protlms

Unified toolkit for inference across a variety of protein language models (pLMs).

protlms is a lightweight client/CLI that gives you one interface for embeddings and likelihoods regardless of the underlying model. The client itself carries no ML dependencies — each model ships as a standalone Docker image, and the client talks to images through a standardized container contract. See docs/VISION.md for the full design.

Installation

pip install -e ".[dev]"

Quick start

Build a model image (the tiny ESM2 demo model), then run inference through the client:

# build the demo image (weights baked in)
docker build --build-arg ESM2_CHECKPOINT=esm2_t6_8M -t protlms-esm2:t6_8M containers/esm2

protlms models list                                          # available models
protlms embed      esm2-8m seqs.fasta -o out/ --pooling mean # pooled embeddings (.npz)
protlms embed      esm2-8m seqs.fasta -o out/ --pooling none # per-residue embeddings (.npy)
protlms likelihood esm2-8m seqs.fasta -o out/                # pseudo-log-likelihoods (.csv)
protlms embed      esm2-8m seqs.fasta -o out/ --gpu          # run on GPU

import protlms

model = protlms.load("esm2-8m")
emb = model.embed("seqs.fasta", pooling="mean")
print(emb.pooled())               # {record_id: (embedding_dim,) array}

ll = model.likelihood("seqs.fasta")
print(ll.rows())                  # per-sequence likelihood / perplexity

Project layout

Development

pytest                       # run the test suite
ruff check src/ tests/       # lint
ruff format src/ tests/      # format
ty check src/                # type check

All tooling is configured in pyproject.toml — there are no separate config files.

License

MIT — see LICENSE.