Interconvertibility of Electronic and Vibrational Circular Dichroism Spectra of Proteins: A Test of Principle Using Neural Network Mapping

Abstract

Electronic circular dichroism (ECD) and vibrational circular dichroism (VCD) are compared with respect to their interconvertibility for protein structural studies. ECD and amide I' VCD spectra of 28 proteins were used with a backpropagation projection neural network with one hidden layer to develop a mapping between the two spectral types. After the network converged, the number of neurons in the hidden layer was optimized by principal component analysis of the synaptic weights of the pilot network topology with redundant hidden neurons. Actual prediction of one spectrum from the other for individual proteins was tested by retraining these networks with 28 reduced training sets having one protein systematically left out. Comparison of network-predicted spectra with experimental ones is used to identify those spectral features which are unique in each method. Similarly, the VCD spectra of 23 proteins measured in both D₂O and H₂O in the amide I region were mapped onto each other with the use of the same type of neural network calculation. The results show that the effects of partial deuteration on the VCD spectra band shape are predictable from the H₂O spectra. An analysis of the synaptic weights of the optimized networks was performed which allowed identification of the linear and nonlinear parts of the obtained mappings. Insight into the details of how the neural networks encode and process the spectroscopic information is derived from a spectral representation of these weight matrices.