Computational machine learning estimation of digitoxin solubility in supercritical solvent at different temperatures...

Area of Science:

• Computational chemistry
• Chemical engineering
• Machine learning applications

Background:

• Supercritical fluid technology is crucial for drug solubility.
• Accurate prediction of drug solubility in supercritical solvents is essential for process optimization.
• Digitoxin solubility in supercritical CO₂ presents a complex modeling challenge.

Purpose of the Study:

• To develop and validate ensemble machine learning models for predicting digitoxin solubility and solvent density in supercritical CO₂.
• To explore the efficacy of AdaBoost, Bayesian Ridge Regression, Gaussian process regression, and K-nearest neighbors algorithms.
• To optimize model performance through hyper-parameter tuning using the Sailfish Optimizer.

Main Methods:

• Utilized ensemble methods, specifically AdaBoost, to combine predictions from Bayesian Ridge Regression, Gaussian process regression, and K-nearest neighbors.
• Employed the Sailfish Optimizer for hyper-parameter tuning to enhance predictive accuracy.
• Trained and validated models using temperature and pressure as input parameters to predict digitoxin solubility and solvent density.

Main Results:

• The AdaBoost combined with Gaussian process regression (ADA-GPR) model achieved the lowest Average Absolute Relative Deviation (AARD%) for both solubility (7.74%) and solvent density (2.76%).
• Demonstrated superior performance of ensemble methods over individual models in predicting supercritical fluid properties.
• Validated the effectiveness of hyper-parameter optimization in improving the accuracy of machine learning models.

Conclusions:

• Ensemble methods, particularly ADA-GPR, are highly effective for accurately modeling digitoxin solubility and solvent density in supercritical CO₂.
• Hyper-parameter tuning using the Sailfish Optimizer significantly enhances the predictive power of these models.
• The study confirms the utility of computational tools for predicting complex chemical properties in supercritical systems.