Abstract
Most active pharmaceutical ingredients (APIs) are poorly water-soluble, limiting their pharmaceutical performance. Co-amorphous (CA) systems offer a promising formulation strategy, yet CA prediction remains constrained by limited labeled data and insufficient molecular representation. Here, a multimodal transfer-learning framework was developed by transferring molecular graph representations from a larger co-crystal dataset and integrating graph, DFT and SMILES information for small-sample CA prediction. The model achieved an AUC-ROC of 0.924 and an MCC of 0.771. Compared with existing AI-assisted CA prediction approaches, this framework combines cross-domain knowledge transfer with complementary molecular representations to address the limited-data challenge. Using afatinib (AF) as a model API, four AF-based CA systems were prepared by ball milling. Solid-state characterization supported amorphous CA formation, and dimer-level quantum-chemical calculations supported the plausibility of favorable local AF-co-former interaction motifs. The CA systems showed pH-dependent apparent solubility advantages and greater dissolution extents than Crystal AF. Under dry storage conditions, the CA systems maintained their amorphous state and morphology. CCK-8 assays revealed differences in HCC827 cell viability inhibition among the tested samples. Overall, this study establishes an AI-assisted approach for CA co-former screening and experimental prioritization, providing a practical strategy for formulation development of poorly soluble drugs.