J. Huuskonen et al., NEURAL-NETWORK MODELING FOR ESTIMATION OF THE AQUEOUS SOLUBILITY OF STRUCTURALLY RELATED DRUGS, Journal of pharmaceutical sciences, 86(4), 1997, pp. 450-454
The ability of neural network models to predict aqueous solubility wit
hin series of structurally related drugs was evaluated. Three sets of
compounds representing different drug classes (28 steroids, 31 barbitu
ric acid derivatives, and 24 heterocyclic reverse transcriptase inhibi
tors) were studied. Topological descriptors (connectivity indices, kap
pa indices, and electrotopological state indices) were used to link th
e structures of compounds with their aqueous solubility. Separate mode
ls were built for each class of drugs using back-propagation neural ne
tworks with one hidden layer and five topological indices as input par
ameters. The effect of network size and training time on the predictio
n ability of the network was studied by the leave-one-out (LOO) proced
ure. 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 ind
ependent networks) were used. The cross-validated squared correlation
coefficient (Q(2)) for 10 averaged predictions was 0,796 in the case o
f the steroid set. The corresponding standard error of prediction (SDE
P) was 0.288 log units. For the barbiturates, Q(2) 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 n
etworks can produce useful models of the aqueous solubility of a conge
neric set of compounds, even with simple structural parameters.