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

Area of Science:

• Computational chemistry and cheminformatics.
• Application of machine learning in drug discovery and chemical property prediction.

Background:

• The logarithmic acid dissociation constant (pKa) is crucial for predicting chemical absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties.
• Existing pKa prediction software is largely proprietary, creating a need for accessible, open-source alternatives.

Purpose of the Study:

• To develop and validate free and open-source quantitative structure-activity relationship (QSAR) models for predicting the strongest acidic and strongest basic pKa values of chemicals.
• To provide a publicly available resource for pKa prediction, facilitating research in computational chemistry and drug discovery.

Main Methods:

• Utilized a dataset of 7912 chemicals with experimentally determined pKa values from DataWarrior.
• Employed KNIME for chemical structure curation and standardization, generating molecular descriptors using PaDEL.
• Developed and compared three machine learning models: support vector machines with k-nearest neighbors (SVM-kNN), extreme gradient boosting (XGB), and deep neural networks (DNN).

Main Results:

• All three developed models achieved comparable performance, with a root-mean-squared error (RMSE) around 1.5 and a coefficient of determination (R²) around 0.80.
• Benchmarking against commercial pKa predictors (ACD/Labs, ChemAxon) showed that the developed open-source models performed favorably.

Conclusions:

• Successfully developed multiple QSAR models for predicting chemical pKa values using publicly available data.
• The models are provided as free and open-source software on GitHub, offering a valuable resource for the scientific community.