Knowledge Discovery in Molecular Structure Databases

In recent years there has been an explosion of information generated on molecular structures. This information is usually three-dimensional in nature, and is normally appended to the existing molecular database in the form of raw coordinate data. There is a recognized need for new tools to structure, manage, and analyze these data. This thesis presents a knowledge discovery technique, based on a theory of conceptual clustering, in which recurrent associations between molecular constitution and conformation can be discovered in three-dimensional molecular structure databases. These associations are de ned in rstorder logic as structured concepts, and are maintained in a taxonomy partially ordered by subsumption. The knowledge discovery method is applied to two large molecular structure databases. In the Cambridge Structural Database the technique is used to perform conformational clustering of several small molecule datasets. The resultant conformational templates are evaluated by comparison with hand analyses and statistical clusterings of the same data. In the Protein Data Bank the technique is used to discover recurrent protein motifs which have sequences predictive of their associated structure. The resultant rules are evaluated with respect to the average structural variation of their instances, and according to a test of statistical correlation between sequence and structure. In both molecular structure databases interesting and signi cant associations were discovered. i Acknowledgements My initial inspiration and excitement for the topic of knowledge discovery in molecular databases developed during intensive research meetings with Frank Allen, Suzanne Fortier, and Janice Glasgow at Queen's University in late 1990. Since then, and during my research, I have become increasingly fascinated by the topic. For their role in maintaining my interest, special thanks is due to my supervisors, Janice Glasgow and Suzanne Fortier. They have provided constant encouragement, enthusiasm, interest, and support for this work. Frank Allen at the Cambridge Crystallographic Data Centre has been a tremendous source of information and inspiration for this work. Much of Chapter 5 would not have been possible without his help in providing me with structural data and listings from statistical clustering experiments. I would also like to thank the Cambridge Crystallographic Data Centre for providing me with the opportunity to use their facilities during several research visits from 1992 through 1995. I would like to thank Gilles Bisson, Diane Cook, Bob Levinson and Kevin Thompson for their generous time and valuable comments on a draft of Chapter 2. Jude Shavlik, as editor of the Machine Learning journal, helped out a lot with the overall presentation and clarity of Chapter 6. Gregory Piatetsky-Shapiro o ered useful advice on the statistical evaluation of predictive rules. Several groups of people have listened to presentations of this thesis material. For their generous time and useful comments, I would like to thank the Department of Information Studies at She eld University, the Ottawa Machine Learning Group, the Biomolecular Informatics group at ZymoGenetics Inc. and the Molecular Scene Analysis Group at Queen's University. The prototype software developed to support this thesis was done in the Q'Nial array programming language. I would like to thank the creator of the language, Mike Jenkins, for taking an interest in my work and for ensuring that I always had the best programming facilities to work with. Chris Walmsley has given me all kinds of technical advice and support; in particular, I would like to acknowledge his MolView software which was used to create some of the gures in Chapters 5 and 6. Figure 5.10 was prepared using the daVinci graph layout system. The diagrams in Figure II.1 were prepared using the QUEST and PLUTO programs from the Cambridge Structural Database. This thesis is dedicated to my grandmother, Phyllis Mobray Hughes (1906-1990). ii