A treatment of stereochemistry in computer aided organic synthesis

This thesis describes the author’s contributions to a new stereochemical processing module constructed for the ARChem retrosynthesis program. The purpose of the module is to add the ability to perform enantioselective and diastereoselective retrosynthetic disconnections and generate appropriate precursor molecules. The module uses evidence based rules generated from a large database of literature reactions. Chapter 1 provides an introduction and critical review of the published body of work for computer aided synthesis design. The role of computer perception of key structural features (rings, functions groups etc.) and the construction and use of reaction transforms for generating precursors is discussed. Emphasis is also given to the application of strategies in retrosynthetic analysis. The availability of large reaction databases has enabled a new generation of retrosynthesis design programs to be developed that use automatically generated transforms assembled from published reactions. A brief description of the transform generation method employed by ARChem is given. Chapter 2 describes the algorithms devised by the author for handling the computer recognition and representation of the stereochemical features found in molecule and reaction scheme diagrams. The approach is generalised and uses flexible recognition patterns to transform information found in chemical diagrams into concise stereo descriptors for computer processing. An algorithm for efficiently comparing and classifying pairs of stereo descriptors is described. This algorithm is central for solving the stereochemical constraints in a variety of substructure matching problems addressed in chapter 3. The concise representation of reactions and transform rules as hyperstructure graphs is described. Chapter 3 is concerned with the efficient and reliable detection of stereochemical symmetry in both molecules, reactions and rules. A novel symmetry perception algorithm, based on a constraints satisfaction problem (CSP) solver, is described. The use of a CSP solver to implement an isomorph‐free matching algorithm for stereochemical substructure matching is detailed. The prime function of this algorithm is to seek out unique retron locations in target molecules and then to generate precursor molecules without duplications due to symmetry. Novel algorithms for classifying asymmetric, pseudo‐asymmetric and symmetric stereocentres; meso, centro, and C2 symmetric molecules; and the stereotopicity of trigonal (sp2) centres are described. Chapter 4 introduces and formalises the annotated structural language used to create both retrosynthetic rules and the patterns used for functional group recognition. A novel functional group recognition package is described along with its use to detect important electronic features such as electron‐withdrawing or donating groups and leaving groups. The functional groups and electronic features are used as constraints in retron rules to improve transform relevance. Chapter 5 details the approach taken to design detailed stereoselective and substrate controlled transforms from organised hierarchies of rules. The rules employ a rich set of constraints annotations that concisely describe the keying retrons. The application of the transforms for collating evidence based scoring parameters from published reaction examples is described. A survey of available reaction databases and the techniques for mining stereoselective reactions is demonstrated. A data mining tool was developed for finding the best reputable stereoselective reaction types for coding as transforms. For various reasons it was not possible during the research period to fully integrate this work with the ARChem program. Instead, Chapter 6 introduces a novel one‐step retrosynthesis module to test the developed transforms. The retrosynthesis algorithms use the organisation of the transform rule hierarchy to efficiently locate the best retron matches using all applicable stereoselective transforms. This module was tested using a small set of selected target molecules and the generated routes were ranked using a series of measured parameters including: stereocentre clearance and bond cleavage; example reputation; estimated stereoselectivity with reliability; and evidence of tolerated functional groups. In addition a method for detecting regioselectivity issues is presented. This work presents a number of algorithms using common set and graph theory operations and notations. Appendix A lists the set theory symbols and meanings. Appendix B summarises and defines the common graph theory terminology used throughout this thesis.