Deep diversification of an AAV capsid protein by machine learning

Area of Science:

• Biotechnology
• Protein Engineering
• Bioinformatics

Background:

• Engineered protein libraries often lack the sequence diversity found in natural protein families.
• Machine learning (ML) models offer a path to explore greater engineered protein diversity without biophysical modeling.
• Adeno-associated virus 2 (AAV2) capsid engineering is crucial for developing effective viral vectors.

Discussion:

• Deep learning models were applied to design diverse AAV2 capsid protein variants.
• The study focused on a specific 28-amino acid segment of the AAV2 capsid.
• Viable engineered capsids were generated, with many exhibiting significantly higher diversity than natural AAV serotypes.

Key Insights:

• Over 200,000 AAV2 variants were generated, resulting in over 110,000 viable engineered capsids.
• A significant portion of these variants displayed 12-29 mutations, exceeding natural AAV diversity.
• Deep neural networks accurately predicted capsid viability even with limited training data.

Outlook:

• This deep learning approach expands access to previously unreachable functional sequence space.
• Potential applications include the development of enhanced viral vectors for gene therapy.
• The methodology holds promise for engineering novel protein therapeutics with improved functionalities.