All data are available on Zenodo and we'd love to see what you can do with it! Cydney also made a nice notebook to run the predictive model (including DMS scan) in Colab!

Preprint: biorxiv.org/content/10.1...
Colab: colab.research.google.com/drive/1KNWvG...
Data: docs.google.com/forms/d/e/1F...

Exchanging for an authentic Wienerschnitzel 🤠

Folding models learn protein stability only implicitly. Without access to negative data, one can in principle make use of the folding free energy (dG) and the change in the free energy upon mutation (ddG). I believe simple aux losses could help for cases where a mutation is clearly disruptive.

Hi Thomas, that's indeed the case! However, this would also mean that one can reliably refold and score redesigned sequences only of those proteins whose structures were deposited post AF2 training cutoff date. This is a big limitation in our opinion!

FAMPNN or PPIpack?

github.com/richardshuai...

colab.research.google.com/github/Kuhlm...

Good luck and enjoy, Seb!

6/6 Amazing team work with great co-authors
@leif-seute.bsky.social, Nicolas Wolf, Simon Wagner, @bioinfo.se, Jan Stühmer and @graeterlab.bsky.social

Try out FliPS and BackFlip yourself, code and Google Colab tutorials are available on GitHub!

5/6 We introduce a framework in which we generate candidate protein structures conditioned on flexibility with FliPS and use BackFlip to select the candidates whose predicted flexibility profile best matches the target before running expensive MD simulations.

4/6 We also introduce BackFlip - an equivariant network that can accurately predict backbone flexibility as derived from MD simulations. Crucially, BackFlip infers flexibility solely from the backbone geometry without requiring evolutionary information, making it useful for de novo protein design.

3/6 In a series of experiments, we demonstrate that FliPS samples novel, realistic proteins with diverse secondary structure composition and a remarkable resemblance to custom target flexibility profiles, as verified in 300ns Molecular Dynamics (MD) simulations of the designed samples.

2/6 Our model FliPS is a conditional flow matching model for protein structure generation. FliPS receives a flexibility profile as conditional input feature and learns how to generate realistic protein structures while respecting target flexibilities.

Hahaha this is hilarious!! I’m now totally convinced I should play with it as well 🤣