Protein quantification across hundreds of experimental conditions

Abstract

Quantitative studies of protein abundance rarely span more than a small number of experimental conditions and replicates. In contrast, quantitative studies of transcript abundance often span hundreds of experimental conditions and replicates. This situation exists, in part, because extracting quantitative data from large proteomics datasets is significantly more difficult than reading quantitative data from a gene expression microarray. To address this problem, we introduce two algorithmic advances in the processing of quantitative proteomics data. First, we use space-partitioning data structures to handle the large size of these datasets. Second, we introduce techniques that combine graph-theoretic algorithms with space-partitioning data structures to collect relative protein abundance data across hundreds of experimental conditions and replicates. We validate these algorithmic techniques by analyzing several datasets and computing both internal and external measures of quantification accuracy. We demonstrate the scalability of these techniques by applying them to a large dataset that comprises a total of 472 experimental conditions and replicates.

Fig. 1.

A representative LC-MS/MS scan. A scan consists of millions of peaks, each with a specific retention time, m/z, and intensity. The intensity of a peak is indicated in grayscale (darker gray corresponds to higher intensity). The first Inset shows several extracted ion chromatograms (XICs) corresponding to the naturally occurring isotopes of a tryptic peptide. The area under an XIC provides a measure of the relative abundance of a peptide. In this example, one peak in the XIC has an MS/MS fragmentation spectrum from which the peptide's amino acid sequence and, hence, the identity of the protein from which it derives can be determined.

Fig. 2.

Comparison of known protein concentration with relative abundance measured by mass spectrometry. Measured protein abundance is plotted on a log–log scale against known femtomole concentration for each of six nonhuman proteins spiked in to human serum by Mueller et al. The lines show best fit by regression, and the corresponding log–log scale slopes and correlation values (R²) for both the log–log scale and the original scale are shown below each plot.