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

Recent breakthroughs in diffusion-based generative models have prompted de novo protein design, notably in generating diverse and realistic structures. Nevertheless, while existing models either excel at unconditional generation or employ residue-wise conditioning for topological control, explorations on a holistic, top-down approach to control the overall topological arrangements is still limited. In response, we introduce TopoDiff, a diffusion-based framework augmented by a structure encoder and a latent sampler. Our model can unsupervisedly learn a compact latent representation of protein global geometry, while simultaneously integrating a diffusion module to leverage this information for controlled structure generation. In benchmark against existing models, TopoDiff demonstrates comparable performance on established metrics and exhibits an improved coverage over the fold modes of natural proteins. Moreover, our method enables versatile control at the global-geometry level for structural generation, under the assistance of which we derived a number of novel folds of mainly-beta proteins with comprehensive experimental validation.