Prediction of physicochemical properties and biological activities from molecular structure and the use of computational neural networks for the analysis of sensor array data

This thesis contains the description of research performed in two areas of computational chemistry. The first half of the thesis concerns the development of quantitative structure-activity relationships (QSAR). A QSAR is a mathematical link between chemical structure and a biological activity or physical property. The latter half of the thesis involves the development and application of methods for the classification of airborne odorants using an array of non-specific chemical sensors. The use of such arrays is modeled on the mammalian olfactory system. These chemical sensing devices are often referred to as artificial noses.; QSAR models are presented for the prediction of the acute mammalian toxicity of substituted anilines. The effect of different methods of feature selection is discussed. Included in this discussion is the development of a feature selection method that utilizes robust regression methods to evaluate the quality of descriptor subsets for predictive models. The toxicity values are expressed as the log(LD50) of the compounds to mice. The QSAR model developed predicts the toxicity for compounds in an external prediction set with a root mean square (rms) error of 0.24 log units. QSAR models were also developed to predict the clearing temperatures of a series of compounds exhibiting a liquid crystalline mesophase. The ability of the models to encode trends in homologous series is presented. The clearing temperatures are predicted with a rms error of 7 K.; The artificial nose uses an array of 19, non-specific, optical sensors with time dependent responses. Each of these sensor responses is encoded using a series of descriptors meant to encode information important to odorant identification. Many of these descriptors capture temporal information. Two pattern recognition methods are developed and applied in this thesis. A number of data sets were collected to demonstrate the utility of these approaches. The ability of both of these methods, learning vector quantization and fuzzy ARTMAP, to accurately classify the odorants is presented and discussed.