Prediction of Adeno-Associated Virus Fitness with a Protein Language-Based Machine Learning Model

Area of Science:

• Biotechnology
• Molecular Biology
• Bioinformatics

Background:

• Adeno-associated virus (AAV)-based gene therapies offer one-time treatment potential for various diseases.
• High manufacturing costs are a significant barrier to the widespread adoption of AAV therapeutics.
• Improving AAV capsid yield, or fitness, is crucial for reducing gene therapy costs.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting the fitness of Adeno-associated virus type 2 (AAV2) capsid mutants.
• To optimize engineered AAVs for manufacturability, addressing a gap in current gene therapy development.
• To create a computational tool that can accelerate the development of cost-effective gene therapies.

Main Methods:

• A hybrid machine learning approach combining a protein language model (PLM) with classical ML techniques was employed.
• The model predicts AAV2 capsid fitness based on the amino acid sequence of the capsid monomer.
• Model performance was evaluated using independent datasets, including multimutant AAV capsids.

Main Results:

• The developed ML model achieved high prediction accuracy for capsid fitness, with a Pearson correlation of 0.818.
• The model demonstrated robustness and generalizability across independent datasets, including complex multimutant capsids.
• The ML model serves as an effective surrogate for time-consuming in vitro experiments.

Conclusions:

• A novel ML model accurately predicts AAV2 capsid fitness from amino acid sequences.
• This computational approach can significantly reduce the time and resources required for AAV capsid engineering.
• The model has the potential to enhance the fitness of clinical AAV capsids, thereby improving the economic viability of gene therapies.