Optical-spectroscopy provides powerful toolkits to decipher proteins dynamic structures. However, the theoretical analysis of spectroscopic signals and connecting them with structural detail is a challenging task. Moreover, the intrinsic complexity of spectroscopic signals of proteins makes it difficult to correlate spectral characteristics with the underlying molecular structure and dynamics. Herein, we have developed data-driven machine learning (ML) protocols that can predict infrared (IR), ultraviolet (UV) and circular dichroism (CD) spectra of proteins with 3 to 5 orders of magnitude reduced computation cost compared to direct quantum chemistry calculations, providing effective tools to identify/probe secondary structure, protein folding, and structure variations with temperature. More importantly, our ML algorithms reduce the time and complexity of connecting two-dimensional (2D) optical-spectra of peptide segments to secondary structures. A convolutional neural network (CNN) model was trained and tested on a dataset consisting 87993 spectra computed from peptide segments with α-helical, β-sheet, and other typical secondary structures. The secondary structure classification accuracy reached near 100% and over 98.7% on spectra sets of new segments extracted from the same and homologous proteins, respectively. We aim to realize cost-effective relations between spectra and structure, i.e. spectra determination/prediction from structural information, and configuration determination/recognition from spectroscopic signals.

Machine learning simulation of two dimensional infrared and ultraviolet spectroscopy of proteins

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.