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

Abstract Background Gene therapy using adeno-associated virus (AAV) as a vector has been successfully applied in treating various monogenic diseases with significant efficacy. However, the presence of neutralizing antibodies (Nab) against AAV in patients' plasma hinders the delivery of AAV to the target tissue. Thereby monogenic diseases like hemophilia A/B, Glycogen Storage Disease Type Ia, and Duchenne muscular dystrophy, requiring systemic administration, are more significantly affected by Nab. In contrast, monogenic diseases suitable for local AAV administration such as eye and ear diseases minimize the impact of Nab. All patients undergoing gene therapy developed Nab, persisting at elevated levels even multiple years, which becomes the main obstacle limiting patients from receiving repeated dosing. Nab bind exclusively to the accessible surface regions of AAV capsid proteins, disrupting viral infection. These antigenic targets are classified as either linear epitopes or conformational epitopes. Importantly, certain antibodies recognizing conformational epitopes may also bind corresponding linear peptide fragments, enabling short peptides to act as effective decoys against Nab binding. This reciprocal recognition provides a mechanistic rationale for using short peptides as decoys to effectively block Nab binding to AAV. Natural Language Processing (NLP) techniques have been successfully applied to protein sequence research, which correspond to the structural, functional, and biochemical properties of proteins. According to prior studies, taking amino acid (aa) sequences as input, transformer-based models or pre-trained models frequently demonstrate excellent performance across various tasks, including contact prediction, structure prediction, generation models, and feature extraction. Immunogenicity is also a critical attribute of protein-based drugs, gene therapy vectors, and vaccines. The immunogenicity classification model trained on millions of aa sequences, leveraging the principle that the human immune system tolerates self-proteins but reacts to pathogen-derived proteins enables rapid identification and optimization of immunogenic regions in therapeutic candidates. Method Using a curated dataset of human- and virus-derived amino acid sequences, we trained peptide classification models based on modified BERT (Bidirectional Encoder Representations from Transformers) architectures with bidirectional Transformer encoders and the pretrained evolutionary-scale model ESM-2. After validating model performance, “virus-polarized” sequences within AAV2 and AAV843 capsids were identified to guide peptide selection. The antibody-blocking efficacy of these AI-predicted decoy peptides was confirmed using a leave-one-out (LOO) strategy and peptide-antibody binding assays. A similar AI-guided decoy development strategy proved effective in Nab in severe hemophilia A with inhibitors (HAI). Finally, the ability of decoy peptides to enhance AAV gene therapy was validated across one primary and two secondary in vivo models, alongside assessment of potential immune responses. To generate novel decoys, we developed a contrastive loss-trained variational autoencoder (VAE) with a BERT encoder to design and validate AI-generated peptide candidates. Results We compare two fine-tuning strategies: full-parameter tuning and low-rank adaptation (LoRA). Across multiple BERT-based encoders, the models achieve over 84% classification accuracy at the optimal aa length. Using these models, we identify short virus-polarized peptides from AAV capsids. A Leave-One-Out strategy validates the importance of these peptides in blocking preexisting and therapy-induced Nab, and in mediating direct or competitive binding to specific antibodies. This decoy strategy proves effective in patients with severe hemophilia A with inhibitors, demonstrating cross-disease potential. We confirm the peptides' inhibitory effects in one primary and two secondary AAV dosing models, and sequential administration of IgG-degrading enzymes (Ides) and decoy peptides blocks high-titer Nab without inducing inflammatory responses or immunogenicity in vivo. Furthermore, a contrastive loss-optimized variational autoencoder (VAE) with a BERT-based encoder designs peptides that effectively block Nab of AAVs. These findings highlight the potential of sequence-based AI algorithms to overcome high-titer Nab barriers in AAV gene therapies.