SE(3)-MeanFlow: Few-Step Protein Backbone Generation on Lie Groups
A novel generative framework named SE(3)-MeanFlow has been developed by researchers for the generation of protein backbones, functioning directly on the Lie group SE(3). This technique adapts MeanFlow from Euclidean settings to the geometric context of protein structures, utilizing the Lie algebra so(3) and R^3. By deriving closed-form identities for average velocities related to rotations and translations, this method eliminates the necessity for simulation-based training targets, thus bypassing the complex numerical ODE integrations typically required over numerous network evaluations. Furthermore, an SE(3) alpha-Flow objective is proposed to facilitate a warm-up phase by removing the Jacobian-vector product from the rotation component, transitioning to a stabilized MeanFlow objective thereafter. This advancement aims to alleviate computational challenges in high-throughput design efforts, enhancing the efficiency of de novo protein design. The research can be found on arXiv with the identifier 2607.27431.
Key facts
- SE(3)-MeanFlow is a few-step generative framework for protein backbone generation.
- It extends MeanFlow from Euclidean space to Lie group geometry of protein frames.
- The method works natively in the Lie algebra so(3) and R^3.
- Closed-form average-velocity identities for rotations and translations are derived.
- The approach provides simulation-free training targets.
- An SE(3) alpha-Flow objective removes the Jacobian-vector product from the rotation branch.
- Training switches to a small-t stabilized MeanFlow objective after warm-up.
- The paper is available on arXiv with identifier 2607.27431.
Entities
Institutions
- arXiv