Data-Driven Point\nAdjusted Jouyban-Acree-Artificial\nNeural Network Hybrid Model for Predicting Solubility of Active Pharmaceutical\nIngredients in Binary Solvent Mixtures

Solubility data of active pharmaceutical ingredients\n(APIs) in\nbinary solvent mixtures are crucial for optimizing solid–liquid\nseparation processes, conducting early solvent screening, and ensuring\nsafety. This study presents data-driven models integrating Monte Carlo\noptimization algorithms, the Jouyban–Acree (JA) model, and\nartificial neural networks (ANN) to comprehensively predict API solubility\nin binary solvent mixtures. A comprehensive database comprising 71,888\ndata points was constructed, encompassing quantitative descriptors\nof the three-dimensional structures of solutes and binary solvent\nmolecules, as well as the molecular interaction energies between these\nsolvents. A hybrid model, Jouyban–Acree–ANN (JAANN),\nwas developed to predict solubility across various temperatures and\nsolvent compositions. This model demonstrated robust predictive performance,\nwith prediction errors generally below 10%. Additionally, we introduced\na Point Adjusted JAANN (PA-JAANN) model that integrates Monte Carlo\nsimulations to refine solubility predictions by calibrating a single\nexperimental data point. This calibration significantly enhances the\nmodel’s accuracy, achieving an average error reduction of over\n20% compared to the standard JAANN model. A comparative Direct Prediction-ANN\n(DP-ANN) model was also constructed, providing rapid solubility predictions\nwithout experimental data, though it had limitations in robustness.\nThe predictive abilities of these three models were thoroughly validated\nbased on experiments involving the mefenamic acid–2-butanol–heptane\nsystem. These models can be used for different predictive needs, offering\nflexible and reliable solubility predictions essential for optimizing\ncrystallization processes in pharmaceutical manufacturing.