One of the coolest parts of this paper is that we managed to obtain structures of these "hinges" in both states, both on their own and in the context of much larger protein assemblies

In this work we designed proteins to switch their shapes contingent on an input (shown as a yellow helical peptide here)

Lastly, we show the PTH binder can be used in diagnostics and that it can be used to detect PTH in human serum, (in both cases with extremely minimal hassle - it kind of "just works"). We always say this should be true of designed proteins but it isn't always the case 😅 (10/11)

We also show that we can use "partial diffusion". Preetham started from experimental hits to resample around their starting point (which we show for Glucagon and neuropeptide Y), gaining orders of magnitude improvement in affinity. (9/11)

Using this approach, we generated sub-nanomolar (i.e. unprecedentedly tight) binders to a peptide from Bim, a protein that is involved in apoptosis, and to PTH. These binders are some of the tightest binders ever designed without any experimental optimization. (8/11)

Around this time, Joseph Watson, Dave Juergens and others were developing a protein diffusion model - a deep learning model trained to "denoise" the structure of a protein. Once trained it can then be asked to generate realistic structures from random noise. (7/11) #diffusion

We also tried "threading" sequences of target peptides into a library of structures I had designed to bind helical peptides. This worked to get a tight binder to secretin. However, this binder wasn't very specific against the closely related glucagon hormone. (6/11)

Another thing we tried was "hallucination" - taking steps in sequence space with the target sequence fixed, and maximizing the alphafold confidence metrics of the resulting prediction. Joseph Watson and Joseph Rogers showed these worked quite well - see figure S4! (5/11)