Generating functional and multistate proteins with a multimodal diffusion transformer.

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Designing functional proteins with natural complexity remains a significant challenge.
• Existing generative models have limitations in creating proteins with specific functional attributes.

Purpose of the Study:

• To introduce ProDiT, a multimodal diffusion model for large-scale functional protein design.
• To unify sequence and structure modeling for enhanced protein generation.
• To enable the design of proteins with multiple functional states and specific effector responses.

Main Methods:

• ProDiT was trained on a dataset of 214 million proteins, incorporating sequences, 3D structures, and annotations.
• The model utilizes a multimodal diffusion approach, integrating sequence and structure information.
• A diffusion sampling protocol was developed to design proteins with multiple functional states.

Main Results:

• ProDiT generates diverse, novel proteins that retain essential active and binding site motifs.
• The model can be conditioned on a broad spectrum of molecular functions, covering 465 Gene Ontology terms.
• Enzymatic active sites were successfully scaffolded and allosterically deactivated by a calcium effector, demonstrating multi-state design capabilities.

Conclusions:

• ProDiT represents a significant advancement in protein design, capable of generating functional proteins at scale.
• The model addresses design specifications previously inaccessible to other generative approaches.
• ProDiT expands the capabilities of the protein design toolkit for future research and applications.