Integrating a partial least squares model with an artificial neural network to discriminate FTIR spectra of virus infected vero cells at 6 hours post exposure

A partial least squares regression (PLSR) model that classified the FTIR spectra of Mock, Coxsackievirus and HSV1 infected vero cells with sensitivities of 85%, 85% and 90%, respectively, was developed. Fourteen discriminators obtained by partial least squares regression performed on difference spectra from 20 independent infection experiments were examined. Sorting the FTIR absorbance of the five best discriminators and dividing them into equal intervals gave an accuracy of 92%. This constitutes proof of concept at 6 hours post exposure for discriminating two infections from each other and from Mock infected cells. When the PLSR model was combined with a 14×7×3 artificial neural network, the best sensitivity and accuracy were obtained with the neural network trained to an RMS error of 0.100 within 400,000 iterations. The sensitivities were Mock = 80%, HSV1 = 90% and Coxsackievirus = 70%, for an overall accuracy of 80% based on 10 Mock, 10 HSV1 and 10 Coxsackievirus test samples.