Exploratory analysis of chemical structure, bacterial mutagenicity and rodent tumorigenicity

There are a great number of chemical compounds in use by humans and there are very limited resources for testing those compounds for potential toxicity. We conduct an exploratory analysis of molecular features to determine if molecular features can be related to bacterial mutagenicity and rodent tumorigenicity. We use numerical molecular topology and the presence of cited toxic functional groups as molecular descriptors. Near neighbor searching is used to find similar compounds to highlight that similar compounds can differ in toxicity. Principal components analysis and linear regression are used in an exploratory analysis. A large data set is divided at random into two groups and the reproducibility of regression results between the two data sets is examined. We determine that many simple `structural alertʹ rules of prediction of toxic effects are overly simplistic and that regression analysis is poorly reproducible; we conclude that prediction of gross toxicology from simple structural and regression rules is questionable.