Neural network modeling for estimation of the aqueous solubility of structurally related drugs

The ability of neural network models to predict aqueous solubility within series of structurally related drugs was evaluated. Three sets of compounds representing different drug classes (28 steroids, 31 barbituric acid derivatives, and 24 heterocyclic reverse transcriptase inhibitors) were studied. Topological descriptors (connectivity indices, kappa indices, and electrotopological state indices) were used to link the structures of compounds with their aqueous solubility. Separate models were built for each class of drugs using back-propagation neural networks with one hidden layer and five topological indices as input parameters. The effect of network size and training time on the prediction ability of the network was studied by the leave-one-out (LOO) procedure. In all three compound groups a neural network structure of 5-3-1 was optimal. To avoid chance effects, artificial neural network (ANN) ensembles (i.e.; averaging neural network predictions over several independent networks) were used. The cross-validated squared correlation coefficient (Q2) for 10 averaged predictions was 0.796 in the case of the steroid set. The corresponding standard error of prediction (SDEP) was 0.288 log units. For the barbiturates, Q2 and SDEP were 0.856 and 0.383, respectively, and for the RT inhibitors, these parameters were 0.721 and 0.401, respectively. The results indicate that neural networks can produce useful models of the aqueous solubility of a congeneric set of compounds, even with simple structural parameters.