FuncDiff is a novel framework that integrates protein sequences, structures, and functions through a unified continuous-discrete diffusion process, introducing innovations in multimodal hybrid diffusion architecture, function-guided conditioning, and co-design optimization.
Key findings
FuncDiff introduces a multimodal hybrid diffusion architecture for simultaneous denoising of continuous atomic coordinates and discrete amino acid tokens.
Incorporate Gene Ontology terms, enzyme commission numbers, and structural motifs as controllable constraints in protein design.
A co-design optimization objective ensures consistency across sequence, structure, and function spaces.
Limitations & open questions
The model's performance on real-world protein design applications needs further validation.
The integration of more complex functional constraints into the generative process is an open challenge.