Prediction of solid state properties of co-crystals using artificial neural network modelling

peer-reviewed The full text of this article will not be available in ULIR until the embargo expires on the 07/11/2018 Using Artificial Neural Networks (ANNs), four distinct models have been developed for the prediction of solid-state properties of cocrystals: melting point, lattice energy, and crystal density. The models use three input parameters for the pure model compound (MC) and three for the pure coformer. In addition, as input parameter the model uses the pKa difference between the MC and the coformer, and a 1:1 MC–conformer binding energy as calculated by a force field method. Notably the models require no data for the actual cocrystals. In total, 61 CCs (two-component molecular cocrystals) were used to construct the models, and melting temperatures and crystal densities were extracted from the literature for four MCs: caffeine, theophylline, nicotinamide and isonicotinamide. The data set includes 14 caffein cocrystals, 9 theophylline cocrystals, 9 nicotinamide cocrystals and 29 isonicotinamide cocrystals. The model–I is trained using known cocrystal melting temperatures, lattice energies and crystal densities, to predict all three solid–state properties simultaneously. The average relative deviation for the training set is 2.49%, 6.21% and 1.88% for the melting temperature, lattice energy and crystal density, respectively, and correspondingly 6.26%, 4.58% and 0.99% for the valdation set. Model–II, model–III and model–IV were built using the same input neurons as in model–I, for separate prediction of each respective output solid–state property. For these models the average relative deviation for the traning sets becomes 1.93% for the melting temperature model-II, 1.29% for the lattice energy model-III and 1.03% for the crystal density model-IV, and correspondingly 2.23%, 2.40% and 1.77% for the respective validation sets