Logistic Regression Modeling on Mass Spectrometry Data in Proteomics Case-Control Discriminant Studies

Abstract

We present an adaption of the logistic regression model for the evaluation of mass spectrometry data in proteomics case-control studies. We parameterize the predictor as a linear combination of Gaussian basis functions along the mass/charge axis. The location of these basis functions is treated as a random variable and must be estimated from the data. A fully Bayesian implementation is pursued, which allows the number of functional components within the regression parameter vector to be specified as a random variable. Calculations are implemented through birth–death process modeling. We evaluate the model on data from a block-randomized case-control designed experiment, as well as on a proteomic model-mouse study, which were both carried out at the Leiden University Medical Center. The first experiment compares mass spectra of serum samples of 63 colon cancer patients with spectra from 50 control patients. We present a-posteriori analyses of the fitted models which allow researchers to select specific spectral regions for further investigation and identification of the associated differentially expressed peptides. A sensitivity study is presented which links some of our results to those which may be obtained through standard maximum likelihood logistic regression on principal components reduction for mass spectral data. The second experiment contrasts proteomic spectra from 18 dystrophin-deficient mdx mice with those from 74 controls.