Prediction of aqueous solubility of organic compounds from molecular structure

Multiple linear regression (MLR) and computational neural networks (CNN) are utilized to develop mathematical models to relate the structures of a diverse set of 332 organic compounds to their aqueous solubilities. Topological, geometric, and electronic descriptors are used to numerically represent structural features of the data set compounds. Genetic algorithm and simulated annealing routines, in conjunction with MLR and CNN, are used to select subsets of descriptors that accurately relate to aqueous solubility. Nonlinear models with nine calculated structural descriptors are developed that have a training set root-mean-square error of 0.394 log units for compounds which span a -log(molarity) range from -2 to +12 log units.