An Automated Workflow Leveraging Machine Learning for Physical Titer Determination from Cryo-TEM Images of Adeno-Associated Virus Capsids.

The application of gene therapies could be life-changing for many previously untreatable health conditions, but it is currently both financially expensive and time-intensive. Controlling the quality of these gene products involves monitoring impurities, such as partially filled and empty viral vectors, aggregates, and debris from the desired productfull viral capsids containing the genetic material for the gene of interest. Transmission electron microscopy (TEM), with its ability to identify nanometer-sized structures, provides a good approximation of ground truth to determine the presence of each of these species in production samples. Unfortunately, TEM methods are limited by the potential for sample damage during preparation and the difficulty in distinguishing viruses from cellular debris. They can also incur significant monetary and labor costs, as the images are often manually screened by trained individuals using subjective classification. Here, computer vision (CV) has been utilized to automate the labeling and classification steps within this analysis process, thereby improving the accuracy of impurity measurements in TEM imaging techniques. The manual analysis step may be superseded by a machine-learned model that predicts the classification of objects in TEM images. In this study, a CV model was trained to classify individual capsid images into one of three categoriesfull, partially full, or emptybased on the amount of DNA they contained. This model was then used to predict capsid category distributions with a 72% accuracy. A graphical user interface, Capsidize, was developed as an accessible and standardized method of labeling full, partial, and empty capsid physical titers in cryo-TEM images, reducing the active working time for labeling by 95%.