Using Machine Learning Models to Predict the Impact of Template Mismatches on Polymerase Chain Reaction (PCR) Assay Performance

Abstract

Molecular assays are critical tools for the diagnosis of infectious diseases. These assays have been extremely valuable during the COVID pandemic, used to guide both patient management and infection control strategies. Sustained transmission and unhindered proliferation of the virus during the pandemic resulted in many variants with unique mutations. Some of these mutations could lead to signature erosion, where tests developed using the genetic sequence of an earlier version of the pathogen may produce false negative results when used to detect novel variants. In this study, we assessed the performance changes of 15 molecular assay designs when challenged with a variety of mutations that fall within the targeted region. Using data generated from this study, we trained and assessed the performance of seven different machine learning models to predict whether a specific set of mutations will result in significant change in the performance for a specific test design. The best performing model demonstrated acceptable performance with sensitivity of 82% and specificity of 87% when assessed using 10-fold cross validation. Our findings highlighted the potential of using machine learning models to predict the impact of emerging mutations on the performance of specific molecular test designs.