Development of Xanthene-Based Fluorescent Dyes: Machine Learning-Assisted Prediction vs. TD-DFT Prediction and Experimental Validation

Abstract

A large number of xanthene-based fluorescent dyes have been reported with unique photophysical properties. Further development of this group of useful chemicals processes challenges due to the massive amount of discrete data and unavoidable human errors in analyzing the data. Given recent advances in data analysis techniques, we integrated machine learning methods with a chemical database to assist identification of useful xanthene dyes in this study. Based on the xanthene dye database a machine learning model (named ATTRNN) was developed and applied in predicting excitation and emission wavelengths of six new dyes. The comparison of machine learning prediction with time-dependent density functional theory (TD-DFT) calculation, as well as experimental validation demonstrated the usefulness of ATTRNN. Moreover, the new dyes were used to develop fluorescent sensors for hydrogen sulfide and cysteine, which further proved the value of data-driven dye discovery.