ABSTRACT
An attempt was made to derive structure–activity relationship (SAR) models allowing the prediction of the mosquito larvicidal activity of benzoylphenylureas (BPUs) known to potentially inhibit the chitin synthesis of insects during their development. Activity values (active/inactive) on Aedes aegypti larvae for about 200 BPUs were obtained under the same experimental conditions. 252Chemicals were described by means of autocorrelation vectors encoding lipophilicity, molar refractivity, H-bonding acceptor ability, and H-bonding donor ability. The data set was randomly split into learning sets and external test sets of 80%/20%, respectively. Feature selection procedures were used to optimally reduce the number of descriptors. A three-layer perceptron was used as statistical tool. The performances of the models were evaluated through the analysis of the prediction results obtained on the different training sets and external test sets. Experimental results on larvae of Ae. aegypti obtained on the few commercialized BPUs and, for two published series of 32 and 73 structurally diverse BPUs, were also used to evaluate and select the most interesting configurations. Two models, presenting 91% of good predictions on the whole data set of BPUs, were selected. They included autocorrelation descriptors but also a descriptor encoding the presence of fluorine atoms in the ortho-position of the benzene ring linked to the carbonyl group of the BPUs. Interestingly, most of the autocorrelation descriptors were of higher order and half of them encoded lipophilicity. Both models are of interest to rationalize the discovery of new BPUs active on larvae of Ae. aegypti.
