We’re very happy to announce that our RFdiffusion manuscript is now on bioRxiv! A lot can change in a week - we’ve now tested over a thousand designs and there’s so much exciting new data! 🧵
GIF
David Juergens
3,452 posts
David Juergens
@DaveJuergens
Machine learning for molecular design. Currently: Postdoc @StanfordUChem @mtzlab. PhD @UWMolES @UWproteindesign. B.S. @UWChemE
Palo Alto, CA
Born October 10, 1996
Joined April 2013
- We have made RFdiffusion free and open source for everyone! Please go check out the GitHub, and the Google Colab! And have fun designing with it!!Today we're making RF Diffusion, our guided diffusion model for protein design with potential applications in medicine, vaccines & advanced materials, free to use. The software has proven much faster and more capable than prior protein design tools. bakerlab.org/2023/03/30/rf-…
GIF - I got my PhD this week. What a wonderful few years it’s been.
Today we celebrate @DaveJuergens with a successful thesis defense! Charcuterie plating lead by our own @Hannahlhan and Ann. - ⚛️ I got a new job ⚛️ After 5 wonderful years at @UWproteindesign and @UWMolES, I’m headed to Palo Alto for a postdoc at @StanfordUChem under Todd Martinez. SUPER excited to dive into deep learning for quantum chemistry and explore the Bay Area!!
- Maybe hot chocolate wants to be called beautiful chocolate just one time.
- Macrocycles are SO COOL. We put together a pipeline for making cyclic peptides that bind to protein targets with RFdiffusion, which we call RFpeptides 😀 Excited to have this online, and so excited for the future of macrocycle design + deep learning!Accurate de novo design of high-affinity protein-binding macrocycles using deep learning @UWproteindesign 🚀 New preprint from David Baker!🚀 • Introducing RFpeptides, a diffusion-based framework for designing high-affinity macrocyclic peptide binders to diverse protein
- Replying to @DaveJuergensThe findings were astonishing. RFdiffusion, alongside better computational filtering methods, gives ~100x the success rate of previous methods!
- Training and inference code for, at a minimum, RFdiffusion version from Anna and Sam’s new enzyme paper will be released (as stated in the supplement).While 9%-88% hit rate is incredible, outdoing all previous methods by a large margin, including physics based and experimental based methods, they yet again are not releasing code. I say we just retrain RFdiffusion using their synthetic data and a flow matching objective.
- Replying to @DaveJuergens@HelenEisenach and @andrewjborst verified by electron microscopy that with RFdiffusion, we can make icosahedral nano-cages! This will be transformative for areas such as vaccine design.
- Replying to @DaveJuergens@spellock22 and Nikita Hanikel then tested our nickel-binding designs. These are C4 symmetric oligomers, designed to match the coordination geometry of the metal. This hasn’t been done before, but nearly half of designs worked! This approach could revolutionize materials design.
- Interesting structure image they’ve inconspicuously placed in Fig 1 panel F… Reminds me of a recent paper about protein diffusion models 🤨
sequence design with a protein LM by adding cross attention to structure encodings, pushing sequence recovery close to ~60%. From Bytedance AI arxiv.org/abs/2302.01649 - Replying to @DaveJuergensFinally, we tested our designs that scaffold the helix from p53. We wanted to achieve higher affinity binding to the MDM2 protein, to block its interaction with native p53 (a goal in cancer therapy). It worked, with some designs binding 1000x as tightly as the p53 helix alone!
- Replying to @DaveJuergensAnd that’s everything! The paper will now go through the peer review process, and we’re going to keep developing RFdiffusion with the code made available to everyone as soon as possible.
