You can deploy this model and start earning money today!

danyalrehman17/
robin-transferable

Name: robin-transferable
Availability: InStock

pytorch

2606.27361

transferable-samplers/many-peptides-md

mit

Views

No views yet

🧬 Autoregressive Boltzmann Generators

Paper

Code

Data

Robin generating equilibrium conformations of unseen peptides

_{🧪 Robin sampling equilibrium conformations of unseen peptides, zero-shot.}

Robin is the 132M-parameter transferable Autoregressive Boltzmann Generator (ArBG) from Autoregressive Boltzmann Generators (ICML 2026 🏆 Spotlight). It's a GPT-style causal transformer that samples equilibrium peptide conformations zero-shot, generalizing to unseen peptides and reducing energy error (E-W2) on 8-residue systems by over 60% vs. the prior state-of-the-art. Code, training, and evaluation live in the official repo.

🚀 Usage

The transferable eval script auto-downloads this checkpoint and evaluates all 90 held-out test peptides:

bash

git clone https://github.com/danyalrehman/AutoBG.git
cd AutoBG && uv sync
sbatch scripts/eval_transferable.sh   # downloads robin.ckpt automatically

To fetch the checkpoint manually:

hf download danyalrehman17/robin-transferable robin.ckpt --local-dir checkpoints

📚 Citation

bibtex

@inproceedings{rehman2026autoregressive,
  title     = {Autoregressive Boltzmann Generators},
  author    = {Rehman, Danyal and Tan, Charlie B. and Bengio, Yoshua and Bose, Avishek Joey and Tong, Alexander},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
  series    = {Proceedings of Machine Learning Research},
  volume    = {306},
  year      = {2026},
  publisher = {PMLR},
  note      = {Spotlight},
  url       = {https://arxiv.org/abs/2606.27361}
}

🙏 Acknowledgements

We thank Hugging Face for hosting the ManyPeptidesMD dataset and the released Robin weights.

📜 License

MIT. The accompanying codebase adapts third-party code under other licenses (Apple, Meta, NVIDIA, Klein & Noé), some non-commercial — see the NOTICE.