Data-driven models and comparison for correlation of pharmaceutical solubility in supercritical solvent based on pressure and temperature as inputs

Abstract

Data-driven models were employed for the solubility correlation, while the focus was on modeling the solubility of raloxifene drug and density of carbon dioxide based on temperature (T) and pressure (P) inputs. Three Machine Learning models, namely Multilayer Perceptron (MLP), Bayesian Ridge Regression (BRR), and LASSO regression, were employed and optimized using the MPHPT method for hyper-parameter tuning. The dataset consisted of experimental measurements of solubility (y), and CO2 density. For the prediction of density, the MLP model exhibited excellent performance with an R2 score of 0.99726, demonstrating a significant level of association between the anticipated and observed values. The mean squared error (MSE) was 9.8721E+01, the mean absolute percentage error (MAPE) was 1.78565E-02, and the maximum error was 1.86395E+01. The LASSO and BRR models achieved slightly lower accuracy, with R2 scores of 0.83317 and 0.83001, respectively. Regarding the solubility of the raloxifene drug, the MLP model demonstrated a strong predictive capability with an R2 score of 0.99343. The MSE was 3.0869E-02, MAPE was 4.02666E-02, and the maximum error was 3.01133E-01. The LASSO and BRR models also provided reasonable predictions, with R2 scores of 0.90955 and 0.8891, respectively. However, they exhibited higher MSEs and MAPEs compared to the MLP model.

Keywords

• Drug production
• Nanomedicine
• Modeling
• Drug design
• Supercritical processing
• Multilayer perceptron