The Applications of Artificial Neural Networks in the Identification of Quantitative Structure-Activity Relationships for Chemotherapeutic Drug Carcinogenicity

We investigate which of two Artificial Intelligence techniques is superior at making predictions about complex carcinogen systems. Artificial Neural Networks are shown to provide good predictions of carcinogen toxicology bands for drugs which are themselves used to treat cancerous cells, by using a novel system of molecular descriptors derived from the molecules’ mass spectrometry intensities, reduced in dimensionality by Principal Component Analysis, to form a series of orthogonal descriptors which retain 95% of the variance of the original data. The creation of molecular descriptors from PCA-resolved mass spectrometry data is shown to be superior to the use of Self-Organising Maps, the selection of a series of modal fragments, or the use of every peak (within the confines of the precepts of Artificial Intelligence). A new system of backpropagation which increases network efficacy in this case is also proposed.