LULA-1 is now available as an open-weight release on Hugging Face. It is a sequence-only protein-ligand scorer: protein amino-acid sequence plus ligand SMILES in, binding score out. No structure input, no docking, and no folding step.
The release is intentionally small. Om publishes the LULA-1 scoring head, verification files, examples, a Colab quickstart, and license terms. The model runs locally through the omtx Python package, using public ESM-2 and ChemBERTa encoders at inference time.
Why open weights
Drug discovery teams often need to rank very large molecule sets before they know which structures are worth deeper modeling. LULA-1 is built for that first-pass ranking problem. The protein is embedded once, and each additional molecule is scored from its ligand representation against the target context.
That shape matters. It makes broad screens practical when a target has no solved structure, when docking is too slow for the first pass, or when a team wants a private local baseline before moving into hosted workflows.
The metric that matters for screening
First-pass screening is a ranking problem: known binders should rise to the top of massive candidate lists, where a scientist can act on them.
That is why the Hugging Face card emphasizes enrichment. In a proteome-scale sweep, LULA-1 scored 10,575,246,882 protein-ligand pairs across 10,569 proteins. Across the top 1,000 ranked molecules per target, it found 267,324 known binders versus 4,877.6 expected under random ranking, an aggregate EF@1000 of 54.8x.
Full secondary validation details remain on the model card for auditability, but the product-facing readout is enrichment: does LULA move likely binders into the reviewable top of a massive candidate list?
Fast, inexpensive triage
LULA is priced and designed for iteration. Hosted LULA-1 scoring starts at 10 Wallet Credits per 1,000 molecules scored, and hosted LULA-2 scoring starts at 50 Wallet Credits per 1,000 molecules scored, with a one Wallet Credit positive-usage minimum. Against public small-molecule design and screen API pricing, LULA-1 is 250x lower and LULA-2 is 50x lower for Om-hosted molecule scoring.
The throughput story is equally important. The Nesso-1 technical report states that Nesso-1 requires roughly one second per prediction on a single GPU. Om positions LULA-2 hosted scoring at about 200 molecules per second for direct molecule ranking. That is the difference between a slow structural pass and a tight score-order-test loop.
How to try it
Install the public SDK and download the release assets:
pip install "omtx[lula]>=2.0.12"
omtx lula download
omtx lula verifyThen load the model locally and score a protein sequence against one SMILES string or a batch of SMILES strings. Batch outputs include score, rank, and top_percentile_in_batch, so the result can be used directly to prioritize a candidate set.
License path
The open-weight release is free for non-commercial research, evaluation, benchmarking, teaching, and publication under the Om LULA Community License 1.2. You can use local LULA scores to choose molecules to order through Om, then use experimental data from Om-purchased molecules commercially for your own program. Commercial local deployment, self-hosting, product integration, resale, or use outside the Om fulfillment path requires a written Om commercial license.
Commercial teams have two paths. Use Om-hosted scoring and fulfillment through paid hosted usage, or contact Om for a private commercial model license when targets, compound libraries, fine-tuning, or deployment need to stay inside your environment.
Open LULA-1 on Hugging Face, review hosted pricing, or start a private licensing inquiry.
