Shaping the uncharted: Revealing the protein structure space from the perspective of generative models.

Abstract

Generative models can now design a diverse set of protein backbones, yet the quantification of distributional similarities of protein structure embeddings revealed that current models fail to capture the full spectrum of structural elements at different hierarchical levels. SHAPES (structural and hierarchical assessment of proteins with embedding similarity) quantifies these gaps and delivers a benchmark to guide next-generation protein design.