The development of predictive models for physical and biological properties from molecular structure and the analysis of data from a conduction polymer chemiresistive sensor array

Research focusing on two aspects of computational chemistry are presented in this thesis. The first area involves the development of QSPR and QSAR models. The goal of this research is to develop models for predicting physical properties or biological activities of compounds.{09}Analogously, classification models are also developed for predicting the class membership of biologically-relevant compounds. The second area of research deals with the analysis of data collected from arrays of conducting polymer microsensors.; An introduction to the methodology of QSPR and QSAR is presented, as are several applications. The first study reports QSPR models for predicting the surface tension, viscosity, and thermal conductivity of 213 organic solvents. The models for surface tension and viscosity compare favorably to previously published QSPR models while the model developed for thermal conductivity is the first such model reported in the literature. The second study involves the development of QSAR models for predicting the inhibitory concentrations of 113 inhibitors of the Na+/H+ antiporter. A five-descriptor CNN model which resulted in an RMS error 0.377 log units for prediction set compounds is reported. In a third application, experimental IC₅₀ data for 314 selective COX-2 inhibitors are used to develop QSAR and classification models as a potential screening mechanism for larger libraries of target compounds. An eight-descriptor committee CNN model was identified as a robust predictor, producing an RMS error of 0.625 log units for the external prediction set of inhibitors. The fourth application involves the development of classification models to predict the genotoxicity of a set of 255 structurally-diverse secondary and aromatic amine compounds. For the prediction set, 70.4% of the nontoxic compounds and 79.2% of the toxic compounds were correctly classified.; The remainder of this thesis describes work analyzing chemical sensor array data to characterize the sensors used. Results of these experiments identified two or three sensors which have poor reproducibility over time. Furthermore, it was shown that collecting data with shorter vapor exposure periods does not significantly reduce the discrimination ability of the arrays. Finally, there is evidence that a deterioration of results may occur when data is collected several months apart.