Nugkta/bgcbench-lcplm

LC-PLM (BGC-Bench mirror)

This repository redistributes the LC-PLM weights released by the original authors, so that BGC-Bench can download them at a fixed tag. It is not an official release of the original authors, and the weights are not retrained. Licence: CC BY-NC 4.0, as upstream; non-commercial use only. Please cite the original paper (below). Intended use: frozen feature extraction in BGC-Bench; see the original paper for the model's training data and limitations.

Upstream

• Paper: Wang, Wang, Sadeh, Zancato, Achille, Karypis & Rangwala (2025). LC-PLM: long-context protein language modeling using bidirectional Mamba with shared projection layers. Transactions on Machine Learning Research. https://openreview.net/forum?id=dWvztQzfy4 (arXiv:2411.08909)
• Code: https://github.com/amazon-science/LC-PLM
• Weights: GitHub repository amazon-science/LC-PLM, commit 80388cad25a8a4f0c1d1f0922fb8a6712e182cb4 (CC BY-NC 4.0; weights in Git LFS)

Format

Files copied byte for byte from that commit (config.json, configuration_lcplm.py, modeling_lcplm.py, layer_norm.py, model.safetensors, LICENSE). layer_norm.py is under Apache-2.0 (Tri Dao, modified by Amazon; its header is kept); the weights and the other files are under CC BY-NC 4.0.

Model

Bidirectional Mamba (BiMamba-S, 48 layers, 1536-d) protein language model; tokenizer: facebook/esm2_t6_8M_UR50D. Load with trust_remote_code=True. BGC-Bench ties the reverse-direction projections after loading, embeds one sequence at a time without padding, and uses fused_add_norm=True. BGC-Bench averages the last layer over residues per CDS, then over CDS (1536-d).

Files

file	bytes
LICENSE	19,346
config.json	773
configuration_lcplm.py	2,147
layer_norm.py	35,139
model.safetensors	3,011,271,376
modeling_lcplm.py	24,759

Use

git clone https://github.com/Nugkta/BGC-Bench && cd BGC-Bench
uv sync
cd envs/gpu && uv sync --frozen
CUDA_VISIBLE_DEVICES=0 .venv/bin/bgcbench embed --encoder lcplm
cd ../.. && uv run bgcbench run --encoder lcplm

Citation

@article{wang2025lcplm,
  title   = {{LC-PLM}: Long-Context Protein Language Modeling Using Bidirectional {Mamba} with Shared Projection Layers},
  author  = {Wang, Yingheng and Wang, Zichen and Sadeh, Gil and Zancato, Luca and Achille, Alessandro and Karypis, George and Rangwala, Huzefa},
  journal = {Transactions on Machine Learning Research}, year = {2025},
  url     = {https://openreview.net/forum?id=dWvztQzfy4}
}