Prediction of aqueous solubility of organic compounds by topological descriptors

Two quantitative models for the prediction of aqueous solubility of 1293 organic compounds were generated by a Multilinear Regression (MLR) analysis, and a Backpropagation (BPG) neural network. The molecules were represented by 18 topological descriptors. The physicochemical relationship between solubility and the descriptors for different individual classes of monofunctional group compounds such as hydrocarbons, ethers, halocarbons, alcohols, aldehydes and ketones, acids, esters, and amines was investigated. The 1293 compounds were divided into a training set of 741 compounds and a test set of 552 compounds based on a Kohonen's self-organizing neural network map. The models obtained show a good predictive power: for the test set, a correlation coefficient of 0.97 and a standard deviation of 0.52 were achieved by the backpropagation neural network approach.