Effect of Different Factors on Predicting Constants of Acidity of Low-Molecular Organic Compounds by Means of Machine Learning

A study is performed of the effect the way of standardizing the molecular structure and parameters of calculating molecular fingerprints has on the accuracy of predicting constants of acidity. It is shown that standardization (i.e., the choice of the tautomeric form and the way of writing the structure of the molecule) using OpenEye QuacPac gives the best results, but the RDKit library allows comparable accuracy to be achieved. It is established that how the charge state is chosen has a great effect on the accuracy of predictions. The accuracy of predictions depending on the radius (size of substructures) of circular molecular fingerprints is studied, and the best results are achieved using radius r = 2. A random forest, a machine learning algorithm, is used. It is also shown that the use of support vectors ensures fairly high accuracy when optimizing hyperparameters.