Mass spectrometry data mining for cancer detection

Early detection of cancer is crucial for successful intervention strategies. Mass spectrometry-based high throughput proteomics is recognized as a major breakthrough in cancer detection. Many machine learning methods have been used to construct classifiers based on mass spectrometry data for discriminating between cancer stages, yet, the classifiers so constructed generally lack biological interpretability. To better assist clinical uses, a key step is to discover ”biomarker signature profiles”, i.e. combinations of a small number of protein biomarkers strongly discriminating between cancer states.

This dissertation introduces two innovative algorithms to automatically search for a signature and to construct a high-performance signature-based classifier for cancer discrimination tasks based on mass spectrometry data, such as data acquired by MALDI or SELDI techniques. Our first algorithm assumes that homogeneous groups of mass spectra can be modeled by (unknown) Gibbs distributions to generate an optimal signature and an associated signature-based classifier by robust log-likelihood analysis; our second algorithm uses a stochastic optimization algorithm to search for two lists of biomarkers, and then constructs a signature-based classifier.

To support these two algorithms theoretically, this dissertation also studies the empirical probability distributions of mass spectrometry data and implements the actual fitting of Markov random fields to these high-dimensional distributions. We have validated our two signature discovery algorithms on several mass spectrometry datasets related to ovarian cancer and to colorectal cancer patients groups. For these cancer discrimination tasks, our algorithms have yielded better classification performances than existing machine learning algorithms and in addition,have generated more interpretable explicit signatures.