Crystallization Solvent Design Methodology for Empagliflozin Based on a Quantitative Prediction Model of Crystal Morphology Using Molecular Dynamics and Generative Adversarial Network

Empagliflozin is a sodium-glucose cotransporter 2 (SGLT2) inhibitor that has received widespread attention in the treatment of diabetes in recent years, and it is usually prepared by solution crystallization. How to choose the appropriate crystallization solvent to obtain empagliflozin crystals with good crystal morphology and high purity is currently a key issue that urgently needs to be addressed. This paper proposes a computer-aided design/screening method for empagliflozin crystallization solvents. First, Hirshfeld surface analysis (HSA) is used to analyze the intermolecular interactions in empagliflozin crystals. Subsequently, the modified attachment energy (MAE) model is used to explore the influence of solvation on the crystal morphology of empagliflozin, and a quantitative prediction model for crystal morphology is established based on the generative adversarial network (GAN) and extreme gradient boosting (XGBoost) model. After data enhancement by GAN, the R2 of the model performance can be increased from 0.4894 to 0.9454, and MAE, RMSE, and MAPE can be reduced to 0.1775, 0.2281, and 5.21% respectively. Next, the established quantitative prediction model is coupled with the computer-aided molecular design (CAMD) model, which is solved by the decomposition solution strategy. Finally, the crystallization solvents that meets all requirements are obtained.