Protein backbone diffusion models consistently undersample catalytically important motifs, and oversample idealized helices, raising questions about the appropriateness of these methods in designing starting points for enzyme design & evolution. From www.biorxiv.org/content/10.1... - ThreadSky

delalamo.xyz • 108 days ago

Protein backbone diffusion models consistently undersample catalytically important motifs, and oversample idealized helices, raising questions about the appropriateness of these methods in designing starting points for enzyme design & evolution. From https://www.biorxiv.org/content/10.1101/2025.01.09.632260v1

Comments

lindorfflarsen.bsky.social•108 days ago

Does this mean that protein structures from genAI are as bland as text or images generated from genAI

kmichie.bsky.social•108 days ago

Is this just due to a bias in the pdb training data of many highly defined helices and much more variance in beta sheets and turns? I expect curating training data would alter the output to be more diverse? 🧬🧶

martinpacesa.bsky.social•107 days ago

I think actually it’s a consequence of the training goals. Machine learning creates the a function that satisfies your loss function in the easiest way. Nature does. Natures goes through (seemingly) unnecessary hoops and loops that results in diversity.

delalamo.xyz•107 days ago

I think that’s it. Secondary structure is mostly risk-free from a design perspective. Loop conformations, particularly long loops, are probably huge sources of error during training

proteinator.bsky.social•107 days ago

Ah yeah, didn’t fit into the previous post: there’s no bias in sec structure distro in the PDB. Also in the paper I attached above we have a plot showing the mean of PDB regards to hels and strands and of different gen models 😅

proteinator.bsky.social•107 days ago

Gen models just learn the locality of the helices very quickly and “cheat” on the metrics by overrepresenting helices. Especially for bigger proteins > 500 aas it really becomes apparent..

proteinator.bsky.social•107 days ago

Not sure it’s that easy. We recently proposed a checkpointing selection criterion (https://arxiv.org/abs/2411.05238) to match better the distribution of secondary structure elements of native proteins from PDB but it doesn’t seem to work too good.

martinpacesa.bsky.social•107 days ago

You call this bland :D
https://bsky.app/profile/martinpacesa.bsky.social/post/3lcrq72kfqk2t

delalamo.xyz•107 days ago

All these beta-sandwiches are making me hungry

lindorfflarsen.bsky.social•107 days ago

Hah. No. What I mean is that perhaps the structures are a bit more boring and idealized compared to some of the more quirky things we see in nature (which I guess is also why single sequence predictions work better for designed proteins)

delalamo.xyz•107 days ago

Bland is a funny but entirely appropriate way to describe them IMO

proteinator.bsky.social•108 days ago

the consequence of that is pretty well known - rock-stable rigid proteins 🤠

Comments

Posting Rules

Reply