Mixture Likelihood Ratio Test of Proteomics Data

• Activity Number: 393
• Type: Contributed
• Date/Time: Tuesday, August 5, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Stephen Erickson*+ and Horace J. Spencer and Ricky D. Edmondson and Samuel G. Mackintosh and Mayumi Nakagawa
• Companies: University of Arkansas for Medical Sciences and University of Arkansas for Medical Sciences and University of Arkansas for Medical Sciences and University of Arkansas for Medical Sciences and University of Arkansas for Medical Sciences
• Keywords: Proteomics ; Mixture distribution ; Genomics

Abstract

Proteomics data produced by mass spectrometry are complicated by a non-zero probability of non-detection, even for proteins which are relatively abundant. In a Phase I clinical trial of a therapeutic HPV vaccine, 5 women with HSILs were treated with the vaccine and provided blood samples at 3 time-points: baseline, post 2, and post 4 injections. PBMCs were isolated, eluted peptides underwent MS/MS analysis, and a total of 1279 proteins were detected in at least one of 15 samples. Among 3837 protein x time-points, we observed a strong relationship between mean log-intensity and the log-odds of detection, and used this relationship to develop a likelihood function which explicitly models non-detections, leading to a likelihood ratio test (LRT) of differential expression. Permutation testing and simulations show nominal control of Type I error, while ensemble information across proteins can be used to regularize the likelihood function in the spirit of empirical Bayes and other shrinkage estimators. Our results suggest that the LRT is more sensitive to changes in expression than commonly used tests, and our approach is computationally cheap and extensible to richer data models.