always amazes me that you can take code written on an A10, change a single line and get power utilization like this in JAX

hack to get more diverse binders out of BoltzGen: sample an unconditional monomer, then sample a binder with secondary structure conditioning to match the monomer. this solves the long-standing open problem of generating 76 AA binders that aren't ubiquitin.

It's a very cool result but IMO there are caveats. Inference is (mostly) slower. There is existing work on faster models (e.g. MiniFold or protenix mini), and also existing work on ensemble prediction. I doubt this works without training on AFDB, which bakes in inductive bias from triangle layers

Recently tested some de-novo minibinders against two targets (thanks Adaptyv!) designed using our open-source design library, `mosaic`; our best method got hit rates of 7/10 and 8/10 and affinities as low as single-digit nanomolar. Wrote up some thoughts here: blog.escalante.bio/minibinder-d...

We added Boltz-2 to our protein design package! Under the hood this relies on a JAX translation, which, thanks to @jeremywohlwend.bsky.social and @gcorso.bsky.social ’s clean code, was fairly easy to write. We’ve been getting great results -- and we have yet to explore the most exciting new features

Designed my (the?) first minibinder using @gcorso.bsky.social 's new Boltz-2 model. Excited to explore how templates and increased MSA dropout impact binder design. This is work in progress but will be open sourced as usual