Development of Computational Tools for Use in Quantitative Structure-Activity and Structure-Property Relationships

Computational tools are developed to relate theoretical aspects of molecular structure to experimental chemical behavior. The assumption that a causal relationship exists between molecular structure and chemical behavior is the fundamental premise that underlies the fields of quantitative structure-activity relationships (QSAR) and quantitative structure-property relationships (QSPR). Generation of a QSAR or QSPR involves the characterization of chemical structure in terms of numerical indices, and the development of mathematical models that correlate these indices with biological activities or physicochemical properties. Tools are developed in this thesis for use at three important stages of this process. A fast quantum mechanical molecular modeling technique based on a modified extended Huckel approach is developed for placing molecules in realistic energy-minimized conformations. The extended Huckel neglect-of-differential-overlap (EHNDO) method utilizes a one-electron approximation, semiempirical parameterization, and an efficient BFGS geometry optimization. The calculated molecular geometries from EHNDO are as accurate as those produced by the two-electron AM1 method, and the EHNDO geometry optimization is approximately three times faster. A new empirical method for calculating atomic charges is developed to characterize electronic structure. These charges are used in combination with the concepts of induction and resonance to arrive at an empirical model for pK _ a estimation in organic acids and bases. The PKACHG method generates partial atomic charges that yield accurate predictions of electric dipole moment, and the pK_ a model is accurate over wide ranges of acidity and basicity. Finally, nonlinear mathematical modeling techniques are investigated as a means of generating QSARs and QSPRs that contain stronger connections between calculated structure and experimental chemical behavior. A computational neural network program, QNET, is developed and outfitted with an efficient BFGS optimization for network training. A quadratic fitting routine, QUADFIT, is proposed as a fast alternative to neural networks. Both QNET and QUADFIT utilize external cross-validation during model development to prevent experimental data from being overfit. In a QSAR study of amine toxicity, QNET significantly outperforms conventional linear regression, and both QNET and QUADFIT yield post-model predictions that are considerably more accurate than those of linear regression.