Machine-guided rapid decision-making on quality of viruses as products and for viral clearance determination
Develop a computational means to integrate results of rapid viral titer assays (e.g., qPCR, ddPCR, optical imaging, probe-capture methods) to best match results of high precision determination of functional titer.
Categories
Vaccines
Assays
Process control
Project status
100% Completed
Industry Need
Develop a model that can better predict functional titer from lower-cost methods of analysis compared to using transmission electron microscopy (TEM) images of viruses
Solution
Carnegie Mellon University and Sartorius aim to develop a computational workflow to integrate results of rapid viral titer assays to best match the outcomes of high precision determination of functional titer.
The two organizations will incorporate and validate their methods by using machine learning (ML) approaches.
Experiments will include:
Analysis of physical and functional virus titers at various concentrations, from bioreactor supernatant in the presence of cell lysate or other confounding matrices
Determination of physical characteristics of infected cells
This project will determine the workflow that optimizes analytical assay reproducibility and repeatability and will predict gaps in knowledge that may limit product analysis.
Outputs/Deliverables
Multiclass Classifier for AAV8 Capsid Detection:
Creation of a classifier to automate the detection of full and empty AAV8 capsids from TEM images.
Achieved 98% accuracy in prediction and 91% selectivity for non-full (non-viable) capsids.
Benchmarking Quantitative Methods:
qPCR analysis was identified as the best method for correlating with TEM images regarding full/empty capsid ratios.
ELISA was found to overpredict capsid numbers as the percentage of empty capsids increased.
The viral lysis method was shown to significantly impact PCR quantification results, with heat lysis yielding the highest values.
Development of New Analytical Tools:
Introduction of Micelle Tagging Electrophoresis (MTE) for measuring AAV8 virus levels effectively in purified samples.
Neural Network Model Development:
Pretrained on synthetic data to improve misclassification and achieved high accuracy and selectivity.
Additional Project Information (Members Only)
Login to the NIIMBL member portal to access more, including: