SE(3)-Stochastic Flow Matching for Protein Backbone Generation

Explore a comprehensive lecture on SE(3)-Stochastic Flow Matching for Protein Backbone Generation presented by Tara Akhound-Sadegh from Valence Labs. Delve into the world of computational protein design and discover FoldFlow, a series of novel generative models based on flow-matching paradigms over 3D rigid motions. Learn about the advantages of FoldFlow models, including their stability, faster training compared to diffusion-based approaches, and ability to map invariant source distributions to invariant target distributions over SE(3). Examine the progression from FoldFlow-Base to FoldFlow-OT and FoldFlow-SFM, understanding how each iteration improves upon the previous model. Gain insights into protein backbone generation for structures up to 300 amino acids, and see how these models produce high-quality, designable, diverse, and novel samples. The lecture covers key topics such as representing protein backbones, Riemannian flow matching, and experimental results, concluding with a Q&A session to address audience inquiries.

- Intro
- Representing Protein Backbones
- Riemannian Flow Matching
- FoldFlow
- FoldFlow-OT
- FoldFlow-SFM
- Experimental Results
- Q&A