Direct Maximization of Protein Identifications from Tandem Mass Spectra

Abstract

The goal of many shotgun proteomics experiments is to determine the protein complement of a complex biological mixture. For many mixtures, most methodological approaches fall significantly short of this goal. Existing solutions to this problem typically subdivide the task into two stages: first identifying a collection of peptides with a low false discovery rate and then inferring from the peptides a corresponding set of proteins. In contrast, we formulate the protein identification problem as a single optimization problem, which we solve using machine learning methods. This approach is motivated by the observation that the peptide and protein level tasks are cooperative, and the solution to each can be improved by using information about the solution to the other. The resulting algorithm directly controls the relevant error rate, can incorporate a wide variety of evidence and, for complex samples, provides 18–34% more protein identifications than the current state of the art approaches. The goal of many shotgun proteomics experiments is to determine the protein complement of a complex biological mixture. For many mixtures, most methodological approaches fall significantly short of this goal. Existing solutions to this problem typically subdivide the task into two stages: first identifying a collection of peptides with a low false discovery rate and then inferring from the peptides a corresponding set of proteins. In contrast, we formulate the protein identification problem as a single optimization problem, which we solve using machine learning methods. This approach is motivated by the observation that the peptide and protein level tasks are cooperative, and the solution to each can be improved by using information about the solution to the other. The resulting algorithm directly controls the relevant error rate, can incorporate a wide variety of evidence and, for complex samples, provides 18–34% more protein identifications than the current state of the art approaches. The problem of identifying proteins from a collection of tandem mass spectra involves assigning spectra to peptides, using either a de novo or database search strategy, and then inferring the protein set from the resulting collection of peptide-spectrum matches (PSMs). 1The abbreviations used are:PSMpeptide-spectrum matchFDRfalse discovery rateHUhidden units. In practice, the goal of such an experiment is to identify as many distinct proteins as possible at a specified false discovery rate (FDR). However, most of the previous work in the context of shotgun proteomics analysis has focused on controlling error rates at the level of PSMs or peptides (1Moore R.E. Young M.K. Lee T.D. Qscore: An algorithm for evaluating Sequest database search results.J. Am. Soc. Mass Spectrom. 2002; 13: 378-386Crossref PubMed Scopus (339) Google Scholar, 2Choi H. Ghosh D. Nesvizhskii A.I. Statistical validation of peptide identifications in large-scale proteomics using target-decoy database search strategy and flexible mixture modeling.J. Proteome Res. 2008; 7: 286-292Crossref PubMed Scopus (100) Google Scholar, 3Käll L. Storey J.D. MacCoss M.J. Noble W.S. Assigning significance to peptides identified by tandem mass spectrometry using decoy databases.J. Proteome Res. 2008; 7: 29-34Crossref PubMed Scopus (449) Google Scholar, 4Choi H. Nesvizhskii A.I. False discovery rates and related statistical concepts in mass spectrometry-based proteomics.J. Proteome Res. 2008; 7: 47-50Crossref PubMed Scopus (170) Google Scholar, 5Keller A. Nesvizhskii A.I. Kolker E. Aebersold R. Empirical statistical model to estimate the accuracy of peptide identification made by MS/MS and database search.Anal. Chem. 2002; 74: 5383-5392Crossref PubMed Scopus (3928) Google Scholar, 6Fenyö D. Beavis R.C. A method for assessing the statistical significance of mass spectrometry-based protein identification using general scoring schemes.Anal. Chem. 2003; 75: 768-774Crossref PubMed Scopus (407) Google Scholar, 7Geer L.Y. Markey S.P. Kowalak J.A. Wagner L. Xu M. Maynard D.M. Yang X. Shi W. Bryant S.H. Open mass spectrometry search algorithm.J. Proteome Res. 2004; 3: 958-964Crossref PubMed Scopus (1170) Google Scholar, 8Perkins D.N. Pappin D.J. Creasy D.M. Cottrell J.S. Probability-based protein identification by searching sequence databases using mass spectrometry data.Electrophoresis. 1999; 20: 3551-3567Crossref PubMed Scopus (6842) Google Scholar, 9Fitzgibbon M. Li Q. McIntosh M. Modes of inference for evaluating the confidence of peptide identifications.J. Proteome Res. 2008; 7: 35-39Crossref PubMed Scopus (31) Google Scholar, 10Elias J.E. Gygi S.P. Target-decoy search strategy for increased confidence in large-scale protein identifications by mass spectrometry.Nat. Methods. 2007; 4: 207-214Crossref PubMed Scopus (2901) Google Scholar, 11Huttlin E.L. Hegeman A.D. Harms A.C. Sussman M.R. Prediction of error associated with false-positive rate determination for peptide identification in large-scale proteomics experiments using a combined reverse and forward peptide sequence database strategy.J. Proteome Res. 2007; 6: 392-398Crossref PubMed Scopus (65) Google Scholar) rather than the protein level FDR. peptide-spectrum match false discovery rate hidden units. This approach creates difficulties for estimating protein level FDRs because the PSM or peptide level error rate may be significantly lower than the protein level error rate, especially in the context of a deeply saturated experiment (12Adamski M. Blackwell T. Menon R. Martens L. Hermjakob H. Taylor C. Omenn G.S. States D.J. Data management and preliminary data analysis in the pilot phase of the HUPO Plasma Proteome Project.Proteomics. 2005; 5: 3246-3261Crossref PubMed Scopus (50) Google Scholar, 13Reiter L. Claassen M. Schrimpf S.P. Jovanovic M. Schmidt A. Buhmann J.M. Hengartner M.O. Aebersold R. Protein identification false discovery rates for very large proteomics data sets generated by tandem mass spectrometry.Mol. Cell. Proteomics. 2009; 8: 2405-2417Abstract Full Text Full Text PDF PubMed Scopus (259) Google Scholar). For example, consider a collection of 1000 spectra that map to 100 distinct peptides with a 1% false discovery rate. This 1% false discovery rate corresponds to 10 incorrectly mapped spectra, each of which is likely to map to a different, incorrect peptide. Thus, the PSM error rate of 1% corresponds to a peptide error rate of 10/110 = 9%. A similar inflation of error rate will occur if we move to the protein level. In general, when the end goal is to find the optimal solution to a protein level problem, it is conceptually and practically beneficial to directly solve the problem of interest rather than artificially dividing the problem into two separate tasks. The two tasks of protein and peptide level optimization are closely related but are likely to have different optimal solutions. Moreover, many machine learning problems involving several subtasks have been shown to benefit from a top-down approach that solves several subtasks simultaneously, in contrast to solving each of them separately. For example, the handwritten document recognition task involves a variety of intermediate problems including extraction of the field of interest, segmentation into characters, and character recognition. An algorithm that combines all the subtasks into a top-down optimization problem substantially outperforms algorithms that treat these subtasks as distinct modules (14LeCun Y. Bottou L. Bengio Y. Haffner P. Gradient-based learning applied to document recognition.Proc. IEEE. 1998; 86: 2278-2324Crossref Scopus (33131) Google Scholar). Similarly, object recognition (on photographs, for example) involves defining a hierarchy of in the such as and an object that learning of the in with in with that a Y. C. and in of on and Scopus Google Scholar). object recognition learning have been used in tasks from for Y. E. in 2005; Google Scholar) to segmentation problems in M. D. M. W. W. R. by optimization with of the on and Google Scholar, that to A for Google Scholar). in a is and involves several intermediate such as of An approach that to all of these tasks simultaneously, as as or than all of the for each R. Bottou L. M. P. from Res. Scholar, R. A for with of the on 2008; Google Scholar). In this we that inferring proteins from peptide-spectrum matches is of a problem that from the top-down a machine learning method to directly the the of proteins identified by the a decoy search strategy, searching each a database of peptides and then a learning algorithm to a on the combined set of and decoy learning the of the model that the of the is with proteins. with A.I. A. Kolker E. Aebersold R. A statistical model for identifying proteins by tandem mass Chem. 2003; 75: PubMed Scopus Google Scholar, P. P. H. in protein inference from shotgun proteomics using peptide of the on 2007; Google Scholar, analysis accuracy and Proteome Res. 2007; 6: PubMed Scopus Google Scholar, Li Y. P. Q. H. A approach to protein inference problem in shotgun M. L. of the on Scopus Google the approach several by an optimization problem that at the protein approach controls the relevant error rate. we the of the problem to and protein level tasks simultaneously, and we that these tasks are such that each task from the solution to the approach PSMs at of the with the that low scoring PSMs can information about the of a protein when in the context of PSMs to this The protein identification problem can be as a with corresponding to spectra, peptides, and proteins An from a to a peptide that the database search a to the peptide. In general, more than may be to a single peptide. is possible to consider more than scoring match for each as we in An from a peptide to a protein that the peptide in the This is because each protein peptides, and each peptide may in more than The to the problem is the with a set of to each peptide-spectrum In this we each PSM using that of the and of the as as the of the match the and The is a on with proteins that are in the the of the used to and current and current and for peptide of the of the match on the mass in and of the in and of and of the of database peptides the specified the peptide by an the peptide have an of of the in the proteins a protein by if then a to end Open in a solve the protein identification problem using a decoy databases have been used in shotgun proteomics for two to false discovery rate for peptide identifications (1Moore R.E. Young M.K. Lee T.D. Qscore: An algorithm for evaluating Sequest database search results.J. Am. Soc. Mass Spectrom. 2002; 13: 378-386Crossref PubMed Scopus (339) Google Scholar, 3Käll L. Storey J.D. MacCoss M.J. Noble W.S. Assigning significance to peptides identified by tandem mass spectrometry using decoy databases.J. Proteome Res. 2008; 7: 29-34Crossref PubMed Scopus (449) Google Scholar, J.E. Gygi S.P. of with tandem mass spectrometry for large-scale protein The Proteome Res. 2003; PubMed Scopus Google Scholar) and to to and incorrect PSMs by a database search algorithm L. J.D. Noble W.S. MacCoss M.J. A machine learning for peptide identification from shotgun proteomics Methods. 2007; 4: PubMed Scopus Google Scholar, M. Bottou L. L. Noble W.S. to the algorithm for peptide identification from shotgun proteomics data Proteome Res. 2009; 8: PubMed Scopus Google Scholar, H. Nesvizhskii A.I. validation of peptide identifications in mass spectrometry-based proteomics.J. Proteome Res. 2008; 7: PubMed Scopus Google Scholar). In the current we a decoy database by the in each then the and decoy and we search each the combined decoy a of scoring peptides for each For the of the proteins are as decoy proteins are as are a set of spectra = and a database of and decoy proteins which we a database The search a set of the set of peptides as = the PSMs are as each a match of peptide to in general, we may to the single scoring peptide for each or a of PSMs of the identified peptides to or more to a set of proteins = that the set of Thus, protein in that has at identified peptide the set of proteins that can the For we a for for this which is in a variety of of the of the peptide-spectrum as as that of the and of the peptide. The model of with to peptides, and proteins the of a PSM to be a of such as A. Nesvizhskii A.I. Kolker E. Aebersold R. Empirical statistical model to estimate the accuracy of peptide identification made by MS/MS and database search.Anal. Chem. 2002; 74: 5383-5392Crossref PubMed Scopus (3928) Google Scholar) and L. J.D. Noble W.S. MacCoss M.J. A machine learning for peptide identification from shotgun proteomics Methods. 2007; 4: PubMed Scopus Google used a of of the a of by are the for the hidden and is the hidden as and are the and for the hidden The of is a that can be by This is the improved model used in M. Bottou L. L. Noble W.S. to the algorithm for peptide identification from shotgun proteomics data Proteome Res. 2009; 8: PubMed Scopus Google Scholar). this we a of hidden units. In preliminary we that or hidden the using hidden to evidence of a single peptide can have several spectra to it we the of a peptide as the to of is the set of PSMs to peptide the the PSMs for each peptide because of the in A.I. A. Kolker E. Aebersold R. A statistical model for identifying proteins by tandem mass Chem. 2003; 75: PubMed Scopus Google that many spectra the peptide are an of the of the the of a protein is in of the of the peptides in that protein as is the set of peptides in protein is the set of peptides in the protein that the MS/MS and is a of the The set is by the protein database with the used to the protein mixture for the mass spectrometry that the peptides have in the of in the peptides, with the that we are to an the to the be as a of the we to in this the of peptides, rather than the of peptides as a because the peptide an of many peptides in the protein but have been by This information to which are more likely to matches the database search = sets of peptides to a In we the = it in validation experiments a protein that on the decoy For each protein we have a it is a or decoy set of proteins and corresponding the goal is to the of the and find we search for the in the that the The of the is using a which the the of and the the using with the C. 20: Scopus Google Scholar). the of the are with to each and the are the is a of by The is in the of the that the PSM are because the PSM is of the protein are the of the learning In this we for peptide and PSM level For peptide we a similar to the protein level we a peptide example, and we this peptide a on the of the corresponding proteins. then a to the on the peptide = Similarly, for PSM level we the at the PSM = In learning we that the data used to the model is from the data used to the to a protein for a data we a that and a collection of we identify in the and we subdivide the into that are in the then a model using of the as a set and as the and we this using each as In the we the proteins from the a on the set of proteins. The which this is as of the the set of proteins identified by we all proteins that are to the spectra, as in analysis accuracy and Proteome Res. 2007; 6: PubMed Scopus Google for protein we into a single all the proteins such that A in of peptide and we A. For that in several a assigning these peptides in a to a single that In for the of with we used the method to all of the in the proteins that than proteins with artificially set to by the For the resulting set of we then and them on or The on the this we the J.D. A approach to false discovery R. Soc. 2002; Scopus Google Scholar) as a statistical confidence to each we a and to PSMs with than as then the is as the of PSMs that are incorrect the peptide in the mass when the The is as the at which a PSM is that the is a general statistical confidence that is to the method for evaluating (1Moore R.E. Young M.K. Lee T.D. Qscore: An algorithm for evaluating Sequest database search results.J. Am. Soc. Mass Spectrom. 2002; 13: 378-386Crossref PubMed Scopus (339) Google Scholar). by using decoy PSMs L. Storey J.D. MacCoss M.J. Noble W.S. Assigning significance to peptides identified by tandem mass spectrometry using decoy databases.J. Proteome Res. 2008; 7: 29-34Crossref PubMed Scopus (449) Google Scholar). the of PSMs and the of decoy PSMs For a the of PSMs is = = and the of decoy PSMs is = = can estimate the at a as The to is then as shown in different data sets from and data sets in L. J.D. Noble W.S. MacCoss M.J. A machine learning for peptide identification from shotgun proteomics Methods. 2007; 4: PubMed Scopus Google and E. D. MacCoss M.J. analysis of from distinct proteomics using of For all of the data the peptides to spectra using the of the algorithm L. MacCoss M.J. Noble W.S. and peptide identification from tandem mass Proteome Res. 2008; 7: PubMed Scopus Google with a of to and mass for of The for and set to and The search a decoy database for each of all and The PSMs for each for the search on of the data with C. and two The of and of The first data set of spectra from a of an and using a reverse phase The spectra a protein database of the from The database search on this data set in a protein data set of proteins and The two data sets in a similar from the but using different and The database search them in and and data sets of proteins and and proteins and The data set is from a C. by and to the data The data set from a analysis of C. proteins spectra, from which spectra spectra a protein database of the from C. and This set which in the protein set of with and the and data sets of from and The protein database of from spectra and which in protein data sets of and and and For the validation of we used protein sets identified by Lee Hengartner M.R. Young the of 1998; Full Text Full Text PDF PubMed Scopus Google Scholar) and protein experiments A. J.S. analysis of protein in 2003; PubMed Scopus Google Scholar). The applied to the data all proteins than the to the and all proteins by and with to the protein The of these of used in the validation For all of the proteins in the data a protein a if it the proteins in the validation set and a false The of proteins on or and the generated on this The set used to the proteins identified by and by in In the sets of proteins by and experiments used to the of and in A.I. A. Kolker E. Aebersold R. A statistical model for identifying proteins by tandem mass Chem. 2003; 75: PubMed Scopus Google analysis accuracy and Proteome Res. 2007; 6: PubMed Scopus Google Scholar, protein with peptide identification Proteome Res. 2009; 8: PubMed Scopus Google and using the data sets that more proteins than and a wide of false discovery rates and all data For example, at an of more proteins than with and more than with for the data the and data and more than with and with and and more than with and with for a model on data set and then the model to a separate data to a of the in are on and on the data However, and when we the data into and on of the on the set than in and evidence that is the because the on the set is similar to the on the we that search the of the that we for In to decoy we the of and to proteins that been identified in phase using methods. For this we a set of proteins in phase is by by analysis Lee Hengartner M.R. Young the of 1998; Full Text Full Text PDF PubMed Scopus Google into the and the of the resulting protein with an and the sequence of protein into the and the resulting A. J.S. analysis of protein in 2003; PubMed Scopus Google Scholar). For all of the proteins in the data a protein a if it the proteins in the validation the data the of proteins in is more with these proteins than that in used the to the proteins by and experiments A. J.S. analysis of protein in 2003; PubMed Scopus Google Scholar) to the to which with protein that when proteins at of 1% are by the of the is with proteins. the the proteins identified by and we the the sets of proteins identified as by the two at a of false discovery rates in For all data and identify many of the proteins. the of the sets in the data sets identified by and at of 1% by them the proteins identified by the For the of proteins identified by the for and for For these and and for and Thus, on these data the validation more the identifications than the identifications for we proteins identified by and in the data A previous of these data sets E. D. MacCoss M.J. analysis of from distinct proteomics using of Scholar) that and at the proteins but that the of a proteins significantly the two focused on a of proteins of the proteins shown to have the and This protein identified with confidence by and in the However, in the the protein identified with confidence by this a low of that the proteins to be in and that that this protein is in it likely that the in for the is of this identification that identified by in the on a single scoring peptide with sequence peptide a very to and made a to the protein The of the peptides that to the of the the that to that the and from from the to which is to in In this peptide to the of proteins and this in of this we the that identified in by and and that the peptide sequence to the identification in this the spectra for the peptide sequence identification in and for the peptide sequence identification in and that these spectra most of the in in two peptide that the identification of the in by and the identification in and, by in the all but a single identified peptide either or to in the the identification of in the a single because it on scoring peptide. This that is proteins with a single identification than which the protein more for the of peptides with low The of this identification by with evidence that at two peptides protein many identifications False discovery rates of protein A the Proteome Res. 2009; 8: PubMed Scopus Google Scholar). In general, a protein identification on a single peptide identification can a of false is to identify proteins on a single scoring peptide because the model the protein by the of peptides in the on proteins of an example, we the and the of proteins in the and in the that identified by at 1% with the proteins identified by at the confidence level. For which peptides with than for protein the proteins identified on a single peptide for the of this For which low peptides if we the proteins with a single scoring peptide as more proteins than in and the proteins identified by have on In the the of proteins is in with the of of the identified by the Similarly, in the the proteins identified by have an of in with the of of the proteins identified by the Thus, by with to the of peptides in the proteins and proteins. focused on a single the of proteins identified from a shotgun proteomics This approach with previous of machine learning to this task A. Nesvizhskii A.I. Kolker E. Aebersold R. Empirical statistical model to estimate the accuracy of peptide identification made by MS/MS and database search.Anal. Chem. 2002; 74: 5383-5392Crossref PubMed Scopus (3928) Google Scholar, L. J.D. Noble W.S. MacCoss M.J. A machine learning for peptide identification from shotgun proteomics Methods. 2007; 4: PubMed Scopus Google Scholar, M. Bottou L. L. Noble W.S. to the algorithm for peptide identification from shotgun proteomics data Proteome Res. 2009; 8: PubMed Scopus Google Scholar, Li W. Noble W.S. A algorithm for the of shotgun peptide in machine of peptide MS/MS spectra and Proteome Res. 2003; PubMed Scopus Google Scholar, J.E. Gygi S.P. protein identification by machine learning from a of tandem mass 2004; PubMed Scopus Google which at the level of PSMs or In general, on optimization or the will on the goal of the proteomics However, in it may be to of peptide and protein the model involves peptide level and PSM level scoring as a of the protein level we can the of on the peptide or PSM identification Moreover, we can the algorithm to directly on the peptide or PSM for the of of at the or protein level. are at the and protein on the and and two data sets with The that protein level optimization as in of peptide and PSM identification as the peptide level and PSM level that the and PSM level optimization tasks are the solution to of the tasks may be improved when to information about the solution to the This is for example, the protein task level information about the of all peptides to the if the goal of the experiment is to peptide it is to this task by the protein level a general optimization the of the protein level is that it more of The optimization and PSM level closely related but are likely to have different optimal solutions. and that when the end goal is to find the optimal solution to the protein level it is beneficial to directly solve the problem of algorithms for inferring a set of proteins from a collection of PSMs the problem into two stages: assessing the of the PSMs and then inferring the protein set A.I. A. Kolker E. Aebersold R. A statistical model for identifying proteins by tandem mass Chem. 2003; 75: PubMed Scopus Google Scholar, P. P. H. in protein inference from shotgun proteomics using peptide of the on 2007; Google Scholar, analysis accuracy and Proteome Res. 2007; 6: PubMed Scopus Google Scholar, Li Y. P. Q. H. A approach to protein inference problem in shotgun M. L. of the on Scopus Google Scholar). that the protein identification problem in this in a of information the of the For example, typically a of spectra are to a peptide the peptide identification information about the spectra is to the protein identification if at most peptide is to each and if for a that to be then information about the peptide is the protein identification if the of the match a peptide and a is using a single such as the by then information about the peptide matches the is In contrast, the machine learning approach directly the of identified into all information to the possible and from to the and with