RoSE: Enhancing SE(3)-based Protein Backbone Generation via Robust Score Estimation

Minzhang Li, Haochen Wang, Weichen Qin, Yifan Qin, Jiakai Zhang, Jiayi Dou, Jingyi Yu

Keywords: Riemannian Diffusion, Protein Design

TL;DR: Our work presents improvements to Riemannian diffusion models for protein structure generation by developing novel heat kernel computation methods on Riemannian symmetric spaces.

Abstract: This work presents improvements to Riemannian diffusion models for protein structure generation by developing robust heat kernel computation methods on $SE(3)$ space. While existing approaches suffer from approximation errors in score-based diffusion, our method enables stable and accurate denoising score matching on the high-dimensional $SE(3)^N$ manifold through theoretically-grounded numerical techniques. The proposed framework achieves competitive performance in protein generation benchmarks, demonstrating superior scores and successfully generating diverse, physically-plausible protein structures. Notably, our model solves 23 out of 24 motif scaffolding problems and designs refoldable nanobodies, significantly advancing the capability to generate functional protein geometries while maintaining mathematical consistency with the underlying manifold structure.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 14084