Advanced feature analysis for enhancing cocrystal prediction

Abstract

The design of novel pharmaceutical crystal forms, including molecular salts and cocrystals, has gained significant attention from pharmaceutical companies due to their ability to modulate key physicochemical and biopharmaceutical properties. The selection of appropriate coformers for cocrystallization, however, remains a challenge, typically relying on labor-intensive trial-and-error methods. This study introduces FeatureMaster, a tool designed to evaluate the representativeness of training sets relative to test sets, thereby enhancing the reliability of machine learning models in predicting cocrystallization outcomes. We employed four key algorithms — feature overlap, quartiles, Cohen's D, and p-value analysis — to a priori assess the predictive accuracy. The efficacy of these methods was evaluated on two systems: piracetam (PRC) and pyridoxine (PN). The test set data were collected from in-house experiments: the PRC and PN test sets were experimentally created with a series of coformers (20 for PRC and 14 for PN) using different synthetic techniques. The experimental tests lead to the formation of 3 new cocrystals for PRC (with quercetin, 2-ketoglutaric acid, and malic acid) and 7 new molecular salts for PN (with 2-ketoglutaric acid, pimelic acid, cinnamic acid, gallic acid, N-acetylcysteine, and caffeic acid). Training sets were collected from literature and features calculated using Hansen Solubility Parameters (HSP), Hydrogen Bond Energy (HBE), Molecular Complementarity (MC), and Quantitative Structure-Activity Relationship (QSAR) methods. Models were developed using the Random Forest algorithm, known for its robustness in handling complex datasets. Our results demonstrate that statistical analyses using overlap, Cohen's D and p-values are fundamental for improving the prediction and for providing a priori insights into the model's reliability. This approach reduces the experimental tests and resource consumption in the cocrystal screening process, offering a promising strategy for future pharmaceutical development.