Isolation of structural constraints by ensemble-conditioned protein sequence design

Abstract

Natural sequences have long been the assumed ground truth for protein sequence design methods. But natural sequences encode more than the structural fold, carrying evolutionary history that reflects ancestry, drift, and function. Deep learning models for protein sequence design inherit this signal, achieving higher native sequence recovery but lower predicted foldability. Motivated by this trade-off, we introduce Caliby, a Potts model-based sequence design method capable of conditioning on structural ensembles. By conditioning on synthetic ensembles generated from an input backbone, Caliby samples sequences consistent with shared structural constraints while averaging out native sequence confounders. This strategy improves designability while reducing native sequence recovery, and generalizes across sequence design models. Mechanistically, designed sequences show reduced covariation between non-contacting residues and lower energetic frustration, indicating more precise structural encoding. We demonstrate experimental applicability by designing thermostable proteins, functional lanthanide binders, and soluble membrane protein variants, including a pentameric assembly confirmed by cryo-EM.