Improving diffusion-based protein backbone generation with global-geometry-aware latent encoding

The global structural properties of a protein, such as shape, fold and topology, strongly affect its function. Although recent breakthroughs in diffusion-based generative models have greatly advanced de novo protein design, particularly in generating diverse and realistic structures, it remains challenging to design proteins of specific geometries without residue-level control over the topological details. A more practical, top-down approach is needed for prescribing the overall geometric arrangements of secondary structure elements in the generated protein structures. In response, we propose TopoDiff, an unsupervised framework that learns and exploits a global-geometry-aware latent representation, enabling both unconditional and controllable diffusion-based protein generation. Trained on the Protein Data Bank and CATH datasets, the structure encoder embeds protein global geometries into a 32-dimensional latent space, from which latent codes sampled by the latent sampler serve as informative conditions for the diffusion-based backbone decoder. In benchmarks against existing baselines, TopoDiff demonstrates comparable performance on established metrics including designability, diversity and novelty, as well as markedly improves coverage over the fold types of natural proteins in the CATH dataset. Moreover, latent conditioning enables versatile manipulations at the global-geometry level to control the generated protein structures, through which we derived a number of novel folds of mainly beta proteins with comprehensive experimental validation.