Improved Normalization of Systematic Biases Affecting Ion Current Measurements in Label-free Proteomics Data

The number of laboratories using MS as a quantitative tool for protein profiling continues to grow, propelling the field forward past simple qualitative measurements (i.e. cataloging), with the aim of establishing itself as a robust method for detecting proteomic differences. By analogy, semiquantitative proteomic profiling by MS can be compared with measurement of relative gene expression by genomics technologies such as microarrays or, newer, RNAseq measurements. While proteomics is disadvantaged by the lack of a molecular amplification system for proteins, successful reports from discovery experiments are numerous in the literature and are increasing with advances in instrument resolution and sensitivity. In general, methods for performing relative quantitation can be broadly divided into two categories: those employing labels (e.g. iTRAQ, TMT, and SILAC (1)) and so-called “label-free” techniques. Labeling methods involve adding some form of isobaric or isotopic label(s) to the proteins or peptides prior to liquid chromatography-tandem MS (LC-MS/MS) analysis. Chemical labels are typically applied during sample processing, and isotopic labels are commonly added during cell culture (i.e. metabolic labeling). One advantage of label-based methods is that the two (or more) differently-labeled samples can be mixed and run in single LC-MS analyses. This is in contrast to label-free methods which require the samples to be run independently and the data aligned post-acquisition. Many labs employ label-free methods because they are applicable to a wider range of samples and require fewer sample processing steps. Moreover, data from qualitative experiments can sometimes be re-analyzed using label-free software tools to provide semiquantitative data. Advances in these software tools have been extensively reviewed (2). While analysis of label-based data primarily uses full MS scan (MS1)1 or tandem MS scan (MS2) ion current measurements, analysis of label-free data can employ simple counts of confidently identified tandem mass spectra (3). So-called spectral counting makes the assumption that the number of times a peptide is identified is proportional to its concentration. These values are sometimes summed across all peptides for a given protein and scaled by protein length. Relative abundance can then be calculated for any peptide or protein of interest. While this approach may be easy to perform, its usefulness is particularly limited in smaller data sets and/or when counts are low. This report focuses only on the use of ion current measurements in label-free data sets, specifically those calculated from extracted MS1 ion chromatograms (XICs). In general terms, raw intensity values (i.e. ion counts in arbitrary units) cannot be used for quantitation in the absence of cognate internal standards because individual ion intensities depend on a response factor, related to the chemical properties of the molecule. Intensities are instead almost always reserved for relative determinations. Furthermore, retention times are sometimes used to align the chromatograms between runs to ensure higher confidence prior to calculating relative intensities. This step is crucial for methods without corresponding identity information, particularly for experiments performed on low-resolution instruments. To support a label-free workflow, peptide identifications are commonly made from tandem mass spectra (MS/MS) acquired along with direct electrospray signal (MS1). Or, in alternative workflows seeking deeper coverage, interesting MS1 components can be targeted for identification by MS/MS in follow-up runs (4). “Rolling up” the peptide ion information to the peptide and protein level is also done in different ways in different labs. In most cases, “peptide intensity” or “peptide abundance” is the summed or averaged value of the identified peptide ions. How the peptide information is transferred to the protein level differs between methods but typically involves summing one or more peptide intensities, following parsimony analysis. One such solution is the “Top 3” method developed by Silva and co-workers (5). Because peptides in label-free methods lack labeled analogs and require separate runs, they are more susceptible to analytical noise and systematic variations. Sources of these obscuring variations can come from many sources, including sample preparation, operator error, chromatography, electrospray, and even from the data analysis itself. While analytical noise (e.g. chemical interference) is difficult to selectively reject, systematic biases can often be removed by statistical preprocessing. The goal of these procedures is to normalize the data prior to calculations of relative abundance. Failure to resolve these issues is the common origin of batch effects, previously described for genomics data, which can severely limit meaningful interpretation of experimental data (6, 7). These effects have also been recently explored in proteomics data (8). Methods used to normalize proteomics data have been largely borrowed from the microarray community, or are based on a simple mean/median intensity ratio correction. Methods applied on microarray and/or gene chip and used on proteomics data include scaling, linear regression, nonlinear regression, and quantile normalizations (9). Moreover, work has also been done to improve normalization by subselecting a peptide basis (10). Other work suggests that linear regression, followed by run order analysis, works better than other methods tested (11). Key to this last method is the incorporation of a variable other than intensity during normalization. It is also important to note that little work has been done towards identifying the underlying sources of these variations in proteomics data. Although cause-and-effect is often difficult to determine, understanding these relationships will undoubtedly help remove and avoid the major underlying sources of systematic variations. In this report, we have attempted to combine our efforts focused on understanding variability with the work initiated by others for normalizing ion current-based label-free proteomics data. We have identified several major variables commonly affecting peptide ion intensities both within and between labs. As test data, we used a subset of raw data acquired during Phase I of the National Cancer Institute's (NCI) Clinical Proteomics Technology Assessment for Cancer (CPTAC) program. With these data, we were able to develop a statistical model to rank bias variables and normalize the intensities using stepwise, semiparametric regression. The data analysis methods have been implemented within the National Institute of Standards and Technology (NIST) MS quality control (MSQC) pipeline. Finally, we have developed R code for removing systematic biases and have tested it using a reference standard spiked into a complex biological matrix (i.e. yeast cell lysate).