Multimodal diffusion for joint design of proteinsequence and structure

Computational design of functional proteins is of fundamental and applied interests.  Data-driven methods, especially generative deep learning, has seen a surge recently.  In this study, we aim at learning joint distribution of protein sequence and structure for their simultaneous co-design. To that end, we treat protein sequence and structure as three distinct modalities (amino-acid type plus positions and orientations of backbone residue frames) and learn three distinct diffusion processes (multinomial, Cartesian, and SO(3) diffusions).  To bridge the three modalities, we introduce a graph attention encoder shared across modalities, whose inputs include all modalities and outputs are projected to predict individual modalities. Benchmark evaluations indicate that resulting JointDiff simultaneously generates protein sequence--structure pairs of better functional consistency compared to popular two-stage protein designers   Chroma (structure first) and ProteinGenerator (sequence first), while being more than 10-times faster.  Meanwhile, they show room to improve in certain self- and cross-consistency.