Since AF3 et al's diffusion model no longer models disordered proteins as spaghetti, it's now become surprisingly useful for generated novel protein backbones de novo. Here, the Chai1-predicted structure of a ChatGPT-generated sequence is redesigned w/ ProteinMPNN, & refolds with ~1.5 Å RMSD - ThreadSky

ddelalamo.bsky.social • 93 days ago

Since AF3 et al's diffusion model no longer models disordered proteins as spaghetti, it's now become surprisingly useful for generated novel protein backbones de novo. Here, the Chai1-predicted structure of a ChatGPT-generated sequence is redesigned w/ ProteinMPNN, & refolds with ~1.5 Å RMSD

1 / 3

Comments

elisafadda.bsky.social•93 days ago

That protein will be N-glycosylated if secreted, which may loosen local helicity, you're welcome 😁

klausenhauser.bsky.social•93 days ago

Now someone make it!

klausenhauser.bsky.social•93 days ago

However I’m quite surprised a random string gives a nice helical bundle even before MPNN

ddelalamo.bsky.social•93 days ago

It’s AF3, everything is a helical bundle!

ddelalamo.bsky.social•93 days ago

(This test was made possible by the Chai-1 web server and @simonduerr.bsky.social's proteinMPNN server here https://huggingface.co/spaces/simonduerr/ProteinMPNN)

kmichie.bsky.social•93 days ago

(Chai1 relaxed their licence didn’t they?) It’s not surprising- the generative transformer ‘wants’ to make something that looks like a real protein 🧬🧶

ddelalamo.bsky.social•92 days ago

Upon further experimentation, this approach is mostly garbage for larger proteins

rommieamaro.bsky.social•92 days ago

omg I love this follow-up. You saved us some effort!

ddelalamo.bsky.social•92 days ago

Yep that's the way it goes - get excited for a few small test cases in the morning, set up a larger test run, check the results six hours later and see a bunch of misfolded junk. On to the next one!

martinvoegele.bsky.social•93 days ago

Smart! 💡

daviddesancho.bsky.social•93 days ago

I wasn't aware of this
And may I add: is it true that we no longer get spaghetti from AF3?
Here is FUS (P35637), which still looks quite spaghetti-ish

ddelalamo.bsky.social•93 days ago

My impression is that they tried to train against spaghettis to recover this behavior but only partially succeeded in avoiding hallucination of secondary structure. Maybe this protein is among the training examples?

sokrypton.org•92 days ago

It depends on the length... they did extra work (by training on af2 predictions) to try get the model to return spaghetti, but once you go past their training limit (length of protein they trained on), it snaps back to returning helices for unknown regions.

daviddesancho.bsky.social•92 days ago

Thanks!
Well, that sounds like a really cumbersome way of getting spaghetti back

sokrypton.org•92 days ago

It is an issue of how diffusion models are trained. You train by adding / removing noise in ordered regions. It has no idea what to do w/ unordered regions its prior is to always be ordered.

For AF2, they included clash loss, so struct module prior is to return extended conformation when unknown.

ddelalamo.bsky.social•92 days ago

Ah didn't know the class loss is what drove that weird (but very useful) behavior!

daviddesancho.bsky.social•92 days ago

Great. That points to an easy way to improve these models (even if only as initial states for MD runs): use AGADIR (or some other secondary structure propensity estimator) as prior for the models. Would be better than having helices where you shouldn't, right?

daviddesancho.bsky.social•92 days ago

Also, with AF2, regions with helical structure that matched IDP-assigned stretches of sequence pointed to interesting stuff like conditional binding. So if everything is a helix, then we lose this "feature"

ajasjaljubetic.bsky.social•92 days ago

Is a clash loss not possible for diffusion models?

sokrypton.org•92 days ago

It's not clear how one would even go about implementing one... You need a resulting clashing structure to propagate the signal that clashes are bad. With how Diffusion models are trained, the final step is always the ground truth (no clash, all ordered).

proteinator.bsky.social•93 days ago

I'm not sure I'm following. How is it useful if seq -> str mapping is not guaranteed anymore? This puts the equality sign between random seq and pMPNN generated seq

ddelalamo.bsky.social•93 days ago

For nonsensical/intrinsically disordered seqs, the seq->str mapping manifested as spaghetti w/ AF2, or fictitious topologies w/ AF3, both low plddt. The latter just happen to be useful for dreaming up new backbone topologies, and MPNN can find appropriate sequences that refold w/ higher plddt

proteinator.bsky.social•93 days ago

That's an interesting assumption though I don't think this will work for something bigger than let's say 150-200aa. And clearly it will hallucinate helical bundles just arranged in slightly (maybe not) different topologies. Not sure it's of any meaning 😅

ddelalamo.bsky.social•93 days ago

I agree! Although it does give make sheets from time to time, just less often than helices (as is the case with RFdiffusion, GENIE, etc). Would need to compare this approach to relaxed sequence optimization from earlier this year

proteinator.bsky.social•93 days ago

Keep me in the loop, I'd be interested in seeing it! We just need someone to run this multiple times xD However, my guess would be that strands will approach 0% on avg in this setup.

proteinator.bsky.social•93 days ago

Typically gen models for protein structure are mode collapsed towards alpha-helices, AF3 won't be an exception here either if used in such a way. The reason why it hallucinates helices is just simply they're easy to learn as an optimal minimization of the diffusion loss fct during training

Comments

Posting Rules

Reply