Adeno-associated viruses (AAVs) are widely utilized in gene therapy, yet their clinical efficacy is often compromised by limited packaging efficiency and pre-existing immunity. To this end, a popular notion is to diversify the entire AAV capsid by, e.g., shuffling domains between different natural AAV serotypes. However, past attempts showed that existing algorithms fail to take into account the complexity of AAV packaging, leading to low proportions of functional variants. To tackle these issues, we first aim to enhance the existing SCHEMA-RASPP structure-guided protein library design algorithm by transforming it into a machine learning (ML) model that incorporates graph neural networks (GNNs). Specifically, this model will be trained on large-scale experimental and synthetic structure-function datasets derived from public databases and in-house chimeric AAV capsid libraries. We further aim to develop a novel, highly scalable in silico pipeline to map B-cell epitopes onto the predicted 60-mer AAV capsid structure by utilizing state-of-the-art biomolecular structure prediction models. We believe this project would lead to the generalization of the advancements in protein design research to AAVs, where progress has been relatively stale.