Clustering of proteomics imaging mass spectrometry data

Abstract

This thesis presents a toolbox for the exploratory analysis of multivariate data, in particular proteomics imaging mass spectrometry data. Typically such data consist of 15000 - 20000 spectra with a spatial component, and for each spectrum ion intensities are recorded at specific masses. Clustering is a focus of this thesis, with discussion of k-means clustering and clustering with principal component analysis (PCA). Theoretical results relating PCA and clustering are given based on Ding and He (2004), and detailed and corrected proofs of the authors' results are presented. The benefits of transformations prior to clustering of the data are explored. Transformations include normalisation, peak intensity correction (PIC), binary and log transformations. A number of techniques for comparing different clustering results are also discussed and these include set based comparisons with the Jaccard distance, an information based criterion (variation of information), point-pair comparisons (Rand index) and a modified version of the prediction strength of Tibshirani and Walther (2005). These exploratory analyses are applied to imaging mass spectrometry data taken from patients with ovarian cancer. The data are taken from slices of cancerous tissue. The analyses in this thesis are primarily focused on data from one patient, with some techniques demonstrated on other patients for comparison.