Grey Wolf Optimizer–tuned support vector regression and SHAP analysis for multi-drug solubility prediction in supercritical carbon dioxide

Accurate solubility prediction of anticancer drugs in supercritical carbon dioxide (scCO 2 ) is indispensable for the rational design of green particle engineering processes for improved bioavailability. This work compiles an integrated experimental dataset (116 measurements) of four hematologic anticancer drugs (Busulfan, Dasatinib monohydrate, Decitabine, and Lenalidomide) across temperatures of 308–338 K and pressures of 120–400 bar. A solvent-state thermodynamic feature space was built through careful engineering of reduced-state parameters, cross-terms, drug-identity terms; 10 features out of the constructed ones were selected based on Random Forest importance. Four machine learning models including Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Support Vector Regression (SVR) and optimized SVR using Grey Wolf Optimizer (GWO–SVR) were built and compared under rigorous repeated K-fold cross-validation and hold-out testing procedures. The optimal model, GWO–SVR, yielded the best generalizability (test R 2 = 0.838, RMSE = 0.793, MAE = 0.542) over RF (test R 2 = 0.811), XGBoost (test R 2 = 0.795) and standard SVR (test R 2 = 0.817). According to the SHAP explainability analysis, the most prominent descriptor was found to be the reduced temperature (T red ) (mean |SHAP|= 0.87), followed by drug identity of Lenalidomide (0.57) and CO 2 density (0.55), depicting the dominating effects of thermodynamics and chemical compounds on solubility in scCO 2 . Sensitivity analysis for perturbations revealed that reduced temperature (T red , S = 40.8%) was the parameter having maximum influence, followed by carbon dioxide density (28.9%) and ln_P (20.9%), while out of eight continuous descriptors, seven parameters had influence greater than 10%, which shows the influence of thermodynamics on solubility. The work demonstrates that GWO–SVR combined with thermodynamically engineered descriptors is a powerful and explainable method for simultaneous multi-drug scCO 2 solubility prediction.