biolm-sdk

Call biological language models from Python or the terminal.

Encode protein sequences, predict structures, generate variants, score antibodies, run DNA models — through one client against BioLM's hosted API or a local biolm-hub gateway.

from biolm import Model

# Predict a structure
result = Model("esmfold").predict(type="sequence", items="MKTAYIAKQRQGHQAMAEIKQ")
print(result["mean_plddt"])

# Embed a library
embeddings = Model("esm2-8m").encode(
    type="sequence",
    items=["MKTAYIAKQRQ", "MKLAVIDSAQRQ", "MENDELMENDEL"],
)

Install: pip install biolm-sdk · Import: import biolm · CLI: biolm

Setup

Python 3.8+

pip install biolm-sdk

Credentials — get a token at biolm.ai, then:

export BIOLM_TOKEN=<token>
# or
biolm account login

Check everything is wired up:

biolm status

Run models

From Python

Bind a model, pass sequences (or PDBs, or other typed inputs), get results back.

from biolm import Model

# Embeddings
esm = Model("esm2-8m")
vecs = esm.encode(type="sequence", items=["MKTAYIAKQRQ", "MDNELE"])

# Generation
progen = Model("progen2-oas")
seqs = progen.generate(
    type="context",
    items="M",
    params={"temperature": 0.7, "num_samples": 5, "max_length": 50},
)

# Structure from sequence
fold = Model("esmfold")
pdb = fold.predict(type="sequence", items="MKTAYIAKQRQ")

Load inputs from disk:

from biolm.io import load_fasta

sequences = load_fasta("library.fasta")
Model("esm2-8m").encode(type="sequence", items=sequences)

Large jobs can stream to JSONL instead of memory:

Model("esmfold").predict(
    type="sequence",
    items=sequences,
    output="disk",
    file_path="structures.jsonl",
)

From the terminal

biolm model list
biolm model show esmfold
biolm model run esmfold predict -i sequences.fasta -o results.json
biolm model example esm2-8m encode   # prints a Python snippet you can paste

The CLI accepts FASTA, CSV, PDB, and JSON. It talks to the same API the SDK does.

Build workflows

For jobs that are more than a single model call.

Protocols

Multi-step jobs defined in YAML — validate locally, submit to the platform, poll until done.

biolm protocol validate design.yaml
biolm protocol list --search design
biolm protocol run my-protocol-slug -i inputs.json --wait

from biolm import run_protocol

results = run_protocol(
    "my-protocol-slug",
    inputs={"sequence": "MKTAYIAKQRQ"},
)

Pipelines

For protein design at scale: generate variants, score them, filter, cluster — with DuckDB caching so re-runs skip work already done.

pip install "biolm-sdk[pipeline]"

Saturation mutagenesis — enumerate single mutants, score, keep the top N:

from biolm.pipeline import GenerativePipeline, SaturationMutagenesisConfig

pipeline = GenerativePipeline(configs=[
    SaturationMutagenesisConfig(
        parent_sequence="MKTAYIAKQRQ",
        scoring_model="esm2-650m",
        score_field="logits",
        top_n=20,
    )
])
df = pipeline.run()

Custom stages — predict → filter → rank, composed explicitly:

from biolm.pipeline import DataPipeline
from biolm.pipeline.filters import ThresholdFilter, RankingFilter

pipeline = DataPipeline(sequences=my_sequences)
# The model slug is intentionally spelled "temberture-regression" in the API.
pipeline.add_prediction("temberture-regression", extractions="prediction", columns="tm")
pipeline.add_filter(ThresholdFilter("tm", min_value=48.0))
pipeline.add_filter(RankingFilter("tm", top_n=10))
df = pipeline.run()

Or use the shorthand:

from biolm.pipeline import Predict

df = Predict("temberture-regression", sequences=my_sequences, extractions="prediction", columns="tm")

See scripts/ in this repo for antibody design, stability engineering, and multi-model examples.

Run models locally with biolm-hub

Point the SDK at a biolm-hub gateway to run open-source models on your own hardware:

bh serve                              # in the biolm-hub repo
biolm hub set http://127.0.0.1:8000   # redirect SDK + CLI
biolm model list                      # discovers models from hub OpenAPI

biolm hub unset returns to the hosted API.

How it works

You write synchronous Python. Under the hood the client is async: it reads each model's schema to pick batch sizes, sends batches in parallel (up to 16 concurrent by default), rate-limits to the API's throttle, retries transient network errors, and gzip-compresses large payloads.

Generators work as items — the client consumes them batch-by-batch without loading everything into memory.

What's in the box

Models include ESM2, ESMFold, ESM-1v, ProteinMPNN, ProGen2, AntiFold, IgBERT, DNABERT2, ABodyBuilder3, and more. Browse with biolm model list or at biolm.ai.

Documentation

Full guides, API reference, and tutorials: biolm.ai/docs

Development

git clone git@github.com:BioLM/biolm-sdk.git && cd biolm-sdk
pip install -r requirements_dev.txt
make install
RS=118 make test

See CONTRIBUTING.rst.

License

Apache 2.0

_{Previously published as biolmai. Migration: biolm.ai/docs/notes/migration-1.0.}