biolm-hub¶

A standardized, agent-first catalog of open biological ML models that deploy on Modal in a couple of commands.

Running an open biological ML model shouldn't mean re-solving the same plumbing every time — dependencies, an undocumented interface, a one-off deployment — before you get a single prediction, for every model and every person or agent who tries. What's still missing is a clean, uniform, documented, deploy-anywhere substrate so nobody has to reinvent the wheel. That's what this repo is: ready-to-run bio-models anyone — human or agent — can pull off the shelf and run, growing as the community adds more.

📦 Source: this site documents the open-source biolm-hub repository on GitHub — browse the code, file an issue, or contribute a model there. (The repo link is also in the top bar of every page.)

Five-minute success¶

git clone https://github.com/BioLM/biolm-hub
cd biolm-hub
make install                               # venv + the `bh` CLI (all extras)

bh setup                                   # checks your Modal config
bh deploy esm2                             # deploy to your Modal workspace

bh serve                                   # local catalog UI + HTTP API → http://127.0.0.1:8000
# then, in another terminal, call it:
curl -X POST http://127.0.0.1:8000/api/v1/esm2-8m/encode \
  -H 'Content-Type: application/json' \
  -d '{"items": [{"sequence": "MKTAYIAKQR"}]}'

No R2 / Hugging Face secrets on your Modal workspace? Prefix deploys with BIOLM_SKIP_MODAL_SECRETS=1 so the build reads public weights anonymously. See the HTTP API for the full calling contract.

What's inside¶

models/ — each model with a uniform layout (app.py, config.py, schema.py, test.py) plus a machine-readable knowledge graph (sources.yaml, comparison.yaml, README.md, MODEL.md, BIOLOGY.md): when to use it, training data, benchmarks, license, alternatives.
cli/ — the bh tool: setup, deploy, serve, r2.
gateway/ — a unified inference endpoint + a catalog web app.
Uniform action verbs (predict, fold, encode, generate, score, log_prob), uniform schemas, structured logging, and a consistent error taxonomy — so an agent that learns one model knows them all.

Start here¶

Quickstart — clone to a running model in a few commands.
Model catalog — every model, with API schema, when-to-use guidance, and license.
Philosophy — the design center.
Contributing — add a model and the house rules.
For agents — the machine-readable API is at /openapi.json (via bh serve or a deployed gateway); each model's comparison.yaml / sources.yaml drives model selection.