We present a novel computational tool which predicts the glass-forming ability of drug compounds solely from their molecular structure. Compounds which show solid-state limited aqueous solubility were selected, and their glass-forming ability was determined upon spray-drying, melt-quenching and mechanical activation. The solids produced were analyzed by differential scanning calorimetry (DSC) and powder X-ray diffraction. Compounds becoming at least partially amorphous on processing were classified as glass-formers, whereas those remaining crystalline regardless of the process method were classified as non-glass-forming compounds. A predictive model of the glass-forming ability, designed to separate between these two classes, was developed through the use of partial least-squares projection to latent structure discriminant analysis (PLS-DA) and calculated molecular descriptors. In total, ten of the 16 compounds were determined experimentally to be good glass-formers and the PLS-DA model correctly sorted 15 of the compounds using four molecular descriptors only. An external test set was predicted with an accuracy of 75%, and, hence, the PLS-DA model developed was shown to be applicable for the identification of compounds that have the potential to be designed as amorphous formulations. The model suggests that larger molecules with a low number of benzene rings, low level of molecular symmetry, branched carbon skeletons and electronegative atoms have the ability to form a glass. To conclude, we have developed a predictive, transparent and interpretable computational model for the identification of drug molecules capable of being glass-formers. The model allows an assessment of amorphization as a formulation strategy in the early drug development process, and can be applied before compound synthesis.
Read full abstract