Efficient protein structure generation with sparse denoising models.

Area of Science:

• Computational Biology
• Biotechnology
• Protein Engineering

Background:

• Proteins are crucial biomolecules with diverse applications in research and medicine.
• Computational protein design aims to create proteins with specific functions.
• Current generative models struggle with large proteins and novel design tasks.

Purpose of the Study:

• To develop a novel computational method for protein structure generation.
• To overcome limitations of existing models in handling large protein sizes and adapting to new design challenges.
• To demonstrate the method's versatility in complex protein design tasks.

Main Methods:

• Introduction of salad, a family of sparse all-atom denoising models for protein structure generation.
• Development of structure editing, a sampling strategy to enable models to tackle unseen tasks.
• Application of the approach to motif scaffolding and multi-state protein design.

Main Results:

• Salad models are smaller, faster, and match or exceed state-of-the-art design quality.
• Successfully generated protein structures up to 1,000 amino acids in length.
• Demonstrated successful application to challenging tasks like motif scaffolding and multi-state protein design.

Conclusions:

• Salad offers a significant advancement in computational protein design, enabling the generation of larger and more complex protein structures.
• The combination of salad and structure editing provides a flexible and powerful platform for diverse protein engineering applications.
• This work expands the capabilities of generative models for designing proteins with tailored properties for biotechnology and biomedicine.