The logarithmic acid dissociation constant (pKa) reflects the ionization of a chemical, which affects lipophilicity, solubility, protein binding, and ability to pass through the plasma membrane. Thus, pKa is a physicochemical property that affects absorption, distribution, metabolism, excretion, and toxicity (ADME-Tox) of chemicals. The goal of this work was to develop a free and open-source computational model to predict pKa values, which could then inform models that predict chemical ADME-Tox properties. Experimental acidic and basic pKa values for 7912 chemicals were collected from the literature. Chemical structures were curated and standardized for quantitative structure-activity relationship (QSAR) modeling and then split into training (80%) and test (20%) sets. Continuous molecular descriptors, binary fingerprints, and fragment counts were generated using PaDEL then used to train predictive models using different machine learning methods, including support vector machines combined with k-nearest neighbors, extreme gradient boosting, and deep neural networks. The resulting computational models showed comparable predictivity on the training and test sets, with a root-mean-squared error around 1.5 and a coefficient of determination around 0.80. Performance of our three best models compared favorably with two commercial pKa predictors from ACD/Labs and ChemAxon that were used for benchmarking. The validated models were implemented in free and open-source packages and are available on GitHub.

Open-source QSAR models for pKa prediction using machine learning approaches

ACS Fall 2020 Virtual Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.