Universal data-driven models to estimate the solubility of anti-cancer drugs in supercritical carbon dioxide: Correlation development and machine learning modeling

The current study aims at modeling the solubility of anti-cancer agents in supercritical carbon dioxide (SC-CO2). An extensive databank, including 893 measured samples for 33 anti-cancer agents were collected from the literature, covering extensive ranges of operating conditions. Eight density-based empirical models were firstly employed to correlate the collected data. After adjusting their constant coefficients, four of them provided satisfactory estimations, with total average absolute relative errors (AAREs) below 10 %. A novel six-parameter empirical correlation was also proposed, with input factors optimized based on the Pearson coefficient analysis. This correlation produced satisfactory results for the analyzed drugs, achieving a total AARE of 7.71 %. Afterward, a generalized and unified model was built using the intelligent method of gaussian process regression (GPR). For the testing data, this model showed excellent results with AARE and R2 values of 2.90 % and 99.87 %, respectively. Furthermore, its estimations for all anti-cancer agents outperformed the empirical correlations significantly. Both empirical and intelligent models accurately described the physical behavior of anti-cancer agents’ solubility in SC-CO2 under various conditions. Subsequently, the most effective factors on the performances of the models were recognized through a sensitivity analysis.