From QbD to Explainable AI: Interpretable Random Forest Surrogates for Design Space Understanding of Voriconazole–β-Cyclodextrin Inclusion Complexes

Background Voriconazole formulation development is often constrained by limited aqueous solubility and variable dissolution behavior. β-Cyclodextrin (β-CD) inclusion complexation prepared by solvent-free co-grinding is a practical solubility-enhancement strategy. A recent Quality-by-Design (QbD) study optimised this system using a central composite design (CCD); however, polynomial response surfaces can be difficult to interpret locally across the design space. Objective To perform an explainable artificial intelligence (XAI) secondary reanalysis of a published QbD CCD dataset for voriconazole–β-CD inclusion complexes, generating interpretable Random Forest (RF) surrogates for design-space understanding and comparing model behavior against published QbD checkpoints. Methods Factor–response data (13 CCD runs) were extracted exactly as reported for β-CD amount (A, mg) and grinding time (B, min) with responses solubility (Y1, mg/mL) and cumulative drug release (Y2, %CDR). Two RF regression surrogates (RF–Y1 and RF–Y2) were trained and evaluated by leave-one-out cross-validation (LOOCV). Published checkpoints were used for benchmarking against QbD predictions. Explainability was implemented using TreeSHAP, permutation feature importance (PFI), partial dependence/ICE plots, and LIME. RF-based response surfaces and a multi-response desirability map were generated to identify high-performance regions. Results LOOCV indicated modest predictive performance (Y1: R²=0.1629, MAE = 11.2217, RMSE = 14.4720; Y2: R²=0.2208, MAE = 12.3883, RMSE = 15.5143). RF design-space mapping indicated increasing Y1 and Y2 with higher A and B, with a broad high-response region. The RF desirability optimum occurred at A = 544.99 mg and B = 26.84 min with predicted Y1 = 66.09 mg/mL, Y2 = 89.08%, and desirability = 0.887. At the published high-performance checkpoint (A = 600 mg, B = 30 min), RF predictions closely matched the experimental results (Y1 ≈ 66.09 vs 65.86 mg/mL; Y2 ≈ 89.08 vs 85.93%), whereas the QbD polynomial overpredicted, especially for Y2. SHAP global importance suggested A dominated Y1 (mean |SHAP|: A = 7.37; B = 3.78), while Y2 depended on both factors (A = 7.48; B = 7.82); PFI supported strong influence of A (ΔMAE: Y1 A ≈ 11.03, B ≈ 5.42; Y2 A ≈ 12.31, B ≈ 5.97). Conclusion Explainable ML did not replace QbD; it augmented a published QbD dataset with transparent, multi-view interpretability and an alternative design-space depiction. RF + XAI triangulated factor priority (carrier-driven solubility; joint carrier–process control of release), highlighted plateau-like high-performance regions, and provided calibration-friendly predictions at the optimised condition. This workflow offers a practical template for integrating explainable AI into formulation-oriented QbD analyses.