Prediction of second-order rate constants between carbonate radical and organics by deep neural network combined with molecular fingerprints

Carbonate radical is among the most important environmental relevant reactive species which govern the transformation and fate of pharmaceutical contaminants (PCs). However, reaction rate constants between carbonate radical and most of the PCs have not been experimentally determined, and quantitative structural-activity relationships (QSARs) have not been established for rate estimation. This study applied MaxMin data processing method and used molecular fingerprints (MF) as the input of a deep neural network (DNN) to predict the rate constants between carbonate radical and organic compounds. MF parameters and the hyper-structure of the DNN were adjusted to yield satisfactory accuracy of rate prediction. The vector length of 512 bits with radius of 1 for MF and 5 hidden layers gave the best performance. The optimized MaxMin-MF-DNN model was compared with some of the most commonly used QSARs and machine learning methods, including random data splitting, molecular descriptors, supporting vector machine, decision tree, etc. Results showed that the MF-DNN model out-performed the other methods by more than 10% increase in prediction accuracy. Applying this MF-DNN model, we estimated reaction rates between carbonate radical and pharmaceuticals used in human medicine (1576) and veterinary practice (390). Among them, 46 drugs were identified as fast-reacting compounds, suggesting the important relations of their environmental fate with carbonate radical. This work combined deep neural network combined with molecular fingerprints to develop a QSAR model which successfully predicted the second-order rate constants between carbonate radical and organics. [Display omitted]