Computer-Aided Molecular Design with Multi-Dimensional Characterization

Abstract

In this work, a methodology for the solution of computer-aided molecular design (CAMD) problems with property models utilizing descriptors of varying dimensionality has been presented. The problems encountered within this field typically require the selection, or design, of pure chemicals, as well as mixtures, exhibiting a desired set of properties and attributes. These properties and attributes are captured through property models, which have widely varying forms. These property models are most often a function of molecular descriptors, which provide a quantitative reference to the structural features in a molecule. There are multitudes of descriptor types, each which can be immediately categorized based on the dimensionality of information they capture. This is one of the strengths of computer aided molecular design, the flexibility to develop a specific model for each property of interest. However, it often leads to the selection of very complex and widely differing property models for each property of interest. An ideal CAMD methodology would not restrict the property modelling stage to certain types of independent variables, and as such, could solve these problems on a single platform. The problem with developing such an algorithm is that the descriptors chosen are often of varying dimensionalities. Inclusion of descriptors beyond two-dimensional requires some consideration of the potential energy surface, or conformational space, for each candidate solution. In addition, the region in which to search for solutions becomes difficult to identify because each property model has its own applicability domain, within which predictions can be made with increased confidence. The approach presented within this dissertation, aimed at solving such problems, utilizes a fragment based descriptor known as the signature descriptor. Previous applications using this descriptor were shown to be successful in terms of solving the problem in an efficient manner while identifying novel solutions. Extension of this descriptor to include spatial information, along with the techniques necessary for using this data, is presented. This has allowed for the estimation of likely local energy minima without the conventional conformational analysis for each potential solution, which has been shown to be computationally intensive. The nature of signature descriptors, being fragment based, allows for an efficient description of which region in chemical space to search for solutions and also facilitates reconstruction of solutions matching a set of descriptor values. A description of previous approaches taken to solve problems of this nature has been outlined such that the benefits of the proposed technique could be exemplified. In addition, several studies have been provided to verify the proposed methodology.