Some findings relevant to the mechanistic interpretation in the case of predictive models for carcinogenicity based on the counter propagation artificial neural network.

The goal of the study was to contribute to a better mechanistic understanding of so-called "general" QSAR models for non-congeneric chemicals based on the counter propagation artificial neural network (CP ANN). Possible mechanisms of action was proofed using the Toxtree expert system based on structural alerts (SAs) for carcinogenicity. We have illustrated how statistically selected MDL descriptors, which refer to topological characteristics as well as to polarizability and charge distribution related to reactivity, are correlated with particular chemical classes (containing carcinogenic SA) with the recognized mechanistic link to the carcinogenic activity and consequently with the carcinogenic potency. Mechanistic insight in CP ANN models was demonstrated using an inherent mapping technique (i.e. Kohonen maps).