"look, i cannot guarantee our proteins will fold better in space, BUT" 😂

But great work! Will hopefully try it soon! :)

What I find super interesting when looking at some of your samples is that (much like chroma) the generated large proteins look somehow more 'brutalistic' or more 'regular' while the hallucinated ones seem more 'natural'. Are there also failed (aka non designable) pdbs in the figshare archive?

Let me/us know what you think about it! Also, the af2cycling can be used on any other sort of input and we are happy to explore more applications!

And one more note: In the beginning we were skeptical about this work, as it posses the real risk of being used to create designs with serious biosecurity risks 😉😅

Code is available through Colabdesign: github.com/sokrypton/Co...
And join the ColabBio Discord for more discussion!

This was/is a super fun collaboration with @sokrypton.org & @hendrikdietz.bsky.social Thanks to the team, Dominik and Lara for all the awesome work!

Now, where is the limit? Well, we went even further, testing 2000 AA designs in silico, but up to this point we haven’t succeeded in finding a designable sequence for our protein pretzel :D

We tested just one sequence, which worked out of the box, and we got clean samples right after the Nickle-NTA purification. TEM analysis also showed that the designed protein seems to agree very well with the designed shape.

We then wanted to push the size limits even further. When designing a 1500 AA triangle shaped protein we noticed that neither ESMFold, nor AF2 with IG & AA could predict the sequences confidently, so we went with AF3, which did a better job.

I was honestly surprised how well this worked. The TEM micrographs are so cool :D Just a bunch of tiny shapes

The 12 proteins expressed well and nsTEM analysis of them revealed that for 10 of them they match their designed shape nearly perfectly, as seen by 2D class averages :D

To investigate if this improvements in in silico success would translate in wet lab results, we went on a very fun journey, designing proteins shaped like letters of the Latin alphabet using Chroma and the af2cycler. The af2cycler improved the designability of the chroma designs strongly

But this small changes result in large improvements in designability. Making even designs of 1000 AAs readily designable. Just like RSO!

The af2cycler changes the backbone only slightly, keeping the overall structure the same, but correcting small inconsistencies.

The af2cycler takes in Chroma designed sequences and structures and performs multiple rounds of denoising, by employing a combination of AF2 predictions and ProteinMPNN sequence design. The resulting backbones are subject to solubleMPNN and then repredicted with ESMFold

We thought that by employing some sort of denoising could rescue these designs and improve designability. We did a lot of tinkering, but are finally able to share the af2cycler, a denoising framework based on a combination of AF2 and ProteinMPNN.

This project started a year ago with the introduction of #Chroma and the observation that the Chroma designs look promising, but often fail when subject to strict in silico filtering.

Gonna repost it later here 🙌