TxGravity-30B-A6B-Open

Model Summary

TxGravity-30B-A6B-Open is a therapeutics-focused, task-specialized model built on the Gravity-30B-A5B base. It predicts a broad range of therapeutic properties — small-molecule ADMET, toxicity, drug–target interaction, protein–protein and peptide–MHC interaction, and more — by converting the Therapeutic Data Commons (TDC) benchmark tasks into an instruction format for LLMs, in the same spirit as Google's TxGemma.

TxGravity is co-developed by SK Biopharmaceuticals (SKBP) and Trillion Labs.

It is trained on TDC tasks reformatted as single-turn instructions (instruction → answer), covering 57 tasks. Answers are formatted as (A)/(B) for binary classification or a normalized 000–1000 bin for regression.

This release is the open-license version, with any datasets restricted to non-commercial use excluded from training.

⚠️ This is a task-specialized property predictor, not a general instruction-tuned or safety-aligned assistant. Its outputs are intended for the TDC-style therapeutic prediction prompts it was trained on. Predictions may be inaccurate, biased, or incomplete and must be independently verified before any experimental, clinical, or decision-making use.

Performance

Evaluated on the TDC therapeutic benchmark against a reproduced TxGemma-27B baseline. Over 60 comparable tasks: 16 wins / 26 ties / 18 losses (6 tasks excluded for lack of a paper-reported comparison). A win/loss requires the margin to exceed the tie threshold; otherwise the task is scored a tie.

Wins (16 tasks) — TxGravity beats TxGemma-27B on: bindingdb_ki, bindingdb_ic50, protein_sabdab, ppbr_az, caco2_wang, lipophilicity_astrazeneca, half_life_obach, clearance_hepatocyte_az, ld50_zhu, bioavailability_ma, buchwald_hartwig, drugcomb_css, drugcomb_loewe, drugcomb_zip, drugcomb_hsa, drugcomb_bliss — spanning binding affinity, ADMET, toxicity, and drug-combination synergy.

Highlights: bindingdb_ki PCC 0.750 vs -0.112, bindingdb_ic50 Spearman 0.766 vs 0.643, protein_sabdab MAE 0.969 vs 2.332, ld50_zhu MAE 0.651 vs 0.776.

Full results

Margin is TxGravity - TxGemma, in percentage points (%p) for score metrics or relative percent (%rel) for error metrics (MAE/MSE, lower is better). Up-arrow = higher is better, down-arrow = lower is better.

Also included in this repo as summary.csv.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("trillionlabs/TxGravity-30B-A6B-Open")
model = AutoModelForCausalLM.from_pretrained(
    "trillionlabs/TxGravity-30B-A6B-Open", dtype="bfloat16", device_map="auto")

messages = [{"role": "user", "content": "<TDC-formatted instruction here>"}]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=16)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

The model expects TDC-style prompts and answers in TDC format: (A)/(B) for classification, or a 000–1000 normalized bin for regression.

License

Apache 2.0

Citation

@misc{txgravity2026,
  title        = {TxGravity-30B-A6B-Open},
  author       = {SK Biopharmaceuticals and Trillion Labs},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/trillionlabs/TxGravity-30B-A6B-Open}}
}

This model builds on the Therapeutic Data Commons (TDC) benchmark:

@article{huang2021therapeutics,
  title   = {Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development},
  author  = {Huang, Kexin and Fu, Tianfan and Gao, Wenhao and Zhao, Yue and Roohani, Yusuf and Leskovec, Jure and Coley, Connor W and Xiao, Cao and Sun, Jimeng and Zitnik, Marinka},
  journal = {Proceedings of Neural Information Processing Systems, NeurIPS Datasets and Benchmarks},
  year    = {2021}
}