Bridging Sequence and Structure: Latent Diffusion for Conditional Protein Generation

Abstract

Protein design encompasses a range of challenging tasks, including protein folding, inverse folding, and protein-protein docking. Despite significant progress in this domain, many existing methods address these tasks separately, failing to adequately leverage the joint relationship between protein sequence and three-dimensional structure. In this work, we propose a novel generative modeling technique to capture this joint distribution. Our approach is based on a diffusion model applied on a geometrically-structured latent space, obtained through an encoder that produces roto-translational invariant representations of the input protein complex. It can be used for any of the aforementioned tasks by using the diffusion model to sample the conditional distribution of interest. Our experiments show that our method outperforms competitors in protein docking and is competitive with state-of-the-art for protein inverse folding. Exhibiting a single model that excels on on both sequence-based and structure-based tasks represents a significant advancement in the field and paves the way for additional applications.