Multistate and functional protein design using RoseTTAFold sequence space diffusion

Area of Science:

• Computational biology
• Protein engineering
• Artificial intelligence in life sciences

Background:

• Protein denoising diffusion probabilistic models generate protein backbones but struggle with sequence-specific attributes and functions.
• Existing models lack the ability to simultaneously generate protein sequences and their corresponding structures.

Purpose of the Study:

• To develop an advanced protein design model capable of simultaneous sequence and structure generation.
• To overcome limitations in guiding protein generation with specific sequence attributes and functional properties.

Main Methods:

• Developed ProteinGenerator (PG), a sequence space diffusion model leveraging RoseTTAFold.
• Employed iterative denoising from a noised sequence representation to generate sequence-structure pairs.
• Guided generation using desired sequence and structural protein attributes.

Main Results:

• Successfully designed thermostable proteins with tailored amino acid compositions and internal repeats.
• Engineered cage structures for bioactive peptides like melittin.
• Created multistate protein triples where a single sequence adopts different structures based on context.

Conclusions:

• ProteinGenerator (PG) offers a powerful, generalizable approach for de novo protein design.
• The model enables integrated computational and experimental optimization of protein function.
• PG facilitates the design of proteins with complex structural and functional properties.