PREDICTION OF REACTION CONDITIONS BY DEEP LEARNING TECHNIQUES

Subject of Research. The paper presents a study of prediction method for various properties of reactions, such as the type of reaction, suitable groups of solvents, catalysts for the reaction. Molecular fingerprint differences between products and reagents were calculated using the RdKit chemical library as a representation of the reactions. Molecular fingerprints are widely used to predict various properties of molecules. Knowledge of the reaction conditions is essential for successful planning of retrosynthesis. Chemical informatics methods can effectively find the relationship between reaction reagents and the necessary conditions for the reaction. At this, the costs of time and resources spent on the determination of the necessary set of conditions for the reaction are reduced. Prediction of solvent groups can significantly improve the quality of models and the applicability of approaches. Method. LightGBM and a neural network with Deep Feature Selection were taken as machine learning models. The results were evaluated with the F1metric. For the models training and evaluation, the data was broken down into chemically dissimilar parts. Bayesian optimization was used to optimize the searching of parameters. Main Results. Experiments were carried out to predict the reaction type, catalysts and solvent groups for the reaction. The obtained results show that the MLP type of reaction can be predicted equal to 0.99, a MLP catalyst equal to 0.7, and MLP group of solvents equal to 0.68 with F1-metric based on the difference in molecular fingerprints between reagents and products of machine learning models. Significant quantity of catalysts and solvents are considered in the paper. Practical Relevance. Automated planning of retrosynthesis is one of the topical areas of research. During planning, a sequence of necessary reactions is drawn up. The considered method can be used for recommendation system development that can suggest a possible group of catalysts and solvents to a chemical specialist, and, thus, reduce the cost of resources and time to determine the necessary reaction conditions.