Predicting mutagenicity of aromatic and heteroaromatic amine mutagens using a newly developed set of laplacian matrix-based graph invariants

This paper presents an effort for development of predictive quantitative structure-activity relationship (QSAR) models to assess mutagenicity of a set of 95 aromatic and heteroaromatic amines using newly developed set of Laplacian matrix-based molecular descriptors using graph convolution. In order to calculate these descriptors we used (1) properties of isolated atoms, (2) properties of the atoms, which arise from their molecules, and (3) properties derived from the graphs that represent the molecules. The QSAR models developed are based on counter-propagation artificial neural network algorithm. The optimization was performed in automated manner using genetic algorithm for (1) selection of the most suitable descriptors as well as for (2) selection of the most suitable parameters of the neural networks. Results presented in this study show that this new type of descriptors, reported in this paper, encode structural information capable of predicting TA98 Ames' mutagenicity of the amines.