Predicting the solubility of drugs in supercritical carbon dioxide using machine learning and atomic contribution

Abstract

The pharmaceutical sector is aware of supercritical CO₂ (SC-CO₂) as a possible replacement for problematic organic solvents. Using a novel artificial intelligence (AI) strategy to predict drug solubility using the SC-CO₂ system mathematically has been deemed an intriguing approach. In this work, the atomic contribution (AC) method and machine learning (ML) models are combined to develop hybrid machine learning models to compute the solubility of several drugs, including anticoagulants, anti-cancers, calcium channel blockers, immunosuppressives, antihistamines, and others. The novelty of the approach lies in using the AC concept to capture molecular details at the atomic level. This enables the model to account for the specific contributions of individual atoms and to provide more precise input features for machine learning. The integration of these molecular insights with ML techniques results in significantly improved predictive performance over traditional ML methods. Throughout the modeling procedure, temperature, pressure, the density of SC-CO₂, and the effect of constituent atoms of the drugs are the input variables, while the solubility of drugs is the output. This study looks into predicting the solubility of these drugs in SC-CO₂ using the least square support vector machine (LSSVM) with radial basis function kernel (RBF) and multilayer perceptron artificial neural network (MLPANN). These models were developed using a database including 2358 experimental solubility data points from 86 solid drugs. The solubility of solid drugs in supercritical CO₂ spans a remarkably wide range in this study, from as high as 3.9 × 10^-2 to as low as 1 × 10^-7. The results demonstrated that this innovative approach could estimate solid drug solubility in SC-CO₂ with AARD% and R² values of 7.20 and 0.99, respectively, under different pressure and temperature conditions. The ability of the models to capture a wide range of solubilities in SC-CO₂ showcases their effectiveness in dealing with both highly and poorly soluble compounds. The developed models, considering their global prediction, accuracy, and being user-friendly, are the best options to be used by researchers for incorporating into software for enabling more efficient design of supercritical extraction processes and reducing the need for trial-and-error experimentation in manufacturing.