Network-based inference from complex proteomic mixtures using SNIPE

Abstract

Motivation: Proteomics presents the opportunity to provide novel insights about the global biochemical state of a tissue. However, a significant problem with current methods is that shotgun proteomics has limited success at detecting many low abundance proteins, such as transcription factors from complex mixtures of cells and tissues. The ability to assay for these proteins in the context of the entire proteome would be useful in many areas of experimental biology.

Results: We used network-based inference in an approach named SNIPE (Software for Network Inference of Proteomics Experiments) that selectively highlights proteins that are more likely to be active but are otherwise undetectable in a shotgun proteomic sample. SNIPE integrates spectral counts from paired case-control samples over a network neighbourhood and assesses the statistical likelihood of enrichment by a permutation test. As an initial application, SNIPE was able to select several proteins required for early murine tooth development. Multiple lines of additional experimental evidence confirm that SNIPE can uncover previously unreported transcription factors in this system. We conclude that SNIPE can enhance the utility of shotgun proteomics data to facilitate the study of poorly detected proteins in complex mixtures.

Availability and implementation: An implementation for the R statistical computing environment named snipeR has been made freely available at http://genetics.bwh.harvard.edu/snipe/.

Contact: ssunyaev@rics.bwh.harvard.edu

Supplementary information: Supplementary data are available at Bioinformatics online.

Fig. 1.

Diagram of the SNIPE method. Spectral counts for each protein are matched to their nodes in a given network. A score for the protein is calculated by summing the node (blue) and its immediate neighbours (grey). A P-value is assigned by permuting the counts in the network and generating a distribution of scores for that node and comparing the observed score to that distribution

Fig. 2.

SNIPE performance. (A–D) Fractions of true- and false-positive proteins recovered at various significance thresholds for (A) SNIPE using permutation-based multiple test correction, (B) SNIPE using FDR for multiple test correction and (C) microarray. (D) Fraction of genes in the MGI database annotated for causing a tooth phenotype recovered at different FDR for SNIPE and Microarray. (E–G) Numbers of proteins in the Helsinki database recovered by (E) SNIPE using permutation-based multiple test correction, (F) SNIPE using FDR compared with microarray. (G and H) Numbers of proteins known to cause a phenotype in the MGI database recovered by (G) SNIPE using permutation-based multiple test correction and (H) SNIPE using FDR compared with microarray

Fig. 3.

Immunofluorescence of E13.5 tooth bud. (A) Diagram of the E13.5 murine first molar tooth bud. The tooth at this stage is composed of invaginating dental epithelium (DE, blue) and surrounding condensed dental mesenchyme (DM, orange). The condensed cells indicate differential cellular fate of the condensed dental mesenchyme from the surrounding non-dental mesenchyme (M). The non-dental oral epithelium (OE) is a cellular bilayer divided by the oral cavity (grey line). (B) Etv5 immunostain shows uniform ubiquitous nuclear expression in the tooth and non-tooth regions. (C) Six1 immunostain shows similarly ubiquitous nuclear expression in the tooth and non-tooth regions with a slight visible enrichment in the dental mesenchyme matching the region previously reported for its gene expression pattern (Nonomura et al., 2010). (D) SF1 (Nr5a1) and (E) Sox11 immunostains show specific expression in the dental mesenchyme and all epithelial tissue. (F) Zeb1 immunostain shows a nuclear localized stain specifically in the mesenchymal tissue with slight upregulation in the dental mesenchyme. Scale bars: