Deep-Learning Prediction of Protein Secondary Structure from Circular Dichroism Spectrum Using Three-Layer Image Recognition

In this study, we developed an image-recognition-based deep-learning method for accurately predicting the DSSP (define secondary structures of proteins) parameters from a circular dichroism (CD) spectrum. Focusing on the inherently high image-recognition capability of deep learning based on a convolutional neural network (CNN), we converted CD spectral numerical data into a three-layer (RGB) image with our unique method and deduced the eight DSSP parameters. We prepared the data set consisting of 243 CD-spectrum RGB images and corresponding DSSP parameters using CD spectra in Protein Circular Dichroism Data Bank and constructed well trained original CNN. The overall correlation coefficient between predicted and ground-truth DSSP parameters was 0.96. The RMSD values between prediction and true values for principal parameters were 0.057, 0.055, and 0.048 for the α-helix, β-strand, and loop-or-irregular contents, respectively. The validity of the proposed deep-learning method was confirmed by experiments in which natively disordered protein α-synuclein exhibited a significant increase in α-helix content upon addition of sodium dodecyl sulfate beyond its critical micelle concentration. We name the method proposed in this paper DACARI (Deep-learning Assisted Circular dichroism Analysis and Recognition Inference).