Thread by @chrisfrank662

Roughly one year ago we started out venturing in the world of #proteindesign. Today I am very happy to share our first preprint ‘Efficient and scaleable de-novo protein design using a relaxed sequence space’ (biorxiv.org/content/10.110…) with @sokrypton & @hendrik_dietz 1/x

Here we propose and experimentally verify a new way to design protein backbones. We use a relaxed sequence representation allowing the free following of gradients during hallucination. This enables a very efficient search for backbones that minimize the loss function. 2/x

By using ProteinMPNN we can create sequences for this backbones and our approach achieves SOTA backbone recovery both for ESM-Fold and AF2 single sequence, with median sub 4 A accuracy for 600AA proteins. Standing in line with recently published diffusion models! 3/x

We experimentally verified our design pipeline by expressing unconditional monomers and were able to produce proteins up to 600AA in size with solubility, SEC and CD success rates of up to 100% for several sizes of tested proteins. Also they are super stable! 4/x

Beside monomers we could also generate homodimers and -trimers as well as functional heterodimers. These heterodimers are expressed separately and when mixed together they form dimers, indicating that our system can efficiently design specific molecular interactions 5/x

Finally @AKhoshouei managed to get some CryoEM structures for three of our 600 AA monomers, perfectly fitting to the designed backbones! 6/X

It was very fun project and I am very thankful to everybody in the team! We learned so much, made so many mistakes, but with a great team and a lot of elbow grease everything is possible! Also special thanks to @ShihaoFeng18 and twitterless Yosta and Dominik! 7/x

This approach is now integrated into the ColabDesign framework and freely available to everybody! We are looking forward to further develop ColabDesign to expand its capabilities! Thanks again to @sokrypton for creating this awesome tool!