Advanced AI modeling and optimization for determination of pharmaceutical solubility in supercritical processing for production of nanosized drug particles

Abstract

This study presents a comparative analysis of three different models, namely Deep Neural Network (DNN), Quantile Regression (QR), and K-Nearest Neighbors (KNN) for the prediction of SC-CO2 density and solubility of rivaroxaban. The models were tuned using the Fireworks Algorithm (FWA) to optimize their hyperparameters. The dataset used in this analysis consisted of temperature (T), pressure (P), SC-CO2 density, and mole fractions of rivaroxaban. The results indicate that the DNN model exhibited outstanding performance in both prediction tasks. For the prediction of SC-CO2 density, the DNN model achieved an impressive R2 score of 0.99667, with a mean absolute error (MAE) of 7.61812E+00. Similarly, for the prediction of mole fractions of rivaroxaban, the DNN model achieved an excellent R2 score of 0.99831, with a very low MAE of 2.49870E-02. The KNN model also demonstrated good performance, with R2 scores of 0.9799 and 0.98029 for SC-CO2 density and mole fractions, respectively. However, it exhibited slightly higher MAE values compared to the DNN model. On the other hand, the QR model showed relatively lower accuracy in both prediction tasks, with R2 scores of 0.90873 and 0.87362 for SC-CO2 density and mole fractions, respectively. The QR model had higher MAE values, indicating larger average deviations from the true values. Overall, the DNN model outperformed both the KNN and QR models in predicting SC-CO2 density and mole fractions of rivaroxaban. These findings highlight the effectiveness of DNN models in accurately modeling and predicting complex chemical properties.

Keywords

• Nano drug
• Solubility
• Green processing
• Deep neural network
• K-nearest neighbors