tattabio/gLM2_650M_bgc_decoder

gLM2 650M BGC Decoder

A 650M-parameter masked discrete diffusion model for conditional generation and redesign of biosynthetic gene clusters. The model supports antiSMASH domain conditioning and sequences up to 16,384 tokens.

import torch
from transformers import AutoModelForMaskedLM, AutoTokenizer

model_id = "tattabio/gLM2_650M_bgc_decoder"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForMaskedLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
).cuda()

Inference code: https://github.com/TattaBio/gLM2_decoder

Citation

@article {Lanclos2026.09.11.750945,
    author = {Lanclos, Nathan and Ibrahim, Kyrellos and Cornman, Andre and Huang, Marco and Gill, Vikram and Jain, Aalini and Abraham, Jonathan and Gin, Jennifer and Chen, Yan and Petzold, Christopher and Baerwald, Justin and Kortemme, Tanja and Keasling, Jay and Hwang, Yunha},
    title = {Generative Design of New-to-nature Biosynthetic Assembly Lines with Genomic Language Modeling},
    elocation-id = {2026.09.11.750945},
    year = {2026},
    doi = {10.64898/2026.09.11.750945},
    publisher = {Cold Spring Harbor Laboratory},
    URL = {https://www.biorxiv.org/content/early/2026/09/21/2026.09.11.750945},
    eprint = {https://www.biorxiv.org/content/early/2026/09/21/2026.09.11.750945.full.pdf},
    journal = {bioRxiv}
}

License

Model weights are released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. The weights are freely available for academic and research purposes.