Статистичний аналіз спеціалізованого набору даних для прогнозування кислотно-основних властивостей і ліпофільності фторованих органічних сполук

Abstract

Fluorine is a key element in pharmaceutical and organic chemistry due to its unique electronic and spatial properties, which significantly influence the physicochemical characteristics of organic compounds. The introduction of fluorine atoms or fluorinated groups alters acid–base properties, lipophilicity, and conformational stability of molecules, complicating their accurate prediction. In this work, we analyzed a specialized dataset of experimental pKa and logP values for 183 saturated fluorinated organic compounds, including primary and secondary amines, carboxylic acids, and their model derivatives, in order to evaluate the patterns of fluorine influence on physicochemical properties and to assess the dataset’s suitability for high-precision modeling. The dataset covers a broad range of structural classes – monocyclic, bicyclic, and acyclic systems – and diverse fluorinated substituents, enabling evaluation of the effects of the number and spatial arrangement of fluorine atoms on acid–base properties and lipophilicity. A comprehensive data preprocessing workflow was performed, including outlier removal using the Shapiro– Wilk test, z-score criterion, and interquartile range method, as well as stratified splitting into training (80 %) and test (20 %) subsets with control for structural similarity based on the Tanimoto coefficient (≥ 90 %). Analysis of the target variables revealed a bimodal distribution of pKa for amines and carboxylic acids and a more uniform distribution of logP. A systematic decrease in pKa was observed for secondary amines with increasing number of fluorine atoms and their proximity to the deprotonation center, whereas the effects of fluorination on carboxylic acids and logP exhibited complex nonlinear patterns, including positional and cooperative effects. Application of graph neural networks to this dataset demonstrated high predictive accuracy for both pKa and logP values. These results confirm the suitability of the dataset for developing high-precision predictive models of acid–base properties and lipophilicity of fluorinated organic compounds, allowing consideration of electronic, spatial, and conformational effects of fluorination and providing prospects for its application in pharmaceutical research and fundamental chemistry.