Filling the Gap in LogP and pKa Evaluation for Saturated Fluorine-Containing Derivatives With Machine Learning

This study addresses the challenge of predicting lipophilicity (logD) and acidity/basicity (pKa) for saturated fluorine-containing derivatives, which are crucial properties for drug discovery. The authors compiled a specialized dataset of fluorinated and non-fluorinated compounds and evaluated over 40 machine learning models, including linear, tree-based, and neural networks. A substructure mask explanation (SME) approach confirmed the importance of fluorine substitutions in these properties. The results were made publicly available via GitHub, pip, conda, and a KNIME node, allowing researchers to use these models for molecular design.

Takeaways:

1. The study focuses on predicting logD and pKa for fluorinated compounds, which significantly impact pharmacological activity, bioavailability, metabolism, and toxicity.
2. Standard prediction methods for these properties struggle with fluorine-containing molecules due to limited experimental data.
3. The authors compiled a dataset of fluorinated and non-fluorinated derivatives with experimental logD and pKa values.
4. More than 40 machine learning models, including linear, tree-based, and neural networks, were trained or fine-tuned for optimal prediction accuracy.
5. A substructure mask explanation (SME) technique validated the role of fluorinated groups in influencing these properties.
6. The models and datasets were open-sourced as a GitHub repository, pip and conda packages, and a KNIME node, making them accessible for further research and application.
7. The study was supported by Blackthorn AI Ltd., Enamine Ltd., and the Ministry of Education and Science of Ukraine.
8. Some authors are employees of Blackthorn AI Ltd. and Enamine Ltd., presenting potential conflicts of interest.