(136e) Prediction of Phase Transition Thermodynamics for Crystals of Pharmaceutical Compounds

Phase transition properties of pharmaceutical crystals are affecting their thermodynamic aqueous solubility, which is considered to be one of the major factors in determining the ultimate success or failure of drug compounds. Solubility is important because it restricts the maximal drug concentration in the human body upon taking the drug. In addition, solubility in organic solvents is a crucial property for both process and medicinal chemistry as well as for formulation design. Many computational models of aqueous solubility have been developed but the challenge of accurately predicting aqueous solubility for diverse molecules as a critical component of drug discovery and development remains in place.

Solubility prediction requires the knowledge of free energy of fusion (∆G_fus) and free energy of mixing (∆G_mix). Using equilibrium thermodynamics and chemical potentials, it is possible to predict ∆G_mix accurately. Unfortunately, there is no rigorous way to predict ∆G_fus, which translates into inaccuracies in predictive models of solubility. Free energy of fusion is often cannot be measured experimentally too.

To enable accurate ∆G_fus prediction we have combined accurate ab initio quantum-chemical calculation of crystals with conventional machine learning (ML) and Quantitative Structure Property Relationships (QSPR) approaches. Based on quantum-chemical calculations of all intermolecular interactions in the unit cell we construct the energy vector diagrams (EVD) of the crystal structure. The EVDs are used to encode intermolecular interactions, basic structural motif of the crystal, and the interactions between such motifs in the form of crystallographic fingerprints that are used as system descriptors in developing QSPR models.

Models have been developed for an aqueous solubility dataset consisting of approximately 40,000 experimental appoints for about 6500 organic molecules after thorough manual curation and standardization; this dataset was constructed from numerous publically available sources. We show that our QSPR models afford accurate prediction of ∆G_fus with less than 1 kcal/mol unsigned error of prediction. In addition to free energy of fusion, we have modeled additional important endpoints such as enthalpy of sublimation, enthalpy of vaporization, heat capacity in liquid and solid phases.