Will Thompson

Followers

Following

Public Views

Ricardo Henao

Duke University

Aimee Zaas

Duke University School of Medicine

Alfred O. Hero

University of Michigan

Gregory Sepich-Poore

Ilkka Julkunen

Leo Lahti

University of Helsinki

University of Manitoba

Interests

Uploads

Papers by Will Thompson

Latent Protein Trees

R. HENAO ET AL. effectiveness for the study of biological and medical questions [Aebersold and Ma... more R. HENAO ET AL. effectiveness for the study of biological and medical questions [Aebersold and Mann (2003), Service (2008), Ping (2009)]. While early work in this field met with a number of notorious failures [Petricoin et al. (2002), Baggerly et al. (2004), Zhang and Chan (2005)] due to overlapping peaks, batch effects and systematic noise, high accuracy spectrometers along with multiple fractionation techniques such as liquid chromatography and ion mobility have led to increased robustness as well as improved qualitative and quantitative results. After summarization, data generated by this technology is typically presented as a p × n dimensional matrix of real-valued intensities where the number of measured features p is typically orders of magnitude larger than n, as in microarray gene expression data. However, there are a number of important characteristics that distinguish mass spectrometry proteomics from gene expression data. First, each feature is a short peptide that has been enzymatically cut out of a parent protein, and parent proteins typically give rise to many such peptides. Second, only the more abundant of these features are typically identified (meaning that the peptide sequence and originating protein are known). Third, features that are present at lower abundances will typically have numerous missing values across the samples. Finally, while the error rate for assigning identifications to features is low, it is not zero, and this leads to some peptides with incorrect identifications. Analysis approaches for these data can be performed at the feature level or at the protein level. The obvious consequence of performing analysis at the feature level is a significant loss of power due to the highly dependent nature of subsets of the features-particularly those that originate from the same protein. We prefer a dimension reduction approach in which groups of features are collected and summarized prior to analysis of associations between features and biological phenotypes. There are a number of approaches to this in the literature, almost all of which rely entirely on the identified features in the data set. The simplest of these approaches involves direct summarization of all or some features that are identified for each protein either through averaging or robust summarization based on quantiles [Polpitiya et al. (2008)]. There are also a number of regression approaches which include fixed effects for protein, peptide and experimental group [Karpievitch et al. (2009)], include an additional random effect for situations in which subjects are measured in replicate [Daly et al. (2008)], or add additional interaction effects between treatment and features [Clough et al. (2009)]. These may assume constant or varying noise levels across isotope groups and have been shown in some cases to exhibit better performance than naive summarization approaches that do not adjust for confounding factors [Clough et al. (2009)]. We are aware of only one approach to the analysis of these data that examines correlation structure between data features [Lucas et al. (2012)].

Will Thompson

Related Authors

Uploads

Papers by Will Thompson

Log In