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

Lipophilicity and acidity/basicity are fundamental physical properties that profoundly affect the compound's pharmacological activity, bioavailability, metabolism, and toxicity. Predicting lipophilicity, measured by $$ (1-octanol-water distribution coefficient logarithm), and acidity/basicity, measured by $$ (negative of acid ionization constant logarithm), is essential for early drug discovery success. However, the limited availability of experimental data and poor accuracy of standard $$ and $$ assessment methods for saturated fluorine-containing derivatives pose a significant challenge to achieving satisfactory results for this compound class. To overcome this challenge, we compiled a unique dataset of saturated fluorinated and corresponding non-fluorinated derivatives with $$ and $$ experimental values. Aiming to create an optimal approach to acidity/basicity and lipophilicity prediction, we evaluated, trained from scratch, or fine-tuned more than 40 machine learning models, including linear, tree-based, and neural networks. The study was supplemented with a substructure mask explanation (SME), which confirmed the critical role of the fluorinated substituents on both physicochemical properties studied and testified to the consistency of the developed models. The results were open-sourced as a GitHub repository, pip, conda packages, and a KNIME node, allowing the public to perform the targeted molecular design of the proposed class of compounds.