Development of several machine learning based models for determination of small molecule pharmaceutical solubility in binary solvents at different temperatures.

Analysis of small-molecule drug solubility in binary solvents at different temperatures was carried out via several machine learning models and integration of models to optimize. We investigated the solubility of rivaroxaban in both dichloromethane and a variety of primary alcohols at various temperatures and concentrations of solvents to understand its behavior in mixed solvents. Given the complex, non-linear patterns in solubility behavior, three advanced regression approaches were utilized: Polynomial Curve Fitting, a Bayesian-based Neural Network (BNN), and the Neural Oblivious Decision Ensemble (NODE) method. To optimize model performance, hyperparameters were fine-tuned using the Stochastic Fractal Search (SFS) algorithm. Among the tested models, BNN obtained the best precision for fitting, with a test R² of 0.9926 and a MSE of 3.07 × 10⁻⁸, proving outstanding accuracy in fitting the rivaroxaban data. The NODE model followed BNN, showing a test R² of 0.9413 and the lowest MAPE of 0.1835. The Polynomial model yielded a lower test R² of 0.8200 and higher error rates, indicating its limitations in unravelling the underlying relationships for the solubility variations. This study shows that advanced machine learning models, particularly BNN and NODE, can predict pharmaceutical solubility and improve crystallization process design and optimization.