Mass spectrometry has emerged as a core technology for high throughput proteomics profiling in biomedical research. However, the complexity of the data poses new statistical challenges for the analysis. Statistical methods and software developments for analyzing proteomics data are likely to continue as a major area of research in the coming years.

In this dissertation, we develop a novel statistical method for analyzing MS data. We propose to use the chemical knowledge regarding isotopic distribution of the peptide molecules along with quantitative modeling to detect chemically valuable peaks from each spectrum. More specifically, a mixture of location-shifted Poisson distribution is fitted to the deamidated isotopic distribution of a peptide molecule in low to moderate molecular weight of the mass spectrum. Maximum likelihood estimation by the expectation-maximization (EM) technique is used to estimate the parameters of the distribution. We then identify the monoisotopic peaks of the spectrum through formal statistical hypotheses testing procedures.

Unlike low to moderate range MS data a Poisson distribution is not suitable for high mass ranges of the spectrum data due to symmetric nature of the isotopic distribution. Also, due to preprocessing and pronounced effect of the additional sources of variability, a Poisson approximation to the binomial model to the isotopic distribution may not hold. Therefore, a mixture of location-shifted Normal model is fitted to model each of the deamidated (possibly) isotopic distribution of a mass spectrum. A nonlinear optimization method to maximize the observed data likelihood is applied instead of EM algorithm to estimate the parameters of the distribution. Similar statistical testing procedures are applied for the peak detection method.

A study of the effectiveness of our features selection method compared to some other relatively new feature selection methods in classifying case and control samples is explored. Superiority of our method is established in terms of the overall classification accuracy, sensitivity, specificity and area under the receptor operative curve (ROC) curve.