Rational design and preparation of sorafenib Co-amorphous systems based on explainable Machine learning and quantum chemistry

Co-amorphous (CA) systems have emerged as an effective strategy for improving the dissolution of poorly water-soluble drugs. However, the screening of co-former for CA systems still relies largely on semi-empirical criteria, providing limited mechanistic insight into co-former selection and CA formation. Herein, an interpretable framework was developed for the rational design of sorafenib-based CA systems by integrating machine learning with quantum chemical analysis. Using a curated dataset and hybrid molecular descriptors, eight classification models were constructed, among which LightGBM achieved the best predictive performance. Four sorafenib CA systems were subsequently prepared and experimentally validated. Solid-state characterization supported the formation of single-phase amorphous CA systems, while FTIR and MEPS analyses suggested the involvement of intermolecular hydrogen bonding and potential electrostatic complementarity between SOR and the selected co-formers. Compared with crystalline sorafenib, the CA systems accelerated initial dissolution in both pH 1.2 HCl medium (0.5% SDS, w/v) and pH 6.8 PBS (0.5% SDS, w/v), and showed medium-dependent apparent solubility enhancement, with sorafenib-indomethacin CA reaching 57.60 ± 0.34 μg/mL after 24 h and 102.69 ± 6.48 μg/mL after 72 h in pH 6.8 PBS, corresponding to 4.1-fold and 7.7-fold increase, respectively. Sorafenib-indomethacin and sorafenib-L-phenylalanine CA systems showed relatively good resistance to recrystallization under the tested storage conditions. SHAP analysis identified descriptor-level contributions to CA formation predictions. This work provides a model-assisted and interpretable strategy for co-former prioritization and rational design of SOR-based CA systems.