RigidSSL

Rigidity-Aware Geometric Pretraining for
Protein Design and Conformational Ensembles
ICLR 2026

¹University of Illinois Urbana-Champaign
²University of Washington
³MPI for Intelligent Systems, Tübingen
⁴National Research Council Canada
⁵University of Ottawa
⁶The Chinese University of Hong Kong

^*Equal contribution

1. RigidSSL use inertial frames to establish a canonical reference system for protein structures.

2. RigidSSL learns protein dynamics by maximizing mutual information between two conformations $g_0$ and $g_1$.

3. RigidSSL variants improve protein structure designability, novelty, diversity, and secondary structure statistics.

4. RigidSSL generalizes to long protein chains of 700-800 residues.

Citation

@inproceedings{
  ni2026rigidssl,
  title={Rigidity-Aware Geometric Pretraining for Protein Design and Conformational Ensembles},
  author={Zhanghan Ni and Yanjing Li and Zeju Qiu and Bernhard Sch{\"o}lkopf and Hongyu Guo and Weiyang Liu and Shengchao Liu},
  booktitle={International Conference on Learning Representations},
  year={2026},
  url={https://arxiv.org/abs/2603.02406}
}