LMR-Long (SHIFU / LMR foundational backbone)

LMR-Long is a foundational RNA language-model backbone (290M parameters) from SHIFU: an integrated framework for deep learning of RNA secondary structure (Galvez and Vicens, Vicens Lab, University of Houston).

It is pretrained by masked language modeling on RNA sequence and serves as the backbone for downstream RNA structure tasks. The three released backbones are:

Model	Params	Context	Repo
LMR-v0	289M	512	GaboG7/LMR-v0
LMR-G	228M	512	GaboG7/LMR-G
LMR-nano	65M	512	GaboG7/LMR-mini
LMR-Long	290M	4,096	GaboG7/LMR-Long

Usage

from huggingface_hub import snapshot_download
path = snapshot_download("GaboG7/LMR-Long")   # LMR-v0 / LMR-G / LMR-mini
# load with the LMR code: https://github.com/Vicens-Lab/LMR

Training code, configs, and the finetuning recipe are in the GitHub repo: https://github.com/Vicens-Lab/LMR

Training data

Pretrained on RNA sequence data; the downstream 2D structure benchmark (SHIFU-Corpus, 254,123 sequences, leakage-audited, family-aware splits) is a separate release.

Intended use and limitations

A general-purpose RNA sequence backbone for finetuning (e.g. secondary-structure prediction). It is a sequence model, not a standalone structure predictor. Inputs cap at the context length in the config.

License

MIT (code and weights). SHIFU-Corpus retains the licenses of its six source databases.

Citation

This work is not yet published; a citation will be added when available.

@misc{lmr_shifu,
  title  = {SHIFU: an integrated framework for deep learning of RNA secondary structure},
  author = {Galvez, Gabriel and Vicens, Quentin},
  note   = {Manuscript in preparation. Citation to be updated.},
  year   = {TODO}
}