A neural network SAR model for allergic contact dermatitis

Allergic contact dermatitis results for 259 compounds were used for deriving a qualitative structure-activity relationship (SAR) model. Chemicals were described by means of one physicochemical descriptor, one topological index, and twelve structural alerts (i.e., a value of 1 indicated the presence of the structural feature in a molecule, 0 the absence). A three-layer feedforward neural network trained by the back-propagation algorithm was used as statistical engine. The comparison of the simulation performances of the obtained model to those produced by a classical linear discriminant analysis clearly revealed the usefulness of the nonlinear methods of modeling shin sensitization.