Prediction of protein secondary structure from circular dichroism spectra using artificial neural network techniques

Abstract

An approach to predict protein secondary structure is presented using circular dichroism (CD) spectra as input to two types of artificial neural networks (ANNs): (i) a three-layer backpropagation network and (ii) a hybrid self-organisation to backpropagation network. The dataset comprised the CD spectra of 22 proteins in the 178-260 nm wavelength range whose secondary structures were known. A total of 22 networks were trained by each method, using the jackknife technique for testing the prediction on each protein in turn. The performance, measured in terms of root mean square residuals and Pearson product-moment correlation coefficients, compares well with that obtained by other statistical and ANN methods, and is likely to improve with the growth of the dataset.