Chemical feature-based machine learning model for predicting photophysical properties of BODIPY compounds: density functional theory and quantitative structure–property relationship modeling

Context

Boron-dipyrromethene (BODIPY) compounds have unique photophysical properties and have been applied in fluorescence imaging, sensing, optoelectronics, and beyond. In order to design effective BODIPY compounds, it is crucial to acquire a comprehensive understanding of the relationships between the structures of BODIPY and the corresponding photoproperties. Fifteen molecular descriptors were identified to be strongly correlated with the maximum absorption wavelength. The developed ML/QSPR model exhibited good predictive performance, with coefficients of determination (R²) of 0.945 for the training set and 0.734 for the test set, demonstrating robustness and reliability. A posterior analysis of some of the selected descriptors in the model provided insights into the structural features that influence BODIPY compound properties; meanwhile, it also emphasizes the importance of molecular branching, size, and specific functional groups. This work shows that applied combined cheminformatics and machine learning approach is robust to screen the BODIPY compounds and design novel structures with enhanced performance.

Methods

In the present study, all the BODIPY models studied were fully optimized, and the corresponding absorption spectrum was obtained at DFT/TDDFT//B3LYP/6-311G(d,p) level. All the above calculations were executed by the Gaussian 16 program. Based upon the theoretical computational results, the machine learning-based quantitative structure–property relationship (ML/QSPR) model was employed for predicting the maximum absorption wavelength (λ) of BODIPY compounds by combining hand-crafted molecular descriptors (MD) and explainable machine learning (EML) techniques using Scikit-learn python library. A dataset of 131 BODIPY compounds with their experimental photophysical properties was used to generate a diverse set of molecular descriptors capturing information about the size, shape, connectivity, and other structural features of these compounds using Chemaxon and Alvadesc software. A genetic algorithm (GA) variable selection together with the multi-linear regression (MLR) method were applied to develop the best predictive model using the Genetic Selection python library.