DEqMS: A Method for Accurate Variance Estimation in Differential Protein Expression Analysis.
Zhu Y, Orre LM, Zhou Tran Y, Mermelekas G, Johansson HJ, Malyutina A, Anders S, Lehtiö J. · 2020
Read DEqMS: A Method for Accurate Variance Estimation in Differential Protein Expression Analysis. by Zhu Y, Orre LM, Zhou Tran Y, Mermelekas G, Johansson HJ, Malyutina A, Anders S, Lehtiö J. (2020) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc7261819.txt.
Introduction
Zhu and co-authors introduce DEqMS, a statistical method for spotting proteins whose levels change between samples. Their methods paper appeared in Molecular and Cellular Proteomics. Mass spectrometry studies measure thousands of proteins at once, and finding real differences depends on estimating how much each measurement naturally varies.
Most methods miss one fact: a protein's variance depends on how many peptide fragments were used to measure it. DEqMS builds that dependence into its estimates. Tested against current approaches, it gave better accuracy and stronger statistical power on both labeled and label-free data. The method ships as a user-friendly R package, so proteomics groups can add it to standard workflows for biomarker hunting and basic biology.
Catalog record
| CREATOR | Zhu Y, Orre LM, Zhou Tran Y, Mermelekas G, Johansson HJ, Malyutina A, Anders S, Lehtiö J. |
|---|---|
| YEAR | 2020 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 52 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
DEqMS: A Method for Accurate Variance Estimation in Differential Protein Expression Analysis* By Yafeng Zhu, Lukas M Orre, Yan Zhou Tran, Georgios Mermelekas, Henrik J Johansson, Alina Malyutina, Simon Anders, Janne Lehtiö Published in: Molecular & Cellular Proteomics : MCP DOI: 10.1074/mcp.TIR119.001646 ABSTRACT Differential analysis of MS-data to identify biomarkers or to understand biology is a cornerstone in proteomics. DEqMS is a robust method for analysis of both labelled and label-free MS-data. The method takes into account the inherent dependence of protein variance on the number of PSMs or peptides used for quantification, thereby providing a more accurate variance estimation. Compared to current methods, DEqMS achieves better accuracy and statistical power in quantitative proteomics. The tool is available as user-friendly R package. Keywords: Quantification, label-free quantification, statistics, data evaluation, mass spectrometry, quality control and metrics, differential analysis, FDR, TMT Mass spectrometry (MS)-based proteomics is widely used for identification and quantification of proteins from complex biological samples (1). In a typical proteomics experiment, proteins are digested into peptides using a proteolytic enzyme, commonly trypsin, prior to MS-analysis. Peptides are usually fragmented multiple times, thus generating multiple peptide spectrum matches (PSMs)1 for the same peptide sequence. Peptide identification and quantification is subsequently performed in data analysis workflows using either label-free or labeled approaches. Both quantitative approaches rely on a hierarchical data structure: PSMs are nested into peptides which are then nested into proteins. Protein level quantification is subsequently generated through summarizing peptide or PSM level information. In previous quantitative proteomics analysis, Student t test, ANOVA (2), Limma (3) and linear mixed models (2, 4–6), have been used to detect differentially expressed proteins (DEPs). In addition, other methods have been developed specifically for quantitative proteomics as exemplified by empirical Bayesian random censoring threshold model (EBRC) (7) and reproducibility-optimized test statistic (ROTS) (8). EBRC, developed for label-free MS-data, applies a random censoring threshold to achieve more accurate protein abundance estimates in the presence of missing values. ROTS is a modified form of t-statistics with a bootstrapping procedure to maximize the reproducibility of top ranked DEPs. In a previous analysis, ROTS was shown to outperform many statistical methods including Limma, t test, SAM and mixed models (9). Limma, originally developed to analyze RNA microarray data, performs “moderated” ANOVA, i.e. it uses an empirical-Bayes approach to shrink gene-wise sample variance toward a common estimate based on all experimental data (3). In an evaluation performed by Kammers et al., Limma detected more truly differentially expressed proteins compared with t test at the same false discovery rate (FDR) (10). Further, in a comparison between generalized linear model (GLM), mixed effects models and Limma in a tandem mass tag (TMT) 10-plex labeled proteomics data set, mixed effects models showed better specificity when combined with median sweeping normalization compared with the other two methods, whereas Limma had better statistical results independent of normalization methods (11). However, there are problems in current methods. For proteins identified with a single PSM/peptide, the estimation of parameters in mixed effect models is less accurate, and in addition, it requires users to understand the correct random effects and to use appropriate statistical tests (11). In Kammers et al.'s study, proteins quantified with a single peptide were excluded from the Limma analysis (10), and D'Angelo et al. showed that omitting proteins quantified by a single peptide reduced false discovery findings (11). In a typical shotgun proteomics experiment, 10–20% of proteins are identified with only one peptide or PSM. Most of these are low abundant proteins that are potentially central in regulation of biological processes or contribute as biomarker candidates and are therefore important to keep in the analysis. Further, t test, ANOVA and Limma do not consider the difference in quantitative variation for proteins identified by different numbers of peptides/PSMs. It has previously been shown by us and others that the accuracy of protein abundance estimates varies by the number of peptides quantified (12–14). In microarray and RNA-seq data analysis, probe intensity and read counts are used to estimate the mean-variance relationship instead of using a common prior variance for all genes, resulting in improved performance over Limma (15, 16). Inspired by this approach, we have developed DEqMS (Differential Expression analysis of quantitative Mass Spectrometry data), to get a more accurate data-dependent estimation of protein variance based on the number of peptides or PSMs used for quantification. MATERIALS AND METHODS Experimental Design and Statistical Rationale DEqMS is a statistical method for identification of differentially expressed proteins in MS-data, that takes into acount the dependence of variance on the number of PSMs or peptides used for protein quantification. The data sets used to evaluate the method and the statistics behind the method is presented in detail below. Data Set D1 - A431 Cells Treated with EGFR Inhibitor Gefitinib Search results of this data set was downloaded from ProteomeXchange with identifier PXD006291 (17). PSM raw intensity table (combining all four IPG strips in the experiment) filtered at 1% PSM and protein level FDR (gene symbol centric search; 1,263,974 PSMs; 203,640 unique peptides; 10,166 genes), was used to calculate protein relative abundance by median sweeping method (18). First, raw intensity values are log2 transformed. Second, for each PSM, the median of log2 intensity is subtracted to get relative log2 ratio. Third, for each protein, its relative log2 ratio is calculated as the median of log2 ratio of all PSMs belonging to this protein. At last, the medians of protein log2 ratios in different samples are subtracted to get equal protein level in all samples, with the assumption that the treatment does not impact the total amount of proteins in different conditions. The generated protein table with log2 ratios without missing values (10,124 proteins) was used for t test, Limma, and DEqMS analysis. Biological triplicates of untreated and gefitinib treated (24 h) cells were used in the analysis for differential protein expression. Data Set D2 - E. coli Spike-in Label-free Data Set Search results of this data (50,016 unique peptides; 6566 proteins) were downloaded from the ProteomeXchange repository (PXD000279) (19). E. coli proteome was prepared into two conditions with 1:3 ratio (10 μg versus 30 μg) in triplicates and added into equal amount of human proteome background (50 μg Hela cells protein extract), see detailed sample preparation in (19). LFQ intensities in “proteinGroups.txt” table, for proteins with quant values in at least two samples per condition (5022 proteins), were log2 transformed and used as input matrix for all methods. To evaluate intensity-based Bayesian approach, trend = TRUE option was enabled in eBayes function provided in Limma. For variance stabilization normalization analysis, original protein LFQ intensity values were normalized by justvsn function in R package vsn before using Limma. ROC analysis was restricted to proteins unique to E. coli or human, with higher ratios in samples spiked-in with the higher amount of E. coli proteins. Data Set D3 - E. coli Spike-in TMT Data Set Sample Preparation One MCF-7 and one E. coli K-12 cell pellet were lysed and sonicated in a buffer containing 4% SDS, 25 mm HEPES pH 7.6 and 1 mm DTT. Total protein amount was estimated (DC protein assay, Bio-Rad, Hercules, California). Samples with different spiked in amounts of E. coli protein extract (3 replicates with 7.5