Sequence-based artificial intelligence-guided inhibitory peptides to counteract AAV neutralizing antibodies in gene therapy.

Adeno-associated virus (AAV) vectors have demonstrated strong clinical efficacy across multiple monogenic disorders, yet neutralizing antibodies (NAbs) remain major barriers to vector redosing and durable transduction. To address this limitation, we developed a sequence-based artificial intelligence framework integrating Bidirectional Encoder Representations from Transformers (BERT) and Evolutionary Scale Modeling with Low-Rank Adaptation (ESM-LoRA) to identify human-tolerant, virus-specific motifs across human and viral proteins using sliding-window analyses and multi-layer fine-tuning. Using capsid-derived segments from AAV2 and AAV843, we prioritized surface-accessible, virus-polarized peptides as candidate decoys and evaluated their ability to block AAV-targeted antibodies. The models achieved high accuracy in distinguishing human versus viral fragments, and selected peptides bound AAV-specific antibodies with nanomolar affinity, competitively restored capsid binding, and recovered >60% of baseline transduction in the presence of NAbs. In mouse models with pre-existing or redosing-induced antibodies, coadministration of decoy peptides and an IgG-degrading enzyme reduced high-titer NAbs to approximately 1:4 and restored hepatic transgene expression without detectable inflammatory or peptide-specific immune responses. A contrastive variational autoencoder (VAE) further generated non-natural peptides with broader epitope coverage. These results support sequence-driven AI design of short, low-immunogenicity decoy peptides as a translational strategy to mitigate NAb interference and enhance the feasibility of AAV redosing.