Machine Learning Identification of Capsid Mutations to Improve AAV Production Fitness

The adeno-associated virus (AAV) holds great potential for gene therapy efforts by providing a viable vector. However, current efforts are constrained by a lack of AAV variants that exhibit specific tropisms or immunogenicity and a lack of sustainable industrial projection. Departing from experimental approaches to addressing these issues, we built a model to predict residue mutations to improve AAV production fitness. Our model leverages the evolutionary paradigm and microenvironment characteristics by analyzing structural AAV data without needing domain knowledge or experimental fitness data for AAV as inputs. When testing our model's predictions for AAV2 residue mutations, we found a threefold increase in the percent of mutations yielding variants with better production fitness than wild type compared to random mutations, achieving a p-value of 7.46x10-12. Given these results, our machine learning approach of using structural data to approximate fitness data has the potential to accelerate AAV development.