Excited to share that I am starting a lab focused on deep learning-based protein design, biophysics and fundamental biology at @MPI_Biochem and @GeneCenter_LMU in Munich with the Emmy Noether Programme this summer. Join us as a PhD or postdoc. More to follow soon!

Our work on hallucinating de novo symmetric protein assemblies w/ #AlphaFold + #ProteinMPNN is out: doi.org/10.1126/science.add1… Major collaborative effort w/ @BasileWicky and A. Courbet, @RobertRagotte @JustasDauparas @EliasKinfu @SamTipps @kribler @minkbaek, at @UWproteindesign 1/5

Biophysical characterization, 7 crystal (@RobertRagotte) and 3 CryoEM (A. Courbet) structures validated these regularized designs (termed HALs), ranging from dimers (e.g. the helix in sheet's clothing in in panel A) to giant rings with over 1500 residues and 10 nm diameters. 4/5

However, experiments on 150 hallucinations were a monumental failure (all insoluble). Activation maximization had generated adversarial sequences. ProteinMPNN sequence design (doi.org/10.1126/science.add2… @JustasDauparas et al.) rescued these backbones, yielding over 70% soluble! 3/5

Currently working on this with both a 6 and 10 amino acid alphabet, first experimental results should be coming in next week.

Excited to see what we can hallucinate next! Code and all design models, as well as experimental data here: github.com/bwicky/oligomer_h… 5/5

Really impressive, did the enzymes show a change in activity after MPNN sequence design?

We develop a MCMC framework to hallucinate protein homooligomers starting from random sequences, guiding them towards cyclic symmetry via a geometric loss function. The results are a set of fully de novo assemblies with intricate and diverse topologies. 2/5

Congratulations! Very much looking forward to the Wicky lab's research.

Structure prediction infers protein structure from protein sequence. Broadly speaking, hallucination methods approach the inverse problem: finding a sequence that encodes a desired structure.