Construction of Local Approximators and Prediction Accuracy Estimation for Prediction of Physicochemical Properties of Molecules

Chemical informatics focuses on storing, indexing, searching, retrieving, and applying information about chemical compounds. Chemical informatics helps chemists investigate new problems and organize and analyze scientific data to develop novel compounds, materials, and processes through the application of information technology. Mathematical modeling and modern machine learning algorithms are the main research tools for approximation of physicochemical properties of molecules. Distinctive features of such problems are the large dimensionality of the input space, significant noise in the experimental data, heterogeneity of the sample. These features require additional analysis and development of new approaches. The aim of this work is to develop a system of interrelated methods (develop new methods and modify existing ones) that would allow to consider features of the problem and provide the needs described above. The main results include the following. A new conformal metric was proposed for construction of conformal predictors for the case of classification problem. A method of the input space partitioning with construction of regression models was used for prediction of physicochemical properties of molecules, a regression method based on classification algorithms was developed. Described methods were applied to the practical problem and numerical simulations were carried out on the melting point dataset and lipophilicity dataset.