIM_URL, and, for hosted mode, NGC_API_KEY from the environment. It sends the key only in the HTTPS Authorization header to the hosted endpoint; local inference sends no key. Keep credentials out of logs and saved artifacts. The output directory contains the full input PDB in request.json and the returned sequences and scores, so use a location appropriate for the input's sensitivity. See references/api.md for endpoint and data-handling details.

Local Docker

For local setup, run the full sequence — env preflight, docker login, docker run, readiness loop, then the no-auth localhost request; do not answer with only a localhost Python request. For the exact preflight (.env sourcing, NGC_API_KEY/NVIDIA_API_KEY handling, and the docker run for nvcr.io/nim/ipd/proteinmpnn:latest), copy the command block in references/api.md under Docker Reference verbatim. This NIM's cache mount is /home/nvs/.cache/nim, not /opt/nim/.cache. When PROTEINMPNN_NIM_URL is supplied, the service is already managed elsewhere; use that URL and do not start another Docker container.

Readiness:

proteinmpnn_nim_url="${PROTEINMPNN_NIM_URL:-http://localhost:8000}"
until curl -sf "${proteinmpnn_nim_url%/}/v1/health/ready"; do sleep 5; done

Instructions

For a request to execute a design, run scripts/design.py and inspect its results. Writing a request script alone does not complete an execution request. If the user asks only for code or setup instructions, provide those without submitting an inference request.

1. Use the user's PDB path and requested sequence count. The client reads the entire PDB into input_pdb; do not replace or truncate the supplied backbone.
2. Select --mode hosted or --mode local and follow Data Transfer and Authorization above before submitting. Hosted mode uploads the PDB to the documented NVIDIA endpoint and requires NGC_API_KEY in the environment. Check only whether the key is set; do not print it, dump the environment, or save authentication headers. Local inference sends no authorization header.
3. Choose a new --output-dir for each request. The client reserves it before submitting, preserves the raw response for diagnostics, and validates the designed sequence count and score alignment before reporting completion.
4. Read summary.json and report the actual results described below. If the request or validation fails, report the failure and diagnostic path; do not substitute example sequences or repeatedly resubmit the same request.

Examples

Run from this skill's directory, or use an absolute path to scripts/design.py. Substitute the user's input path and a new output directory:

python scripts/design.py --mode hosted \
  --pdb /path/to/backbone.pdb --num-sequences 10 \
  --temperature 0.1 --output-dir /path/to/new-design-run

For a running local NIM, use --mode local; the client honors PROTEINMPNN_NIM_URL. To design only chain A, exclude cysteine, or request the soluble model, add --chains A, --omit-aas C, or --soluble respectively. --seed sets random_seed; --ca-only selects the CA-only model. The helper uses one temperature per request; run separate output directories for a temperature sweep. For advanced JSONL controls or a custom batch request, use references/api.md and the post-response example in references/examples.md.

Save And Report Output

The client writes request.json, response.raw, response.json, designed_sequences.fa, and summary.json into the requested output directory. The FASTA preserves the complete returned mfasta, including a native/WT entry when present. The summary contains only designed sequences, each paired with its actual score, and records whether scores came from the JSON array or FASTA headers. It is also printed after the artifacts are saved and checked.

In the final response, report:

• The number of designed sequences, excluding the native/WT reference.
• Each design's identifier and actual returned score, plus its sequence (for long sequences, give a clearly labelled preview and link to the full FASTA).
• The saved FASTA and summary paths, and the raw response path for provenance.
• That these are inverse-folding candidates, with no fold-back validation performed unless it was actually requested and run.

Do not treat a score as proof that a sequence folds or binds. Further validation with Boltz2 or OpenFold3 is an optional next step. For FASTA/score sanity checks, read references/validation.md.

Limits And Troubleshooting

• Minimum GPU VRAM: about 3 GB.
• sampling_temp must be a list, even for one value.
• Empty mfasta: check non-empty input_pdb and num_seq_per_target >= 1.
• PDB parse errors: use valid PDB ATOM records.
• Local URL 404 usually means an accidental /v1/ prefix.
• Cache mount error: use /home/nvs/.cache/nim inside the container.