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

Area of Science:

• Biophysics
• Computational Biology
• Structural Biology

Background:

• Circular dichroism (CD) spectroscopy is a key technique for analyzing protein secondary structure.
• Predicting detailed secondary structure parameters (DSSP) from CD spectra traditionally involves complex analysis.
• Deep learning offers advanced image recognition capabilities applicable to spectral data analysis.

Purpose of the Study:

• To develop a novel deep learning method for accurate prediction of DSSP parameters from CD spectra.
• To leverage convolutional neural networks (CNNs) for image-based analysis of CD spectral data.
• To introduce DACARI (Deep-learning Assisted Circular dichroism Analysis and Recognition Inference) for enhanced protein structure analysis.

Main Methods:

• Converted numerical CD spectral data into three-layer (RGB) images.
• Utilized a custom-trained convolutional neural network (CNN) for image recognition.
• Compiled a dataset of 243 CD spectrum RGB images and corresponding DSSP parameters from the Protein Circular Dichroism Data Bank.

Main Results:

• Achieved an overall correlation coefficient of 0.96 between predicted and ground-truth DSSP parameters.
• Obtained low Root Mean Square Deviation (RMSD) values for key parameters: 0.057 (α-helix), 0.055 (β-strand), and 0.048 (loop/irregular).
• Validated the method by observing increased α-helix content in α-synuclein upon SDS addition.

Conclusions:

• The developed deep learning method (DACARI) accurately predicts protein DSSP parameters from CD spectra.
• Image-based analysis using CNNs provides a powerful approach for interpreting CD spectral data.
• DACARI demonstrates potential for advancing the study of protein structure and conformational changes.