ML-compatible experimental approaches to accelerate AAV engineering
Gene therapy has great potential for treating genetic diseases, but efficient and safe gene delivery remains a challenge. Recent advances in adeno-associated viral (AAV) vector engineering through directed evolution and rational design show promise in transforming gene therapy. However, conventional AAV capsid engineering relies on selections in model systems that are inefficient at identifying variants with enhanced functionalities that translate to human patients. The key challenge is searching vast capsid sequence spaces for variants that possess multiple traits of value to gene therapy that are ideally maintained across preclinical models and human patients. In this primer, I will present two data-driven approaches recently developed in my group that leverage ML to identify capsids with greater translational potential. The first approach, Fit4Function, enables a more systematic and efficient search for capsids that are optimized across multiple functions. Fit4Function is based on the design of capsid libraries that enable the collection of high quality, ML-compatible screening data for the purpose of training accurate models that can be used in combination. The second approach utilizes high-throughput in vitro receptor-targeting screens to rapidly identify thousands of capsid variants that can engage receptors on cell types of interest. This in turn makes it possible to more efficiently explore the sequence space using saturation mutagenesis and generative models trained on these data. These approaches demonstrate how wet lab experiments can be designed to enable ML that efficiently predicts capsids with greater translational potential that target cell types of interest in preclinical models and humans.