lula-1.1 / README.md

Publish LULA-1.1 open-weight release

34d20a4 verified about 23 hours ago

9.54 kB

license: other
license_name: om-lula-community-license-1.1
license_link: LICENSE
pipeline_tag: other
tags:
  - biology
  - drug-discovery
  - small-molecule-discovery
  - protein-ligand
  - binding-affinity
  - open-weight
  - lula-1
  - lula-1.1

License notice. By downloading, accessing, or using LULA-1.1, you agree to the Om LULA Community License 1.1. LULA-1.1 is licensed for research, evaluation, benchmarking, teaching, and other non-commercial research uses only. Any commercial work requires a separate Om commercial license, including internal commercial discovery, commercial drug discovery, screening, hit finding, lead optimization, portfolio decisions, production R&D, product candidate identification, patent or therapeutic program work, hosted inference, paid API/SaaS access, resale, support/deployment, product bundling, and competing model services. For commercial licensing, contact dmc@omtx.ai.

Protein scaffold with a ligand bound in a highlighted pocket

+ + +

omtx.ai

Open-weight release track

LULA-1.1 sequence-only protein-ligand scoring.

Protein amino-acid sequence plus ligand SMILES in, binding score out. No structure input, no docking, no folding step.

Sequence-only Local inference Open weights

Model at a glance

1.7M parameters 6.8 MB 6.14M training pairs 13,368 proteins Sequence-only Local inference Open weights

LULA-1.1

LULA-1.1 is a lightweight, sequence-only protein-ligand binding scorer from Om Therapeutics. It takes a protein amino-acid sequence and ligand SMILES and returns a binding score. There is no structure input, docking, or folding step.

This release uses the same ConPLex-style two-tower scoring architecture as the original LULA-1 open-weight release, with an updated target-balanced training recipe and expanded training coverage. The customer-facing model name is LULA-1.1.

What Changed From LULA-1

LULA-1.1 keeps the LULA-1 two-tower architecture while updating the weights, training coverage, sampling, and protein-context handling.

Compared with LULA-1, LULA-1.1 increases supervised protein-ligand training coverage from 2,763,260 to 6,137,835 pairs, adding 3,374,575 protein-ligand training pairs.

Coverage	LULA-1	LULA-1.1
Supervised protein-ligand training pairs	2,763,260	6,137,835
Binder-labeled training pairs	2,132,861	4,816,392
Non-binder-labeled training pairs	630,399	1,321,443

The updated sampling recipe is target-balanced to avoid letting high-row-count targets dominate the update stream. LULA-1.1 also uses complete protein-context inference: 1,022-residue ESM windows with 256-residue overlap, C-terminal coverage, overlap-averaged residues, and full-sequence mean pooling excluding BOS/EOS tokens.

The validation evidence for this release is mixed across panels. LULA-1.1 is published as the next open-weight release for research and evaluation; users should benchmark it against their own targets before relying on rank ordering.

Protein Coverage

LULA-1.1 represents 13,368 protein source entities across model-ready release inputs.

Example proteins represented include EGFR, JAK2, RET, CDK2, MAPK1, GSK3B, DRD2, OPRM1, CHRM2, HTR2A, ESR1, AR, PPARG, BACE1, and thrombin.

Example protein classes include kinases, GPCRs, nuclear receptors, proteases/peptidases, ion channels and transporters, phosphatases, epigenetic/chromatin regulators, immune/complement/coagulation proteins, and cell-surface receptors.

This release reports aggregate coverage only. It does not include protein-level source manifests, amino-acid sequence tables, ligand rows, per-pair training rows, source object paths, internal private dataset/vintage identifiers, or customer data.

What This Release Contains

LULA-1.1 ships as a compact scoring head that runs on top of two public pretrained encoders. Om distributes the LULA-1.1 scoring head in this repository; the third-party encoders remain governed by their own upstream terms.

Component	Parameters	Source
LULA-1.1 scoring head	1,705,984	this repository
ESM-2 650M protein encoder	652,358,616	`facebook/esm2_t33_650M_UR50D`
ChemBERTa-77M-MTR ligand encoder	about 3,500,000	`DeepChem/ChemBERTa-77M-MTR`

Usage

pip install "omtx[lula]>=2.0.14"

hf auth login
omtx lula download --model lula1.1
omtx lula verify

Version Selection

Use the public model selector to choose the release:

omtx lula download --model lula1     # original LULA-1 open-weight release
omtx lula download --model lula1.1   # LULA-1.1

from omtx.lula import load_model

model = load_model("lula1")    # original LULA-1
model = load_model("lula1.1")  # LULA-1.1

from omtx.lula import load_model

model = load_model("lula1.1")
rows = model.score(
    protein_sequence="MSHHWGYGKHNGPEHWHKDFPIAKGERQSPVDIDTHTAKYDPSLKPLSVSYDQA",
    smiles=["CCO", "CC(=O)Nc1nnc(s1)S(N)(=O)=O"],
)
print(rows)

Batch scoring returns, per molecule: score, rank, and top_percentile_in_batch. Scores are intended for relative prioritization within a candidate set and are not calibrated binding probabilities.

Data Locality

Local scoring and fine-tuning run on your machine. Om does not receive your targets, compounds, labels, checkpoints, or scores when you use the local model.

Hosted Om scoring is a separate product surface.

Files

File	Bytes	SHA256
`model/best.pt`	6,826,303	`1933bbdf4aa335d52498c754dc961141d10631e38fb4cc1daa9a908e7e4601ba`
`model/model_config.json`	310	`de22bc8182d7f5f46454455e0e316296b59fb67f3aa420ce78c2728815806ff6`
`model/inference_config.json`	436	`115a343fef54f0d0a108998bb3ce86bd493e29a6518055e5707d896a1ddac36b`

License

LULA-1.1 is distributed under the Om LULA Community License 1.1. It is open-weight, not OSI open source. Commercial use requires a separate written Om commercial license. Contact dmc@omtx.ai for commercial licensing.

Attribution

LULA-1.1 uses a ConPLex-style scoring architecture. See the ConPLex reference implementation at https://github.com/samsledje/ConPLex and the publication DOI 10.1073/pnas.2220778120.