Isolation of structural constraints by ensemble-conditioned protein sequence design

Caliby, a Potts model-based protein sequence design method conditioned on structural ensembles, improves designability while reducing native sequence recovery across sequence design models.

• Why it matters: Natural sequences embed evolutionary history that deep learning design methods can exploit, yielding higher native recovery but lower predicted foldability; this work targets that trade-off.
• What they did: Caliby conditions sequence sampling on synthetic structural ensembles generated from an input backbone, averaging out native sequence confounders while enforcing shared structural constraints.
• The result: Designed sequences show reduced covariation between non-contacting residues and lower energetic frustration, and experimental designs include thermostable proteins, functional lanthanide binders, soluble membrane variants, and a pentameric assembly confirmed by cryo-EM.