GeoBPE 0.1.0

Protein Structure Tokenization via Geometric Byte Pair Encoding (GeoBPE)

Protein Geometric Byte Pair Encoding

This repo contains our implementation of Protein Structure Tokenization via Geometric Byte Pair Encoding (ICLR 2026).

Overview

Protein structure is central to biological function, and enabling multimodal protein models requires joint reasoning over sequence, structure, and function. We introduce GeoBPE, a geometry-grounded PST that transforms continuous, noisy, multi-scale backbone conformations into discrete ``sentences" of geometry while enforcing global constraints.

Analogous to byte-pair encoding, GeoBPE generates a hierarchical vocabulary of geometric primitives by iteratively (i) clustering Geo-Pair occurrences with k-medoids to yield a resolution-controllable vocabulary; (ii) quantizing each Geo-Pair to its closest medoid prototype; and (iii) reducing drift through differentiable inverse kinematics that optimizes boundary glue angles under an SE(3) end-frame loss.

Run GeoBPE

GeoBPE supports two sub-commands: encode and induce. Run geobpe --help for a description.

Run geobpe encode --help and geobpe induce --help to see detailed arguments.

We include the following resources to make it easy to use GeoBPE:

• GeoBPE API and Usage Guidelines Doc -- descriptions, intuitions, and guidelines on how to effectively and efficiently use GeoBPE
• Experiment Logs -- collection of past experiments varying hyperparameters settings; quickly lookup settings & performance to save future iteration time.

Citation

If you use GeoBPE in your research, please cite our paper:

@inproceedings{sun2025protein,
  title={Protein Structure Tokenization via Geometric Byte Pair Encoding},
  author={Sun, Michael and Yuan, Weize and Liu, Gang and Matusik, Wojciech and Zitnik, Marinka},
  booktitle={International Conference on Learning Representations},
  year={2026},
  url={https://arxiv.org/abs/2511.11758}
}

Contact

Please contact [e-mail] if you have any questions.