Інформаційна технологія оцінювання LogP та pKa фторвмісних насичених похідних на основі моделей машинного навчання

Abstract

The physicochemical properties of organic compounds, in particular lipophilicity (LogP) and acid–base characteristics (pKa), play a crucial role in drug discovery and medicinal chemistry. Accurate prediction of these parameters is essential for assessing the bioavailability, distribution, and toxicological profiles of potential drug candidates. However, despite the availability of numerous computational methods, their predictive efficiency significantly decreases when applied to structurally specific and narrow classes of compounds, particularly saturated fluorine-containing derivatives. Fluoroorganic fragments can substantially modify the electronic, spatial, and energetic characteristics of molecules, yet universal machine learning models built on heterogeneous datasets do not fully account for these peculiarities. Therefore, the developed information technology for evaluating LogP and pKa of fluorine-containing saturated derivatives was based on a specially constructed representative open-access dataset comprising 183 saturated fluorinated and nonfluorinated compounds, which ensured the highest predictive validity. For acid–base property prediction, the best performance was achieved using a Graph Isomorphism Network (R² ≈ 0.989), whereas for LogP prediction the optimal approach was a Graph Convolutional Network with AttentiveFP descriptors (R² ≈ 0.955). The proposed information technology has been implemented as a web-based system that includes a chemical structure input module (SMILES, SDF/MOL, interactive editor), an automated descriptor calculation block, integrated neural network models, and a results visualization subsystem. In addition to the predicted pKa and LogP values, the system displays supplementary physicochemical characteristics such as molecular weight, polar surface area, dipole moment, and topological indices, thereby enhancing interpretability of the results. The functional versatility of the solution is ensured by its support for singlestructure input, batch processing of compound libraries, and interactive molecular design. This enables its use in virtual screening, preliminary analysis, experimental data verification, educational purposes, and scientific research. The developed web application does not require installation of specialized software or knowledge of programming languages, making it accessible to a broad audience of chemists, pharmacists, and researchers in related fields.