Presentations describe use of in vivo screening and machine learning to accelerate the analysis, design, and validation of improved AAV vectors.
Cambridge, MA – May 11, 2021 – Dyno Therapeutics, a biotechnology company applying artificial intelligence (AI) to gene therapy, will describe ongoing enhancements to its machine learning platform in three oral presentations at the 24th annual meeting of the American Society of Gene and Cell Therapy (ASGCT), taking place virtually from May 11-14, 2021.
Dyno’s presentations, summarized below, highlight advancements of the company’s CapsidMap™ platform that enable increased capsid design efficiency, optimized experiments for in vivo capsid validation, and improved quality of in vivo measurements.
Efficient design of optimized AAV capsids using multi-property machine learning models trained across cells, organs and species, Tuesday May 11, 2021 at 5:45-6:00pm EST
Machine learning models improve AAV capsid design efficiency, defined as the probability that a designed variant will have improved function. We synthesized and barcoded capsid libraries containing 803,041 designed sequence variants of three natural AAV serotypes and measured their properties as delivery vectors both in vitro and in vivo. Single-property machine learning models trained on these data can improve the efficiency of library design by at least several hundred-fold. Furthermore, models trained on multiple properties help overcome data sparsity and measurement error, thereby improving model accuracy, and providing a more reliable interpretation of experimental results.
Risk-Adjusted Selection for Validation of Sequences in AAV Design Using Composite Sampling, Tuesday May 11, 2021 at 6:00-6:15pm EST
Using our machine learning models, billions of promising AAV variants can be designed and scored computationally for predicted fitness. Accurate high-throughput experimental testing is practically limited, however, to hundreds of thousands of variants. From these high-throughput measurements we must then choose on the order of hundreds of variants for more extensive experimental validation, to identify the best possible individual capsid variants.
To address this question of which capsid variants should be tested experimentally, we developed a novel optimization algorithm called Composite Sampling (CS). This algorithm reduces a pool of hundreds of thousands of variants in a way that maximizes the chances of including the best candidates in the validation set. We show the value of the method through two datasets, demonstrating that Composite Sampling is consistently better than conventional approaches for selecting high performing sequences for the validation set.
AAV Capsid Property Estimation Is Improved by Combining Single-Molecule ID Tags and Hierarchical Bayesian Modeling of Experimental Processes May 13, 2021 at 6:15-6:30pm EST
Data-driven capsid engineering can only be as good as the quality of the data. While our high-throughput barcoded approach to measuring capsid properties allows generation of large datasets needed to train machine learning models, measuring many capsids at once introduces biases and noise in the data. To remove these biases and noise from in vivo measurements, we built a probabilistic model of AAV packaging and transduction to allow inference of the distributions for the underlying properties being measured. This presentation will describe validation of improved property value estimates from the model in comparison to naive estimates using experimental controls.
More information on our presentations may be found in the ASGCT program
About Dyno Therapeutics
Dyno Therapeutics is a pioneer in applying artificial intelligence (AI) and quantitative in vivo experiments to gene therapy. The company’s proprietary CapsidMap™ platform rapidly discovers and systematically optimizes Adeno-Associated Virus (AAV) capsid vectors that significantly outperform current approaches for in vivo gene delivery, thereby expanding the range of diseases treatable with gene therapies. Dyno was founded in 2018 by experienced biotech entrepreneurs and leading scientists in the fields of gene therapy and machine learning. The company is located in Cambridge, Massachusetts. Visit www.dynotx.com for additional information.