Towards machine learning prediction of the fluorescent protein absorption spectra

Abstract

We demonstrate that machine learning models trained on a set of features obtained from QM/MM molecular dynamic trajectories of fluorescent proteins can be used to predict the chromophore dipole moment variation upon excitation, the quantity related to the electronic excitation energy. Linear regression, gradient boosting, and artificial neural network- based models were considered using cross-validation on the training dataset. Gradient boosting approach proved to be the most accurate for both internal (R² = 0.77) and external (R² = 0.7) test sets.