UniDiff introduces a novel framework for joint sequence-structure generation in proteins, leveraging Riemannian diffusion on product manifolds to respect protein structure geometry while accommodating discrete sequences. The method offers a coupled generation process, avoiding discretization or sequential staging, with theoretical guarantees for likelihood computation and uncertainty quantification.
Key findings
UniDiff treats sequences and structures as elements of a joint continuous latent space.
The framework uses Riemannian diffusion on product manifolds to respect protein structure geometry.
UniDiff introduces a multimodal score network with SE(3)-equivariant and permutation-equivariant branches.
The noising process corrupts both modalities while preserving conditional dependencies.
Coupled denoising procedure enables iterative refinement of sequences and structures.
Limitations & open questions
The paper is a research proposal and does not include experimental results or validation of the proposed methods.