Machine Learning for the Label-free Detection of Viruses via Surface-enhanced Raman Spectroscopy
Open Access DepositedEarly virus identification is essential for both treatment management and spread reduction. Currently, nucleic acid tests, direct antigen, and antibody detection tests dominate viral diagnostics. However, each of these assays relies on the detection of a specific target, limiting their use when a pathogen presents atypically, during instances of coinfection, or when too many mutations occur. Surface-enhanced Raman spectroscopy coupled with machine learning can be used to rapidly detect and identify viruses in a label-free manner. This dissertation investigates both the mechanical parameters important for virus classification as well as the machine learning (ML) component, including exploration of model types, their feature importance and interpretability, classification confidence, and the use of ML-derived synthetic spectra. The results of this work demonstrate the robust, rapid, and label-free detection of both human polyomavirus 2 (aka JCPyV), a non-enveloped DNA virus, and influenza A virus, an enveloped RNA virus, using our platform consisting of gold-coated carbon nanotubes for virus capture, surface-enhanced Raman spectroscopy, and machine learning. While the case of JCPyV was less complex, with samples obtained from a single culture, limiting variation in both the viral concentration and mutations occurring, influenza A provided a more realistic classification problem. With influenza A, a high detection accuracy was achieved despite a 10-fold difference in viral concentrations, mutations arising from separate culture growths, and a 1,000- to 10,000-fold lower concentration than that of JCPyV. Ultimately, cross-validated AUCs reached 0.99 for JCPyV and 0.96 for influenza A identification. When differentiating between two closely related virus strains belonging to two subtypes, influenza A/H1N1 and influenza A/H3N2, the cross-validated AUC reached 0.95. By averaging 10 spectra versus assessing only 1, the overall accuracy increased to 0.98, effectively reducing the effect of samples that contained some non-representative spectra. It was determined that in each case, the convolutional neural networks (CNN) models reached the highest cross-validated and testing accuracy, though when utilized in the simpler JCPyV classification task, the traditional models also performed with greater than 0.96 AUC and were more easily interpreted for feature importance. For the task of synthesizing spectra to aid in model training when training samples were limited, Generative Adversarial Networks were able to produce highly precise virus-positive spectra, increasing the testing accuracy to greater than 90%. Ultimately, this dissertation substantiates the use of Raman spectroscopy and machine learning for the label-free and rapid identification of both JCPyV and influenza A viruses and provides an assessment of machine learning models for their use in spectral classification studies.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.