Open source application for small molecule pKa predictions

The acid-base dissociation constant (pK_a) of a drug has a far-reaching influence on pharmacokinetics by altering the solubility, membrane permeability and protein binding affinity of the drug [1,2]. To the best of our knowledge, there is no publicly available, open source and license-free pK_a prediction tool that can reach the quality of commercial tools. Our goal is to develop a highly accurate pK_a prediction tool based on a mixture of freely accessible and commercial data which is however still free to use for everyone.

To do so, we identified multiple freely available experimental pK_a datasets, including data from DataWarrior, ChEMBL and various scientific publications. Additionally, we have access to several commercial data sets, for example from Novartis [3] and OpenEye [4].

We started with monoprotic molecules obtained from ChEMBL and DataWarrior to evaluate the dataset quality and our modelling concepts. After preprocessing 5994 unique structures with a pK_a value between 2 and 12 were used for machine learning. We tested seven different machine learning configurations including four different basic regressors together with six unique descriptor/fingerprint sets resulting in a total number of 42 trained and 5-fold cross-validated models. Additionally, we evaluated the models with two external test datasets. The results have been published in March 2020 [5]. Furthermore, we investigated how Graph Convolutional Networks and QM-based approaches can be used to further improve prediction quality.

To be able to predict the pK_a values of multiprotic molecules, two major problems had to be solved: Localization of the titratable groups without licensed software and the once-only assignment of the experimental values to the corresponding groups for all datasets. For the localization part we evaluated the results of Marvin [6] and Dimorphite-DL [7] to compile a list of 24 SMARTS pattern that catch almost 90% of all groups in our combined dataset of over 17000 unique molecules. Finally, the Marvin [6] predictions were used to assign the experimental values to the corresponding group while removing outliers. The resulting data set can be used as a starting point for machine learning in a following step.

All data and code can be found at https://github.com/czodrowskilab

[1] Charifson, P. S., & Walters, W. P. (2014). Acidic and Basic Drugs in Medicinal Chemistry: A Perspective. Journal of Medicinal Chemistry

[2] Manallack, D. T. (2007). The pKa Distribution of Drugs: Application to Drug Discovery. Perspectives in Medicinal Chemistry

[3] Richard A. Lewis, Stephane Rodde, Novartis Pharma AG, Basel, Switzerland

[4] pKa COMPLETE_DATABASE v1.13: OpenEye Scientific Software, Santa Fe, NM.

[5] Baltruschat M and Czodrowski P. Machine learning meets pKa [version 2; peer review: 2 approved]. F1000Research 2020, 9(Chem Inf Sci):113

[6] Marvin 20.1.0, 2020, ChemAxon Ltd, http://www.chemaxon.com

[7] Ropp PJ, Kaminsky JC, Yablonski S, Durrant JD (2019) Dimorphite-DL: An open-source program for numerating the ionization states of drug-like small molecules. Journal of Cheminformatics