Add ZAO 4D molecular foundation model skills (zao-embed, zao-predict)

[kokishimada]

Add ZAO — 4D molecular foundation model skills

This PR adds agent skills for ZAO, a 4D molecular foundation model from SyntheticGestalt. ZAO converts SMILES into 2048-dimensional molecular embeddings by processing multiple 3D conformers per molecule, capturing spatial arrangement and conformational flexibility that 1D/2D representations miss. The embeddings are drop-in features for downstream property/activity prediction; the ZAO embedding + CatBoost recipe ranks #1 on several TDC ADMET Benchmark Group datasets.

Skills

• zao-embed — send SMILES to a subscribed ZAO endpoint and return 2048-dim embeddings (.npz/.npy/.csv/.jsonl).
• zao-predict — embed a labeled set via the endpoint, train a downstream CatBoost model on the user's labels, and predict on query molecules.

Deployment model

ZAO is distributed as a managed marketplace endpoint: the user subscribes to the ZAO model, deploys an endpoint in their own cloud account, and the skill calls that endpoint. All preprocessing (SMILES standardization, GPU conformer generation, feature extraction, model forward pass) runs inside the endpoint container — the caller only sends SMILES and receives embeddings. Two modes share an identical request/response contract:

• AWS SageMaker (default) — a subscribed ZAO SageMaker Model Package, invoked via sagemaker-runtime:InvokeEndpoint (AWS SigV4 auth).
• Local container — the same container run locally over HTTP (ZAO_ENDPOINT_URL).

This is a new integration shape for the toolkit: a customer-subscribed marketplace endpoint, distinct from the NVIDIA-hosted nim-skills/ and the locally-built open-models-skills/. We placed it under open-models-skills/zao/ as the closest fit and would welcome guidance on the preferred category (or an MCP-wrapper form) — see the open question below.

Contents

open-models-skills/zao/
├── README.md
├── config/defaults_predict.json
├── scripts/{invoke_endpoint,train_catboost,predict_catboost}.py
├── skills/{zao-embed,zao-predict}/SKILL.md
└── tests/test_skill_frontmatter.py

Also registered in skills.sh.json (Open Models), added to the root README skill catalog, and mirrored into plugins/bionemo-agent-toolkit/skills/.

Testing

• tests/test_skill_frontmatter.py (14 checks) passes: both skills conform to the agentskills.io frontmatter spec and the project's metadata requirements, and stay within the 500-line / 5000-token budget.
• End-to-end functional test with the real ZAO model: zao-embed → zao-predict trained a downstream regressor on 118 real drug compounds and predicted held-out molecules, beating the mean baseline. Per-molecule error handling (out-of-range MW, unparseable SMILES) returns null embeddings index-aligned and does not fail the batch.

Note on generated artifacts

.skillsource.json (per-skill) is produced by the maintainers' build_repo/autogenerate.py tooling, which is not in this public repo, so it is omitted here — please regenerate it (and any derived manifests) as part of merge.

Open question for maintainers

Where should a "customer-subscribed marketplace endpoint" skill live, given it is neither an NVIDIA-hosted NIM nor a locally-built open model? Happy to move it or wrap it as an MCP server if that is the preferred pattern.