Rational Coformer Screening of Multicomponent Crystals of Abiraterone Acetate Using Machine Learning (ML) and Multicomponent Hydrogen Bond Propensity (MCHBP): A Comprehensive Structural and Computational Investigation

Abstract

Multicomponent crystalline systems offer a viable approach to improving the physicochemical and pharmacokinetic properties of drugs, thereby enabling the rational optimization of solid forms through coformer identification. The conventional process relies on empirical knowledge and chemical intuition. To overcome traditional challenges, we developed a novel predictive framework that combines machine learning (ML) with molecular complementarity hydrogen-bond propensity (MCHBP) for efficient screening of coformers. The effectiveness of the method was confirmed in a proof-of-concept study of the pharmaceutical multicomponent system for abiraterone acetate (AA). The success of this process lies in the generation of high-quality three-dimensional molecular descriptors derived from accurate structures. The novelty of this study is that it uniquely obtains high-quality 3D descriptors using optimized structures with the MMFF94 force field. For each molecule, 1,826 three-dimensional Mordred descriptors were generated with RDKit software. We used two types of descriptor fusion strategies: concatenation (M1) and summation (M2). The performance of the ten machine learning models was evaluated; gradient boosting performed well among the other nine tested models. The predicted API-coformer pairs were ranked in descending order of their probabilities. However, the predicted probabilities of the top 20 candidates were very close, making further prioritization difficult. To refine the selection, we used the MCHBP ranking, which enhanced the matching of hydrogen-bond networks. Following the MCHBP analysis, we selected the top eight unreported coformers for experimental investigation. From these, seven new multicomponent phases were identified, and six multicomponent crystals of AA were confirmed through SCXRD: AA−2HBA, AA−23DHBA, and AA−24DHBA as cocrystals and AA−25DHBA, AA−26DHBA, and AA−246THBA·H₂O as salts. The type and nature of intermolecular interactions and their contributions were analyzed by Hirshfeld surface and fingerprint analyses. The electronic structure, charge transfer, and hydrogen bonding were analyzed by density functional theory at the DFT/B3LYP/6-311++G(d,p) level. The cocrystal and salt forms were differentiated using the quantum theory of atoms in molecules (QTAIM) and natural bonding orbital (NBO) analysis. Noncovalent interactions were analyzed by noncovalent interaction (NCI) and reduced density gradient (RDG) graphical user interfaces.