Employment of artificial intelligence approach for optimizing the solubility of drug in the supercritical CO2 system

Abstract

This paper investigates the solubility behavior of digitoxin in supercritical carbon dioxide (CO2) through a comprehensive analysis employing ensemble learning techniques and various regression models. The dataset consists of temperature and pressure as input variables, with solvent density and digitoxin solubility as output variables. Utilizing bagging as the ensemble method, Gaussian process regression (GPR), Bayesian Ridge Regression (BRR), Orthogonal Matching Pursuit (OMP), and Polynomial Regression (PR) were employed as regression models. Hyper-parameter tuning was achieved through gradient-based optimization. Results revealed that Polynomial Regression (PR) emerges as the most effective model for predicting both solvent density and digitoxin solubility. For solubility prediction, BAG-PR yields an R2 score of 0.98527, with MSE and MAE of 1.4290E-03 and 3.40547E-02, respectively. Concerning solvent density, BAG-PR achieves an outstanding R2 of 0.99766, along with MSE and MAE of 7.4759E+01 and 7.33964E+00, respectively. These results show that ensemble learning, and polynomial regression can accurately predict solubility and solvent density, revealing digitoxin's behavior in supercritical CO2.

Keywords

• Drug solubility
• Bagging regression
• Polynomial regression
• Gradient-based optimization