Kevin K. Yang 楊凱筌#996
Protein seq/structure diffusion-based codesign improves when the two design modes evolve independently, with sequence design starting only after structure is mostly decided. One of those things that we all already knew from experience, but now formally discovered algorithmically
Users appreciate GeoCoupling learning optimal temporal coupling for multimodal biomolecular design because it effectively denoises alpha carbon coordinates faster than latent variables, similar to the proteina complexa paper.
@DdelAlamo Does that count as “co-design” though? Same model I suppose
@DdelAlamo Nice. Also effectively in the proteina complexa paper tbf where the alpha carbon coords are denoised at a faster rate than the the latent variables
@DrJoshuaBox I was thinking this and DISCO where they flatten the distribution of AAs in earlier steps
Kevin K. Yang 楊凱筌#996Protein seq/structure diffusion-based codesign improves when the two design modes evolve independently, with sequence design starting only after structure is mostly decided. One of those things that we all already knew from experience, but now formally discovered algorithmically