Abstract
Structure-conditioned sequence design models aim to design a protein sequence that will fold into a given target structure. Deep-learning-based approaches for sequence design have proven highly successful for various protein design applications, but many non-idealized backbones still remain out of reach for current models under typical in silico success criteria. We hypothesize that training objectives prioritizing native sequence recovery unintentionally push models to reproduce non-structural signals (e.g. phylogenetic relatedness, neutral drift, or dataset sampling biases), rather than a broadly generalizable structure-sequence mapping. Inspired by recent work bridging sequence likelihood and fitness prediction in protein language models, we introduce Caliby, a Potts model-based sequence design method capable of conditioning on an ensemble of structures. Conditioning on a synthetic ensemble generated from an input backbone allows sampling of sequences consistent with the structural constraints of the ensemble while averaging out undesired biases towards the native sequence. Ensemble-conditioned sequence design with Caliby reduces native sequence recovery while substantially improving AlphaFold2 self-consistency, outperforming state-of-the-art models ProteinMPNN and ChromaDesign on both native and de novo backbones. Finally, we train a variant of Caliby on only soluble proteins and demonstrate in silico that Protpardelle-1c binder designs that were previously deemed undesignable by SolubleMPNN are actually designable under SolubleCaliby, highlighting limitations of existing filtering pipelines. These results suggest that Caliby can expand the de novo design space beyond highly idealized backbones.