Smart formulation: extending AI-based stability prediction to compounded liquid dosage forms using KNIME

Reliable prediction of beyond-use dates (BUDs) for compounded oral liquid formulations remains challenging because stability depends on complex interactions between active pharmaceutical ingredient (API) properties, formulation composition, packaging, and storage conditions. This issue is increasingly relevant in the evolving European regulatory landscape, where certain magistral preparations traditionally compounded extemporaneously may, under defined conditions, be prepared in advance, reinforcing the need for scientifically justified BUDs. Smart Formulation Module II was developed as a data-driven decision-support tool combining BUD prediction with formulation optimization. Within this context, we propose the "Prediction-Optimization-Production" (POP) framework linking predictive modeling, pharmaceutical formulation optimization, and hospital pharmaceutical production. The model was developed from a curated Advanced Molecular Formulation Database (AMF-DB) comprising 536 formulation-condition records for 137 APIs derived from 180 documentary sources. Molecular and physicochemical API descriptors were combined with formulation composition, structured vehicles, packaging, and storage conditions using a Tree Ensemble Regression model implemented in KNIME. External evaluation comprised 230 formulation-condition records for 101 APIs, all involving API-vehicle combinations absent from the learner dataset, and an additional evaluation included 74 records for 26 APIs not represented in the learner dataset. External predictive performance was moderate (R² = 0.56, RMSE = 20.1 days, MAE = 13.4 days) and was higher in formulation domains well represented in the learner dataset. For the 26 unseen APIs, molecular similarity to learner APIs was generally low (median nearest-neighbor Tanimoto similarity = 0.240; 92.3% <0.50), while experimentally demonstrated stability durations and application-predicted BUDs showed comparable median values of 90 days (IQR 60-90) and 84 days (IQR 75.5-88), respectively. Pharmaceutical utility was illustrated during an aprepitant shortage by comparing 17 structured vehicles according to predicted BUD and formulation characteristics relevant to patient suitability. Predicted BUDs ranged from 85 to 129 days, while marked differences in osmolality, sucrose, and paraben content influenced formulation selection. SyrSpend® SF pH 4 combined a predicted BUD of 86 ± 6 days with very low osmolality (<50 mOsm/kg) and absence of sucrose and parabens, whereas OraSweet® SF provided the highest predicted BUD (129 ± 12 days) but had substantially higher osmolality (∼2150 mOsm/kg). Smart Formulation Module II provides a multicriteria approach to BUD prediction and formulation optimization for compounded oral liquids within the POP framework. Predictions are intended to support pharmaceutical decision-making and require experimental confirmation, particularly for poorly represented APIs and formulation contexts. The application is available for open evaluation through a dedicated web link provided upon request.