Analysis of Microarray Data Using Inductive Logic Programming and Ontological Background Information

Abstract

The use of formal logic in chemometric analysis has been sparse. However, there are some problems which are well suited for analysis with inductive logic programming (ILP), a class of machine learning methods based on first-order predicate logic. The principal strength of ILP is that input, background knowledge and output can contain first-order logic statements. Two results of this is that ILP in contrast to other common methods is capable of learning common substructures in a set of graphs, and that it has the ability to use graph-structured background knowledge in analysis. Particularly the growing set of biomedical ontologies is an interesting research area in this respect.

This thesis first contain an introduction to ILP together with a review of chemical and biochemical problems where the method have been applied until now. Most papers on the topic have been published in the machine learning literature, and this paper serves as an overview to the parts of the literature that are most relevant for chemistry. The most well-studied problem is structure–activity relationships (SAR) for diverse sets of molecules, because the relational atom–bond structure is well suited for relational representation and hence also ILP learning.

The next part of the thesis is a further enhancement of using ILP for SAR problems. A method is implemented which use descriptors from quantum topology instead of atoms and bonds as the molecular representation. This method performs equally well to the previous atom–bond representation, but have the advantage of being rooted in quantum theory while keeping the representational simplicity of the concepts of atoms and bonds.

Two articles are devoted to the main aim of this thesis; the use of the gene ontology as background information in microarray analysis. The first article introduces the method, presents the setup and compare it to traditional analysis methods. ILP performs similar to other methods in terms of predictive performance. The advantage of ILP is in the possibility of a classification theory to include both ontological information and expression information for single genes.

After asserting that ILP can be used for such analysis, the next article studies how the variable selection performed on the microarray data before running ILP will affect the result. Three different variable selection methods are tested, and the result was that ILP is not very much affected by variable selection performed on the dataset before analysis, as long as enough relevant information is included