Robust Summarization and Inference in Proteome-wide Label-free Quantification

Abstract

Label-Free Quantitative mass spectrometry based workflows for differential expression (DE) analysis of proteins impose important challenges on the data analysis because of peptide-specific effects and context dependent missingness of peptide intensities. Peptide-based workflows, like MSqRob, test for DE directly from peptide intensities and outperform summarization methods which first aggregate MS1 peptide intensities to protein intensities before DE analysis. However, these methods are computationally expensive, often hard to understand for the non-specialized end-user, and do not provide protein summaries, which are important for visualization or downstream processing. In this work, we therefore evaluate state-of-the-art summarization strategies using a benchmark spike-in dataset and discuss why and when these fail compared with the state-of-the-art peptide based model, MSqRob. Based on this evaluation, we propose a novel summarization strategy, MSqRobSum, which estimates MSqRob's model parameters in a two-stage procedure circumventing the drawbacks of peptide-based workflows. MSqRobSum maintains MSqRob's superior performance, while providing useful protein expression summaries for plotting and downstream analysis. Summarizing peptide to protein intensities considerably reduces the computational complexity, the memory footprint and the model complexity, and makes it easier to disseminate DE inferred on protein summaries. Moreover, MSqRobSum provides a highly modular analysis framework, which provides researchers with full flexibility to develop data analysis workflows tailored toward their specific applications. Label-Free Quantitative mass spectrometry based workflows for differential expression (DE) analysis of proteins impose important challenges on the data analysis because of peptide-specific effects and context dependent missingness of peptide intensities. Peptide-based workflows, like MSqRob, test for DE directly from peptide intensities and outperform summarization methods which first aggregate MS1 peptide intensities to protein intensities before DE analysis. However, these methods are computationally expensive, often hard to understand for the non-specialized end-user, and do not provide protein summaries, which are important for visualization or downstream processing. In this work, we therefore evaluate state-of-the-art summarization strategies using a benchmark spike-in dataset and discuss why and when these fail compared with the state-of-the-art peptide based model, MSqRob. Based on this evaluation, we propose a novel summarization strategy, MSqRobSum, which estimates MSqRob's model parameters in a two-stage procedure circumventing the drawbacks of peptide-based workflows. MSqRobSum maintains MSqRob's superior performance, while providing useful protein expression summaries for plotting and downstream analysis. Summarizing peptide to protein intensities considerably reduces the computational complexity, the memory footprint and the model complexity, and makes it easier to disseminate DE inferred on protein summaries. Moreover, MSqRobSum provides a highly modular analysis framework, which provides researchers with full flexibility to develop data analysis workflows tailored toward their specific applications. Label-free quantitation (LFQ) mass spectrometry (MS) based workflows have become standard practice in quantitative proteomics (e.g. (1Goeminne L.J.E. Gevaert K. Clement L. Experimental design and data-analysis in label-free quantitative LC/MS proteomics: A tutorial with MSqRob.J. Proteomics. 2018; 171: 23-36Crossref PubMed Scopus (26) Google Scholar, 2Tebbe A. Klammer M. Sighart S. Schaab C. Daub H. Systematic evaluation of label-free and super-SILAC quantification for proteome expression analysis.Rapid Commun. Mass Spectrom. 2015; 29: 795-801Crossref PubMed Scopus (15) Google Scholar)). This technology typically starts with protein extraction followed by an enzyme digestion step to produce peptides of a convenient length. The thus obtained peptide mixture is then analyzed in a mass spectrometer where intact peptide masses and their intensities are measured, resulting in an MS1 spectrum. In typical LFQ, the intensities of the thus recorded peaks are taken as proxies for peptide abundance. To identify the peaks observed in the MS1 spectrum, these peaks are first isolated in the instrument, and then subjected to fragmentation. Each of the resulting fragmentation (MS2) spectra is then used for peptide identification. In LFQ, each sample is separately analyzed on the mass spectrometer, and differential expression is obtained by comparing relative intensities between runs for the same identified peptide (1Goeminne L.J.E. Gevaert K. Clement L. Experimental design and data-analysis in label-free quantitative LC/MS proteomics: A tutorial with MSqRob.J. Proteomics. 2018; 171: 23-36Crossref PubMed Scopus (26) Google Scholar). However, this workflow also induces challenging data analysis problems. First, different peptides from the same protein often have very distinct physio-chemical properties, leading to large differences in their MS1 intensities even though these peptides are of similar abundance (supplemental Fig. S1A1). Second, because of technological constraints not all peptides can be subjected to fragmentation. Indeed, only those peptides with the highest MS1 intensities within a certain retention window are typically selected for fragmentation (3Tu C. Li J. Sheng Q. Zhang M. Qu J. Systematic assessment of survey scan and MS2-based abundance strategies for label-free quantitative proteomics using high-resolution MS data.J. Proteome Res. 2014; 13: 2069-2079Crossref PubMed Scopus (34) Google Scholar). As a result, the identification in any given run depends not only on the abundance of that peptide, but also on the abundances of any co-eluting peptides. There can thus be context-depending missingness in a given run. Moreover, there are many other potential sources of (random or non-random) missingness, including peptide misidentification, ambiguous matching of MS1 peaks, and poor quality MS2 spectra (4Lazar C. Gatto L. Ferro M. Bruley C. Burger T. Accounting for the multiple natures of missing values in label-free quantitative proteomics data sets to compare imputation strategies.J. Proteome Res. 2016; 15: 1116-1125Crossref PubMed Scopus (128) Google Scholar). Hence, there is considerable variation in terms of the peptides that are identified in each of the different MS runs in an experiment. Taken together, the identification issue and the peptide specific effects on quantification have a severe impact on the downstream summarization of peptide intensities toward protein abundances (5Goeminne L.J.E. Argentini A. Martens L. Clement L. Summarization vs peptide-based models in label-free quantitative proteomics: performance, pitfalls, and data analysis guidelines.J. Proteome Res. 2015; 14: 2457-2465Crossref PubMed Scopus (21) Google Scholar). Indeed, because of these issues, simple summarization methods such as the mean or median peptide intensity are known to give unreliable protein abundance estimates (5Goeminne L.J.E. Argentini A. Martens L. Clement L. Summarization vs peptide-based models in label-free quantitative proteomics: performance, pitfalls, and data analysis guidelines.J. Proteome Res. 2015; 14: 2457-2465Crossref PubMed Scopus (21) Google Scholar) and more advanced summarization strategies have therefore been proposed for LFQ data in the literature (6Silva J.C. Gorenstein M.V. Li G.Z. Vissers J.P.C. Geromanos S.J. Absolute quantification of proteins by Lcmse.Mol. Cell. Proteomics. 2005; 5: 144-156Abstract Full Full PubMed Scopus Google Scholar, J. M. label-free quantification by and peptide Cell. Proteomics. 2014; 13: Full Full PubMed Scopus Google Scholar, H. M. identification of using 2018; 13: PubMed Scopus Google Scholar, M. T. T. an for analysis of quantitative mass 2014; PubMed Scopus Google Scholar). In Fig. we the of these different data analysis strategies on a benchmark we differences in between the different summarization which are by the and any differences in this of a protein between Moreover, of the summarization strategies the all To these summarization issues, peptide-based such as L.J.E. Gevaert K. Clement L. and in quantitative label-free Cell. Proteomics. 2016; 15: Full Full PubMed Scopus (34) Google for expression (DE) of proteins directly from the observed peptide intensities. The is that these methods outperform methods L.J.E. Gevaert K. Clement L. and in quantitative label-free Cell. Proteomics. 2016; 15: Full Full PubMed Scopus (34) Google Scholar) and Fig. Indeed, by peptide intensities for differences in peptide and for differences in the of identified peptides for a given protein in each resulting in a and a on the However, the also from drawbacks compared with summarization sample effects to for between the peptide intensities for a given protein in the same This makes data analysis computationally more of of the test and even these is because of in the peptides The of effects also makes it to disseminate the toward non-specialized as the of the more Moreover, not provide protein summaries for each which are important for to and the and for in downstream applications. therefore a novel for using a two-stage which we MSqRobSum provides protein summaries that for peptide specific which are then using Hence, MSqRobSum the of MSqRob's with the of which and modular data analysis workflows. In these workflows from the visualization and of the protein that is by the of MSqRobSum on a spike-in why it state-of-the-art for DE in LFQ quantitative and it on a of state-of-the-art for DE analysis of proteins on a benchmark spike-in compared peptide-based and and analysis of data all we to the workflow as by the also novel summarization for DE MSqRobSum, which to MSqRob's superior while also providing useful protein expression summaries. The of MSqRobSum and other state-of-the-art for differential expression analysis using a dataset S. Li J. Q. L. C. proteomics quantification in large 2018; PubMed Scopus Google Scholar)). different and in a The resulting run on an mass data with J. M. peptide identification mass and protein PubMed Scopus Google using the and proteome with the of as a and of and of the protein as In to between runs also using The resulting by In all proteins are as DE and all proteins as To benchmark and of these different quantification the and of a of proteins by the with a of DE proteins as the proteins with a then a certain The is then the of proteins in the of and proteins the is the of all proteins also MSqRobSum on In the first from of The dataset of LFQ MS data from and and is to as the dataset A. K. M. J. M. analysis of label-free and for quantitative 2015; PubMed Scopus Google Scholar). The data with by The M. L. on the spectra that by of quantification data in 2018; Scholar). In a C. M. M. of in and protein Cell. Proteomics. 2015; 14: Full Full PubMed Scopus Google Scholar) the in the and this the To this and in and each analyzed in with LFQ The data with C. M. M. of in and protein Cell. Proteomics. 2015; 14: Full Full PubMed Scopus Google Scholar). This dataset is to as the the workflow in the from the values as in the in M. C. an for downstream analysis of 2018; Scholar). that are only identified as or are from the data The intensities of in a given sample that can be to the same peptide are intensities are to protein intensities using the (6Silva J.C. Gorenstein M.V. Li G.Z. Vissers J.P.C. Geromanos S.J. Absolute quantification of proteins by Lcmse.Mol. Cell. Proteomics. 2005; 5: 144-156Abstract Full Full PubMed Scopus Google Scholar). to their leading intensities are to the and median intensities in each sample are intensities are DE of proteins is analyzed in with using the differential expression for and Res. 2015; PubMed Scopus Google Scholar). that we an for protein in all model for In the analysis the models are with the protein intensity in sample of the the of spike-in the which are to be with mean and The are with which the estimates by for multiple using the a standard workflow from the protein summaries by protein summaries are and intensity values for each protein in each can the as The median of the peptides from a protein in all sample is on these is used to an protein expression This is to the peptide intensities from this protein in all J. M. label-free quantification by and peptide Cell. Proteomics. 2014; 13: Full Full PubMed Scopus Google Scholar). protein summaries, as in in that are only identified by a and are from the data on the values are for all for multiple using the A standard M. T. T. an for analysis of quantitative mass 2014; PubMed Scopus Google workflow starts from the peptide intensities in with only or all and peptides that in more protein are a peptide is multiple in a only the intensity is The peptide intensities are median and missing values are using an intensities are to protein intensities using median analysis and Google Scholar). models based on these protein summaries. with the protein intensity in sample of the the of spike-in the which are to be with mean and The multiple is using the values are analyzed with the standard workflow in the H. M. identification of using 2018; 13: PubMed Scopus Google Scholar). that are or that from are from the data proteins with missing values in are The data are using A. A. M. and for Google Scholar). values are for proteins that are missing and proteins that are missing not (4Lazar C. Gatto L. Ferro M. Bruley C. Burger T. Accounting for the multiple natures of missing values in label-free quantitative proteomics data sets to compare imputation strategies.J. Proteome Res. 2016; 15: 1116-1125Crossref PubMed Scopus (128) Google Scholar). proteins are as proteins observed in for and these are with proteins are to be missing abundance and are thus are when missing in and are with a In a is from a the first of all observed protein in the and with a standard as the median standard DE of proteins is analyzed in with using the differential expression for and Res. 2015; PubMed Scopus Google like the is using an as in the The data is using the L. for mass spectrometry data and PubMed Scopus Google Scholar). The analysis is using the peptide intensities as in the in The spike-in data and the data are using as in the The are using M. A of methods for data based on and PubMed Scopus Google like the analysis in the of L.J.E. Gevaert K. Clement L. and in quantitative label-free Cell. Proteomics. 2016; 15: Full Full PubMed Scopus (34) Google Scholar). that are only identified by a and are from the data To peptide to and proteins are that are only observed all are also in which a protein is only observed in are in the DE analysis for this is a peptide-based the with the intensity of peptide in sample with the the of spike-in a that for the in expression between the peptides from the same protein in on the sample the of peptide the is to be with mean and only peptide is for a protein in all the model reduces The parameters for and peptide are using by the between effects and in the is by toward when there are only This and makes the more and The of is by using The of the terms are with is using the MSqRob's model can also be a two-stage analysis models for in Scholar). we first peptide intensities to the protein and test for DE on these protein summaries. The same is used as for the analysis in In the first we aggregate all peptide intensities of a protein using with using the the intensity of peptide in sample and the of peptide the peptide as a can be as the mean intensity in sample for this The is to be with mean and In the we an analysis on protein intensities with the the protein intensity in sample of the the of spike-in the is to be with mean and for multiple using the can a in in MSqRobSum compared with because we on the intensities and variation also do not that the of the sample estimates is highly dependent on the of peptide intensities in the sample models for in Scholar). However, we that the resulting impact on is in analysis and using the is for The been in L.J.E. Gevaert K. Clement L. and in quantitative label-free Cell. Proteomics. 2016; 15: Full Full PubMed Scopus (34) Google Scholar). However, we in and it to also for proposed two-stage strategy, we a model for each protein we the which the and MSqRobSum analysis. In a full we typically for in the using for of the model However, the analysis in MSqRobSum only by This while The and are as a MSqRobSum The summarization is also to the for summarization in the L. for mass spectrometry data and PubMed Scopus Google Scholar). methods and novel are all using a dataset where an proteome different in a first compare for DE analysis of LFQ based quantitative and why the of we that novel MSqRobSum, maintains the of the based MSqRob. that MSqRobSum toward modular data analysis workflows and we and why MSqRobSum this by MSqRobSum in In this we compare methods and and peptide-based model The workflow for missingness of peptides abundance by using the which provide intensities by the mean intensity of the peptides (6Silva J.C. Gorenstein M.V. Li G.Z. Vissers J.P.C. Geromanos S.J. Absolute quantification of proteins by Lcmse.Mol. Cell. Proteomics. 2005; 5: 144-156Abstract Full Full PubMed Scopus Google Scholar, M. C. an for downstream analysis of 2018; Scholar). However, this not for peptide specific effects and by only using the peptide intensities. This and in the protein summaries (supplemental Fig. Indeed, the peptides typically between leading to a compared with all other methods The workflow is based on summarization and analysis with using J. M. label-free quantification by and peptide Cell. Proteomics. 2014; 13: Full Full PubMed Scopus Google Scholar). for peptide specific effects by abundance of peptides between However, the in often considerable which to missingness and summaries (supplemental Fig. that spike-in often have peptides between and these are unreliable for though for peptide by for it to The of also in analysis as their only the of the data for the that are in the The summarization with the downstream analysis often in a compared with the other methods The analysis of data based analysis by a imputation for missing protein intensities that missingness or missingness because of abundance H. M. identification of using 2018; 13: PubMed Scopus Google Scholar). also provides a more downstream DE analysis using models with the differential expression for and Res. 2015; PubMed Scopus Google Scholar)). Hence, more as as more estimates and the analysis M. T. T. an for analysis of quantitative mass 2014; PubMed Scopus Google Scholar) is for proteomics data analysis. peptide-based using However, workflows in which peptide intensities are first to protein and is then on the protein The of summarization in is median which for peptide specific effects and is in the of missing data but this is in by missing peptide intensities by using an However, imputation strategy, that all intensities are missing because of thus to other sources of Indeed, it that when the imputation step is in with spike-in (supplemental Fig. that are The median summarization in more estimates compared with However, based workflows and Fig. and in with and This because the are by a of spike-in with missingness by the a of spike-in with missingness from other The missingness by abundance of is therefore more to be for leading to a and a of for these In which also for missingness, not for these be that more for proteins and thus proteins which the of in a Indeed, when only then (supplemental Fig. MSqRob, a peptide-based that provides and by and a procedure for L.J.E. Gevaert K. Clement L. and in quantitative label-free Cell. Proteomics. 2016; 15: Full Full PubMed Scopus (34) Google Scholar). thus estimates with and all models The in is even more and In this we that we can the model in a two-stage with impact on In the first of MSqRobSum we peptide intensities in a sample to protein intensities using This summarization is and more and summarization (supplemental Fig. In the MSqRobSum provides and to (supplemental Fig. MSqRobSum a similar to for to highly proteins and Fig. and the of MSqRobSum is that of for and it be that all summarization methods from a in A to the of MSqRobSum is protein which protein summaries for the in because of intensities of peptide in all of (supplemental Fig. As a result, this protein a very from the MSqRobSum analysis for and is thus as a (supplemental Fig. The models the the peptide and the between sample and this The is the and for and MSqRobSum the of in differential expression in and (supplemental Fig. The of in the because spike-in between the peptide in the sample data by and and 5: PubMed Google Scholar, L.J.E. Argentini A. Martens L. Clement L. Summarization vs peptide-based models in label-free quantitative proteomics: performance, pitfalls, and data analysis guidelines.J. Proteome Res. 2015; 14: 2457-2465Crossref PubMed Scopus (21) Google Scholar). This in peptides with abundance in to be in the sample with the protein resulting in a intensity for those peptides in that This is as the of the proteins as the (supplemental Fig. more proteins thus to be and these the (supplemental Fig. The for to MSqRobSum of is based on with MSqRob. The first issue is that it is which of be used for the test with MSqRob. the of of the the peptide sample but these do not to the of of the standard on the Indeed, these standard the within sample and the between sample and the of therefore between those of the within sample and those that be obtained for a that models the data protein the of be and the of are thus to be The issue is as the large models in a workflow is computationally In the summarization in the first of MSqRobSum is a to the two-stage in MSqRobSum, analysis is to a of the of methods maintains this while even Fig. Moreover, MSqRobSum also protein summaries that are useful for visualization and for downstream which are not in as it models the peptide intensities The MSqRobSum workflow of summarization with and DE analysis with each step can have an important impact on the of the data analysis the of summarization and provides flexibility to each of the MSqRobSum with other in modular workflows. To the of such modular workflows, we from the workflow and we each step in the MSqRobSum workflow the The workflow of summarization with for is and the is not the or the However, data analysis a the which is in the effects be for the analysis but it is not and are in (supplemental Fig. strategies for these sample can thus the and of the analysis by the intensities with The in is based on which are when with more and other more can therefore by intensities with MSqRob's which for and that is not in and because of as also (supplemental Fig. summarization strategy, based on between is for with which to summaries with this a imputation strategy, which the of the protein there is missingness and the of imputation be MSqRobSum, we for peptide-specific effects a summarization which models the peptide intensities directly This all protein intensities and imputation such as used in is therefore not (supplemental Fig. The full MSqRobSum workflow thus while Moreover, this workflow all other modular In this we MSqRobSum on The that the proteome of and and the that the impact of the on the proteome of the Fig. the of MSqRob, MSqRobSum, and in the that is more and many more proteins a the summarization based DE MSqRobSum and and DE proteins and MSqRobSum However, MSqRobSum more that a more which in a multiple only proteins that analyzed with MSqRobSum more proteins the (supplemental Fig. This that the in DE proteins between MSqRobSum and is by the strategy, and that MSqRobSum is the the we the differences between and MSqRobSum (supplemental Fig. methods similar estimates similar test and similar standard on the estimates for the of the a of by to the or very distinct estimates are The of are for for MSqRobSum for many proteins in we on the DE proteins by MSqRob, we can identify why more DE proteins are with The test of for the of of the standard on the which of the within and between sample is therefore based on the of the within sample which is for proteins with many peptides. the MSqRobSum analysis is run with the the of DE proteins from to and in Fig. The test is for those proteins that are only for MSqRob. This is because of a standard for in Fig. and that the for the between sample can be to have a large in with and that it therefore be often by in Fig. and such the test of as all observed data are which to an of the standard on the a the of the is also in MSqRobSum, leading to a more (supplemental Fig. and thus that can be to produce protein in with sample the of L.J.E. Gevaert K. Clement L. and in quantitative label-free Cell. Proteomics. 2016; 15: Full Full PubMed Scopus (34) Google we also evaluate the of the in the dataset of C. M. M. of in and protein Cell. Proteomics. 2015; 14: Full Full PubMed Scopus Google Scholar). In Fig. we that the analysis many DE of these DE have to and are in the as compared with the (supplemental Fig. This is because the induces which can be to many proteins that the analysis all as because for Hence, the analysis can be to produce and many DE proteins with also that the proteins that by and not by not for differential expression between and The and the MSqRobSum analysis DE as from of these DE proteins are in as compared with (supplemental Fig. MSqRobSum proteins that MSqRob. However, in this with more MS we can that the between is by the in of Indeed, the of and MSqRobSum is when are based on the (supplemental Fig. In this work, we MSqRobSum, a novel for LFQ which protein intensity and protein DE analysis. a benchmark of different for summarization based LFQ methods and the state-of-the-art peptide-based model, MSqRob. the on all peptides and all which can only on the protein summaries. However, models are computationally expensive, can be hard to understand by with of and do not provide protein summaries for visualization and downstream processing. drawbacks are not in Indeed, summarization is a and reduces the of data the obtained protein summaries of the The of protein summaries also reduces model and with that have of However, many methods a considerable in compared with analysis that this in is dependent on with the summarization that do not peptide specific such as the in a in performance, a like peptide specific but is based on and is not very data MSqRobSum, we on for which for peptide-specific all data in model based and is Taken together, the is a considerable in in the DE analysis when compared with also that is for the of a DE The first of such is which can have a large impact on DE analysis The of is imputation of missing and this can be However, because different imputation methods and because each of these to different sources of missingness, the are typically when using a where missing values and values missing abundance are H. M. identification of using 2018; 13: PubMed Scopus Google Scholar). be that the in MSqRobSum can imputation (supplemental Fig. in LFQ is the model for DE standard but these are compared with based models in more Moreover, the model to it and to (5Goeminne L.J.E. Argentini A. Martens L. Clement L. Summarization vs peptide-based models in label-free quantitative proteomics: performance, pitfalls, and data analysis guidelines.J. Proteome Res. 2015; 14: 2457-2465Crossref PubMed Scopus (21) Google Scholar). In the MSqRobSum we therefore MSqRob's model of on the protein summaries. This considerably of the DE a to for a of DE proteins for in DE all summarization methods from a in in these often more severe that of MSqRobSum Fig. and Moreover, we in the analysis of the that can be in with because of of the of and an of the between sample the summarization the that the analysis workflow become the protein abundance estimates can be used for visualization and in other for can also from protein summaries by other This considerable flexibility to develop modular workflows that are tailored toward their specific and when novel and more summarization become the of the in the of with differential expression label-free quantitation mass spectrometry differential analysis of proteomics data model missing missing not of

Topics & Concepts