Enhancing truncation event prediction in AAV vector genome designs

Summary: Truncation events in recombinant adeno-associated virus (rAAV) vectors pose a major challenge in gene therapy, affecting therapeutic efficacy and increasing production costs. To address this, we developed a deep learning model that predicts truncation hotspots with high precision, facilitating rational vector optimization. Our model was trained on a high-resolution dataset generated from PacBio SMRT and nanopore sequencing, capturing genome-wide truncation patterns across diverse rAAV constructs, including single-stranded (ssAAV) and self-complementary (scAAV) vectors. Key genomic features—GC content, DNA secondary structures, and nucleotide composition—were extracted to train a multi-output deep learning model integrating convolutional neural networks (CNNs) and bidirectional gated recurrent units (Bi-GRU). This approach accurately identifies truncation-prone regions, aligning closely with experimental data. By leveraging PacBio data for training, our model offers a transformative computational framework for optimizing rAAV vector design, improving stability, and enhancing the reliability of next-generation gene therapies.