Kai Yi

1/7 Excited to share GuideFlip! We use guided discrete flow matching to co-design protein sequences and structures for flexible interactions, with experimental validation across three systems: α-synuclein, RBX1 and active-state GPCR nanobodies.

6/7 We further apply GuideFlip to a curated set of 285 human DisProt targets. For 177 (62.1%), at least one design passed our computational filters. These are predicted candidates. Explore complexes and download sequences: huggingface.co/spaces/ykiii...

1/7 Excited to share GuideFlip! We use guided discrete flow matching to co-design protein sequences and structures for flexible interactions, with experimental validation across three systems: α-synuclein, RBX1 and active-state GPCR nanobodies.

4/7 For RBX1’s disordered N-terminus, GuideFlip gave 5/12 hits across two rounds (41.7%). After 2nd place in the GEM × Adaptyv competition, follow-up designs reached KD = 4.5 nM. Thanks to GEM and @adaptyv.bio for organizing, and @adaptyv.bio for testing!

1/7 Excited to share GuideFlip! We use guided discrete flow matching to co-design protein sequences and structures for flexible interactions, with experimental validation across three systems: α-synuclein, RBX1 and active-state GPCR nanobodies.

2/7 GuideFlip combines an all-atom inverse-folding prior (ADFlip) with gradients from AlphaFold2. Binder residues are assigned progressively as the complex is re-predicted, allowing sequence and the binding interface to evolve together.

Thrilled with the our results with @kjamali.bsky.social and @sjorsscheres.bsky.social from the recent binder competition! 🎉 2 of our 7 designs successfully bound to the target, with affinities ranking 2nd and 3rd overall. All of these designs targeted the disordered region of the protein.