Counter-Propagation Artificial Neural Network Models for Prediction of Carcinogenicity of Non-congeneric Chemicals for Regulatory Uses

Abstract

The evaluation of carcinogenic hazard of chemicals to human is nowadays one of the most challenging tasks. Quantitative structure–activity relationship (QSAR) models are welcome tools to cope with complex, expensive and time consuming experimental methods for evaluation of carcinogenic potency. Therefore, in last decade, vast effort was involved to introduce new in silico models for prediction of carcinogenicity of non-congeneric chemicals that can be effectively used for regulatory purposes in the scope of new legislation REACH (Registration, Evaluation, Authorisation and Restriction of Chemicals). In this chapter we focus on models developed in the scope of CAESAR and PROSIL projects which were implemented in on-line available internet platform VEGA (http://www.vega-qsar.eu/use-qsar.html). These QSAR models for prediction of carcinogenic potency are based on counter propagation artificial neural network algorithm (CPANN). CP ANN algorithm represents a suitable tool for modeling of complex biological data like carcinogenicity. We emphasized on the representation of key development steps needed to be involved in model construction to meet requirement of five OECD principles. First of all, it reported the description of carcinogenicity endpoint and analysis of quality of chemical and biological data (principle 1), followed by an explanation of the CPANN algorithm selected for modelling (principle 2). Next, the interpretation of domain of applicability for non-congeneric chemicals (principle 3) was given. Furthermore, the statistical performance characteristics of models in sense of its goodness-of-fit, robustness and predictivity was reported (principle 4), and finally, the mechanistic interpretation of models on the basis of selected types of descriptors and structural alerts of studied chemicals was represented (principle 5).