I4 - Molecular and Cellular Methods (Including assess vector integration, genome integrity, and outcomes)
347: Predicting Truncation Events in AAV Vector Genome Designs Using Deep Learning
Type: Oral Abstract Session
Presentation Details
Session Title: AAV Vector Integration
The presence of truncated genomes found in recombinant adeno-associated virus (rAAV) vector preparations can have significant impact on the effectiveness of gene therapy treatments that are based on these biodrugs. The packaging of truncated genomes is known to be caused by template-switching events centered on inverted repeats and sequences with high GC content. Importantly, they may result in the production of non-functional vectors with unknown consequences. The propensity of some vector design elements to drive the formation of truncated genomes is not completely known and can vary from batch-to-batch. Therefore, the ability to predict truncation hotspots in AAV vector designs is crucial for developing future vectors. In this study, we aimed to use deep learning (DL) to identify truncation events in AAV vector genome designs, de novo. We have accumulated multiple datasets related to the detection of truncated genomes from previous and ongoing studies that profile vector genome heterogeneity in packaged rAAVs using single molecule, real-time (SMRT) and nanopore sequencing technologies. Additionally, recognizing the intricate variability in AAV vector design, our study encompasses an extensive diversity of design features including different promoters, transgenes, single-stranded (ss) versus self-complementary (sc) configurations, diverse capsid types, and various preparation methods. We hypothesized that using a large dataset of detected truncation events found among a diversity of rAAV designs to train a DL model, we can develop an accurate truncation hotspot prediction algorithm. We have defined as features across the training data, the relative positions of the truncation events and truncation abundances centered at 200-nt sliding windows. Using this metadata, we trained a combination of convolutional neural network (CNN) and bi-directional long short-term memory (LSTM) models to predict truncation events. The model is designed to classify whether a sequence will house a truncation hotspot and predict the “strength"" of the truncation. The model incorporates both regression and classification components and achieved remarkable classification accuracy (98%). Crucially, the algorithm can accurately predict truncations in vector designs tested thus far. This study illustrates the potential for DL models to predict genome truncations in rAAVs. These advancements will not only aid in designing more effective gene therapy vectors, but also deepen our understanding of AAV’s biophysical properties.
G.G. and P.T. are corresponding authors.
Sandhiya Ravi1,2, Mitchell Yip1, Ngoc Tam Tran1,2, Suk Namkung1, Guangping Gao1,2, Phillip Wl Tai1,2
1Horae Gene Therapy Center, UMass Chan Medical School, Worcester, MA,2Department of Microbiology and Physiological Systems, UMass Chan Medical School, Worcester, MA"
Find This Session