Reproduce “Generating Novel, Designable, and Diverse Protein Structures by Equivariantly Diffusing Oriented Residue Clouds” on OpenResearch

About this paper

Yeqing Lin, Mohammed AlQuraishi

Researchers at Columbia University developed Genie, a generative diffusion model that produces novel, designable, and diverse protein structures by combining a simple C α coordinate diffusion process with an SE(3)-equivariant denoiser operating on oriented residue frames. This dual representation strategy yielded 58.3% confidently designable structures for short proteins, a substantial improvement over prior methods, and demonstrated competitive performance in novelty and diversity against state-of-the-art models for longer proteins.

Key insights

• Combining simple coordinate diffusion with a sophisticated frame-based SE(3)-equivariant denoiser effectively addresses the trade-off between noise application robustness and precise geometric modeling.
• Explicitly incorporating reflection sensitivity through Invariant Point Attention on oriented residue frames is crucial for generating physically correct protein structures and avoiding non-physical mirror images.
• The 'geometric asymmetry' in the dual representation allows for both ease of noise addition (on C α coordinates) and high-fidelity geometric reasoning (via Frenet-Serret frames) during the generation process.

Summary by alphaXiv.