SCFIA: a statistical corresponding feature identification algorithm for LC/MS
© Cui et al; licensee BioMed Central Ltd. 2011
Received: 15 June 2011
Accepted: 11 November 2011
Published: 11 November 2011
Identifying corresponding features (LC peaks registered by identical peptides) in multiple Liquid Chromatography/Mass Spectrometry (LC-MS) datasets plays a crucial role in the analysis of complex peptide or protein mixtures. Warping functions are commonly used to correct the mean of elution time shifts among LC-MS datasets, which cannot resolve the ambiguity of corresponding feature identification since elution time shifts are random. We propose a Statistical Corresponding Feature Identification Algorithm(SCFIA) based on both elution time shifts and peak shape correlations between corresponding features. SCFIA first trains a set of statistical models, and then, all candidate corresponding features are scored by the statistical models to find the maximum likelihood solution.
We test SCFIA on publicly available datasets. We first compare its performance with that of warping function based methods, and the results show significant improvements. The performance of SCFIA on replicates datasets and fractionated datasets is also evaluated. In both cases, the accuracy is above 90%, which is near optimal. Finally the coverage of SCFIA is evaluated, and it is shown that SCFIA can find corresponding features in multiple datasets for over 90% peptides identified by Tandem MS.
SCFIA can be used for accurate corresponding feature identification in LC-MS. We have shown that peak shape correlation can be used effectively for improving the accuracy. SCFIA provides high coverage in corresponding feature identification in multiple datasets, which serves the basis for integrating multiple LC-MS measurements for accurate peptide quantification.
Liquid Chromatography-Mass Spectrometry/Tandem Mass Spectrometry (LC-MS/MS) is a powerful tool for protein identification and quantification . One important task in LC-MS/MS processing is the identification of corresponding features (peaks registered by identical peptides) in multiple datasets, which is critical for the integration of quantification information to reduce measurement variation .
Before other discussions, we first introduce some definitions that are used throughout the paper. A feature is the two dimensional (retention/elution time - m/z) signal registered by a single charge variant of a peptide. When we consider extracted-ion-chromatograms (XICs), a feature is represented by its LC elution peak in an LC-MS/MS run. If a peptide is picked up by Tandem MS, then its LC elution peak can be located exactly in LC-MS. We refer to such LC peaks as "features with identity". If a peptide is not picked up by Tandem MS, then its elution peak location would be unknown, and its LC peak is called "a feature with unknown identity".
If several datasets are collected in an experiment, then each dataset has an associated list of Tandem MS identified peptides. We simply refer to the peptides associated with a dataset Q1, for example, as Q1 peptides. The union of all peptides from all datasets is noted as the "union peptide set". When corresponding features of a peptide is found in all datasets, we say that the peptide is "completely identified for quantification", or simply "completely identified/quantified" in different context.
Current alignment approaches focus on correcting the mean of elution time shifts between datasets using warping functions. Warping function based methods can be categorized as profile- or feature-based. Profile-based approaches align total-ion-chromatograms (TIC) or higher-resolution profiles based on the full, unprocessed data obtained in LC-MS experiments. The most basic profile-based methods compare the difference in the TICs . A method called correlation optimized warping (COW) was proposed by Nielsen . Bylund proposed many modifications to COW . Parametric time warping (PTW) was proposed by Eilers . Van showed an extension of PTW called semi-parametric time warping (STW) . Prince generated the warping function based on dynamic time warping with a one-to-one (bijective) smooth warp-function called Obi-warp .
Feature-based approaches focus on either aligning chromatogram peaks, aligning features or significant features in images [9, 10]. In an initial feature detection step, these approaches try to distinguish relevant features of peptides and irrelevant noise in the data. Among these methods, a very sophisticated algorithm called LCMSWARP has been published by Jaitly . Another paper  compared six freely available alignment algorithms, and found that OpenMS  performs the best on both proteomics and metabolomics data. Most recently, Voss  proposed a method which combines hierarchical pairwise correspondence estimation with simultaneous alignment and global retention time correction. Voss's paper focuses on the alignment of multiple datasets at the same time. However, the performance is slightly worse than that of OpenMS on proteomics data.
To address the proposed problem, we develop a Statistical Corresponding Feature Identification Algorithm (SCFIA) which identifies corresponding features not only based on matching elution times but also elution peak shapes. We build statistical models which can be used to evaluate the probability of candidate feature pairs as corresponding ones. The identification of corresponding features can be applied to various LC-MS datasets under different experimental conditions without user supplied information. Testing results show that SCFIA improves accuracy and complete quantification coverage significantly.
The proposed algorithm is designed for instruments with high mass resolution which have very few overlapping LC elution peaks within XICs. For example, with a mass resolution of 60, 000FWHM on a Orbitrap instrument, there are very few overlapping elution peaks using a mass window of 10 parts-per-million (ppm) for extracting XICs, and the proposed algorithm can be applied. Such a resolution and mass accuracy is routinely available nowadays.
20090608_Orbi6_TaGe_SA_TUMOR_5mix1_01.raw (Group1 Q1)
20090608_Orbi6_TaGe_SA_TUMOR_5mix1_02.raw (Group1 Q2)
20090608_Orbi6_TaGe_SA_TUMOR_5mix1_03.raw (Group1 Q3)
the warping function is non-linear; and
the elution peaks are crowded.
Comparing to the simple protein mix datasets in , where LC elution peaks are sparse, the super-SILAC datasets are more complex in protein composition, which lead to crowded XICs because many peptides with similar masses are eluted out within a short period of time. If we only correct the mean of time shifts between corresponding features, many peaks will be wrongly matched because there are LC peaks in the close vicinity of the true corresponding ones.
Tandem peptide identification
We use X!Tandem  in Trans-Proteomic Pipeline (TPP)  and MaxQuant for Tandem MS identification. In both TPP and X!Tandem, we select the International Protein Index(IPI)-human database version 3.68 as the source of protein sequences. In the TPP, X!Tandem with Kscore is applied as the search engine. Parent mass and fragment ions are searched with maximal mass errors of 7ppm and 0.5 Dalton respectively. Methionine oxidation and N-terminal acetylation are considered as variable modifications and cysteine carbamidomethylation is selected as the fixed modification. SILAC labeling is also considered as a variable modification. In the analysis, the minimum length of peptides is set to 6, and the maximum number of missed cleavage sites is set to 2. Finally PeptideProphet  of TPP is used to validate the search results, and peptides are annotated with PeptideProphet true positive probabilities (short noted as PeptideProphet probabilities). MaxQuant (version 22.214.171.124) is used with the same settings as that of the X!Tandem, except that the validation step is done by the decoy method. The IPI human database is decoyed by Andromeda , and the false discovery rate is set to 0.01.
Ground truth list generation
In each dataset, select one retention time for each unique peptide identification. Sometimes, we find that a unique peptide is identified multiple times. In such cases, we pick the identification with the highest PeptideProphet probability.
We filter peptides by applying a PeptideProphet probability threshold.
We select peptides that are identified in both Q1 and Q2 to form the ground truth list with information of retention_time_sec, m/z value, and peptide sequence.
The ground truth list is further divided to a training and a testing set. The training set is used for statistical model training and the testing set is used for performance evaluation. Since features with higher intensities are less corrupted by noise, features with top 20% intensities are selected to form the training set. In Q1 and Q2 datasets from Group1, the training set contains 270 peptides, and the testing set contains 1425 peptides, which are annotated with their retention times in both Q1 and Q2.
Note that a pair of non-corresponding features can be obtained by replacing one of the features in a corresponding pair with a random feature from the same XIC of the replaced one. In this way, we can construct a non-corresponding feature training set.
Note that the higher the PeptideProphet probability threshold is, the purer the ground truth list. The threshold will affect the calculated accuracy in corresponding feature identification. For example, at a threshold of 95%, at least around 5% of the testing peptides are false positives, which cannot be matched to their LC peak intervals recorded in the ground truth. Consequently, the calculated corresponding feature identification accuracy can not exceed 95% significantly.
In contrast to the ground truth list selection process in , we do not filter features based on retention time to avoid introducing bias to the training set. The "pureness" of the ground truth list is controlled by the threshold on PeptideProphet probability. The threshold can be raised to reduce the number of outliers.
Performance evaluation based on the testing set
Before we describe the algorithm, we want to clarify the performance evaluation method used in this paper. After we get the testing set, we pretend that we know the identities of the testing peptides in Q1, but not in Q2. We then apply SCFIA. If an identified corresponding feature has an elution time that differs from what has been recorded in the ground truth of Q2, then an error is registered. Finally, we calculate accuracy as the ratio between the total number of correctly identified corresponding pairs over the total number of peptides in the testing set. Note that this accuracy measurement is equivalent to the precision rate in  when considering pair-wise alignment.
In SCFIA, we first use the training set to construct the statistical models. We then evaluate the performance based on the testing set. Finally, we compare the performance SCFIA to that of OpenMs (which is described as the best ), and a warping function based method (Gwarping).
Statistical corresponding feature identification algorithm (SCFIA)
Step 1: Pre-processing of LC/MS data
To identify possible LC peak intervals for a given peptide in both Q1 and Q2, we first calculate its XICs at its mono- and first isotope m/z values in the charge state that it has been identified in Tandem MS.
We use the XIC at the higher isotope position to detect up to n high intensity regions by applying a threshold at three times the background noise standard deviation above the median noise level. Only one interval corresponds to the elution interval of the peptide. In Q1, the exact interval is known by selecting the interval that includes the retention_time_sec recorded in the ground truth.
In Q2, we employ the same process as that in Q1. However, without identification information, the exact elution interval is unknown, and we treat all detected intervals as corresponding feature candidates, which should include the true corresponding one.
Given an identified LC peak interval in Q1, there are the n candidates in Q2, which form n candidate corresponding feature pairs.
Step 2: Mean elution time shift correction
To evaluate the performance of Gwarping, we first use the Gwarping function for mean time shift correction, then we assign the nearest features in Q2 as corresponding ones. We find that the alignment performance of Gwarping exceeds that of OpenMS, which is considered as the best in . This suggests that warping based on Tandem MS identification is reliable. Due to this reason, we use Gwarping as a representative warping function based method to compare with SCFIA.
Step 3: SCFIA models
In the third step, we build statistical models of corresponding features.
Parameters considered in the model
After pre-processing both Q1 and Q2, we obtain a training set of corresponding features, based on which, we can train our statistical models. The parameters considered are elution time shift and LC peak shape correlation between corresponding features. These two parameters are independent. Elution time shift is mainly affected by varying experimental conditions, and LC peak shape depends on the physicochemical characteristics of a peptide.
Elution time shift has been used as the most important parameter for LC peak alignment traditionally. In SCFIA, the time shift is assumed to have a Gaussian distribution  after mean correction, whose parameters can be estimated from the training set.
LC peak shape of peptides is another important parameter. Under similar experimental conditions, identical peptides form similar LC peaks, while different peptides form different LC peak shapes. Similarity between two LC peaks can be measured by the R2 statistics, which indicates how well a regression line approximates the observed data points. An R2 of 1.0 means that the regression line perfectly fits the data, while 0 means the poorest fit. For details please see . When we regress an LC peak in Q1 to one in Q2, the resulted R2 statistic is noted as the alignment R2(AR) statistics. ARs can be calculated using the Matlab function regress(·).
where P(y) represents the probability that the considered feature pair with time shift AT and peak shape correlation AR is corresponding. Both P(AT) and P(AR) are given by the statistical models we constructed from the training set. Our goal is to find the corresponding feature pair that maximizes the likelihood probability function in (1).
Elution time shift model
In Figure 4, we plot the warping function estimated from elution time shifts of corresponding features from the training set of Q1 and Q2 from Group1. After applying the warping function, we calculate the remaining time shifts between corresponding features.
In the past, the time shifts are assumed as Gaussian , and we can write P(AT | μ, σ2) ~ N(μ, σ2), where μ represents the mean and σ2 the variance. We estimate (μ, σ2) from the remaining time shifts between corresponding features using the Matlab function normfit(·).
AR statistic model
where k and θ are parameters of the Gamma distribution, which can be estimated using the Matlab function gamfit(·). In Figure 6(a), we plot the fitted Gamma distribution with the normalized histogram of AR for corresponding features. The normalized histogram and fitted model of AR for non-corresponding features are plotted in Figure 6(b). In Figure 6(c), we compare the difference in fitted distributions of AR between corresponding and non-corresponding features. We can see a notable difference in this example. Note that Group1 is composed of datasets from different peptide fractions, thus there exist significant concentration variations which do not lead to significant deterioration of peak shape correlations between corresponding features. This indicates that AR is a valuable parameter for corresponding feature detection. The fitted Gamma distribution is then used as P(AR) in (1).
Step 4: Estimate probabilities of candidate corresponding feature pairs
In the fourth step, between any pair of candidate features, we first calculate its AT and AR, which are plugged in (1) subsequently. The candidate pair with the highest likelihood probability will be reported as the corresponding one.
Accuracy in corresponding feature detection
To compare the performance of SCFIA with other methods, we use the testing ground truth list of Q1 and Q2 from Group1. We apply various algorithms for corresponding feature identification. When applying SCFIA, we use a PeptideProphet probability threshold of 95% and a 10ppm mass window for calculating the XICs. This mass window is selected based on the mass accuracy of the instrument, which should be adjusted for different instruments.
Comparison with OpenMS
OpenMS(Version 1.7.0) is evaluated to have the best performance in . The details of the simulation process of OpenMS can be found in the [Additional file 1]. We set RT to two possible values, 500 and 700, while MZ is set as 0.01. We try different settings to ensure the best result. OpenMS achieves a 79.79% and 80.35% accuracy under two different settings which have little difference.
Comparison with Gwarping
To estimate the improvement of SCFIA over warping function based methods, we want to compare the performance of SCFIA to Gwarping. After applying the Gwarping function, the elution time of each peptide in Q1 is mapped to Q2, then the LC peak which is the closest to the mapped time point is considered as the detected corresponding feature in Q2. By employing this simple method, the accuracy is 89.89%, which is higher than that of OpenMS. This result is not surprising because OpenMS does not consider non-linear warping functions. There are a total of 144 peptides that are not aligned correctly out of 1425 testing peptides. We inspect manually and find that these peptides have interfering LC peaks that are closer to the mapped time points than the true corresponding ones in Q2. The proportion of such peptides strongly depends on experimental settings. If shorter elution time is desired, then more peptides will have close neighbors, and warping function based methods will be less effective in finding corresponding features.
Performance of SCFIA
Corresponding feature identification accuracy
Testing set of Q1 and Q2 from Group1
We manually inspect the 49 peptides that are not aligned by SCFIA. We find that for these 49 peptides, their corresponding features specified by the ground truth do not agree in elution time and peak shape as well as interfering ones. We show an example in [Additional file 1]. We suspect that these peptides are false positives in Tandem MS identification. If this assumption is true, we should be able to observe an increased accuracy rate as we raised the threshold on PeptideProphet probability.
Corresponding feature identification accuracy v.s. PeptideProphet probability threshold
Testing set of Q1 and Q2 from Group1
No. of peptides tested
Accuracy of SCFIA
Corresponding feature identification accuracy
Testing set of different dataset pairs
Q1 Q2 Grp.1
Q2 Q3 Grp.1
Q1 Q3 Grp.1
Q1 Q2 Grp.2
Q2 Q3 Grp.2
Q1 Q3 Grp.2
Complete quantification coverage
We employ the following procedure for the complete identification of all peptides in the union set. We first select Q1 peptides, and find their corresponding features in Q2 if their identities are unknown. Then the same procedure is repeated from Q1 to Q3. Subsequently, we focus on Q2 peptides, and we find all corresponding features in Q1 and Q3 if their identities are not known yet. Lastly, we focus on Q3 peptides. This procedure is repeated for all peptides in all datasets with unknown identities until complete identification. Using this procedure, a total of 11124 (93.78%) peptides are completely identified in all three datasets, 590 are identified in at least two datasets.
We then investigate Group2 datasets (Q1, Q2, Q3) from three technical replicates. The PeptideProphet probability threshold we choose is still 0.95. In Figure 8(c), we show that the union has 9239 peptides, and the intersection has 3467. For testing purposes, peptides detected in different charge states and different datasets are removed from the list, which leaves 5628 peptides without complete identifications. Using SCFIA, a total of 5655 (99.52%) peptides are completely identified in all three datasets, and 22 are identified in at least two datasets. Since Group2 datasets are from replicates, a higher complete identification rate is expected than that of Group1.
With complete identification, these peptides can be quantified completely. Since peptide quantification is a lengthy topic, we leave it out of this paper.
Comparison with MaxQuant
MaxQuant  is a popular tool that provides both Tandem MS identification and quantification. We want to compare the peak identification coverage of SCFIA with that of MaxQuant. To this end, we employ MaxQuant (Version 126.96.36.199) to process super-SILAC datasets Q1, Q2, and Q3 in Group1 and Group2. The size of peptide identification results is summarized in the Venn-diagram in Figure 8(b). We can see that the union set of Group1 contains a total of 10511 peptides, and the intersection between them is 700. Thus based on Tandem MS identification information, only 556 peptides can be completely quantified in all three datasets. In contrast, after applying SCFIA, a total of 8938 peptides are identified in all three datasets in the first group.
The same process is repeated in Group2, and the results are reported in In Figure 8(d). Significant advantage of SCFIA is reported again.
Complete identification coverage
Complete Identification Coverage
Xtendem (Data Group1)
Maxquant (Data Group1)
Xtendem (Data Group2)
Maxquant (Data Group2)
Q1 ∪ Q2 ∪ Q3-
Q1 ∩ Q2 ∩ Q3
SCFIA identified completely
SCFIA identification coverage
Through testing, we can see that SCFIA can be applied in the alignment of both technical replicates and datasets collected from different LC/MS runs.
The accuracy results based on Group1 and Group2 data suggest that SCFIA is more effective when there are high elution time and concentration variations. In such cases, using peak shape correlation improves the performance. However, the improvement changes with experimental conditions. When elution time variation is small, and there exists long gaps between elution peaks, then alignment based on elution time is sufficient. However, when elution time variation is large, and gaps between LC peaks are small, peak shape correlation becomes useful in performance improvement. Users can always decide if using peak shape correlation will provide performance gain by inspecting the ROC curves estimated by SCFIA. In experiments where peak shape reproducibility is not strong, or when the XICs are not crowded, then it may be sufficient or necessary to use AT alone.
SCFIA requires a number of "common" identifications for training the statistical models. Generally the more common identifications the better. Preferably, there are around 200 common identifications. We observe no obvious difference in performance in our experiments when the size of the training set increases beyond this number.
In this paper, we propose a new Statistical Corresponding Feature Identification Algorithm (SCFIA) for the identification of corresponding features in different LC-MS/MS datasets. The main innovation of the algorithm is the use of statistical models for both elution time shifts and peak shape correlations, which provides maximum likelihood estimates of corresponding features. The algorithm allows accurate corresponding feature identification with crowded elution profiles. We verify the algorithm on two groups of super-SILAC datasets, and the performance is shown to be better than warping function based methods including OpenMS. SCFIA is shown to have very high accuracy in corresponding feature identification and the performance is near optimal.
SCFIA can be utilized for the complete identification of elution peak intervals of Tandem MS identified peptides in multiple datasets. We have verified that SCFIA provides high coverage in complete identification which will lead to more accurate quantification in differential analysis for biomarker discovery.
Availability and Requirements
Project name: SCFIA project; Operating system(s): Windows XP/vista/7; Programming language: Matlab; Licence: GNU GPL; Any restrictions to use by non-academics: licence needed. The related material including the testing dataset can be found at the project webpage http://compgenomics.utsa.edu/SCFIA.html.
This work is supported by San Antonio Life Sciences Institute Research Enhancement, and a grant from National Institute of Health (NIH 2G12RR013646-11). We thank the Computational Biology Initiative (UTSA/UTHSCSA) for providing access and training to the analysis software used.
- Aebersold R, Mann M: Mass spectrometry-based proteomics. Nature 2003, 422(6928):198–207. 10.1038/nature01511View ArticlePubMedGoogle Scholar
- Vandenbogaert M, Li-Thiao-Té S, Kaltenbach H, Zhang R, Aittokallio T, Schwikowski B: Alignment of LC-MS images, with applications to biomarker discovery and protein identification. Proteomics 2008, 8(4):650–672. 10.1002/pmic.200700791View ArticlePubMedGoogle Scholar
- De Boor C: A practical guide to splines. Springer Verlag; 2001.Google Scholar
- Nielsen S, Andersen J, Hvelplund P, Jørgensen T, Sørensen M, Tomita S: Triply charged bradykinin and gramicidin radical cations: their formation and the selective enhancement of charge-directed cleavage processes. International Journal of Mass Spectrometry 2002, 213(2–3):225–235. 10.1016/S1387-3806(01)00543-7View ArticleGoogle Scholar
- Bylund D, Danielsson R, Malmquist G, Markides K: Chromatographic alignment by warping and dynamic programming as a pre-processing tool for PARAFAC modelling of liquid chromatography-mass spectrometry data. Journal of Chromatography A 2002, 961(2):237–244. 10.1016/S0021-9673(02)00588-5View ArticlePubMedGoogle Scholar
- Eilers P: Parametric time warping. Anal Chem 2004, 76(2):404–411. 10.1021/ac034800eView ArticlePubMedGoogle Scholar
- Van Nederkassel A, Xu C, Lancelin P, Sarraf M, MacKenzie D, Walton N, Bensaid F, Lees M, Martin G, Desmurs J, et al.: Chemometric treatment of vanillin fingerprint chromatograms:: Effect of different signal alignments on principal component analysis plots. Journal of Chromatography A 2006, 1120(1–2):291–298. 10.1016/j.chroma.2005.11.134View ArticlePubMedGoogle Scholar
- Prince J, Marcotte E: Chromatographic alignment of ESI-LC-MS proteomics data sets by ordered bijective interpolated warping. Anal Chem 2006, 78(17):6140–6152. 10.1021/ac0605344View ArticlePubMedGoogle Scholar
- Walczak B, Wu W: Fuzzy warping of chromatograms. Chemometrics and Intelligent Laboratory Systems 2005, 77(1–2):173–180. 10.1016/j.chemolab.2004.07.012View ArticleGoogle Scholar
- Katajamaa M, Orešič M: Processing methods for differential analysis of LC/MS profile data. BMC bioinformatics 2005, 6: 179. 10.1186/1471-2105-6-179PubMed CentralView ArticlePubMedGoogle Scholar
- Jaitly N, Monroe M, Petyuk V, Clauss T, Adkins J, Smith R: Robust Algorithm for Alignment of Liquid Chromatography- Mass Spectrometry Analyses in an Accurate Mass and Time Tag Data Analysis Pipeline. Anal Chem 2006, 78(21):7397–7409. 10.1021/ac052197pView ArticlePubMedGoogle Scholar
- Lange E, Tautenhahn R, Neumann S, Gröpl C: Critical assessment of alignment procedures for LC-MS proteomics and metabolomics measurements. BMC bioinformatics 2008, 9: 375. 10.1186/1471-2105-9-375PubMed CentralView ArticlePubMedGoogle Scholar
- Sturm M, Bertsch A, Gröpl C, Hildebrandt A, Hussong R, Lange E, Pfeifer N, Schulz-Trieglaff O, Zerck A, Reinert K, et al.: OpenMS - An open-source software framework for mass spectrometry. BMC bioinformatics 2008, 9: 163. 10.1186/1471-2105-9-163PubMed CentralView ArticlePubMedGoogle Scholar
- Voss B, Hanselmann M, Renard B, Lindner M, Köthe U, Kirchner M, Hamprecht F: SIMA: Simultaneous Multiple Alignment of LC/MS Peak Lists. Bioinformatics 2011.Google Scholar
- Meek J, Rossetti Z: Factors affecting retention and resolution of peptides in high-performance liquid chromatography. Journal of Chromatography A 1981, 211: 15–28. 10.1016/S0021-9673(00)81169-3View ArticleGoogle Scholar
- Bellew M, Coram M, Fitzgibbon M, Igra M, Randolph T, Wang P, May D, Eng J, Fang R, Lin C, et al.: A suite of algorithms for the comprehensive analysis of complex protein mixtures using high-resolution LC-MS. Bioinformatics 2006, 22(15):1902. 10.1093/bioinformatics/btl276View ArticlePubMedGoogle Scholar
- Bantscheff M, Schirle M, Sweetman G, Rick J, Kuster B: Quantitative mass spectrometry in proteomics: a critical review. Analytical and bioanalytical chemistry 2007, 389(4):1017–1031. 10.1007/s00216-007-1486-6View ArticlePubMedGoogle Scholar
- Cox J, Mann M: MaxQuant enables high peptide identification rates, individualized ppb-range mass accuracies and proteome-wide protein quantification. Nature biotechnology 2008, 26(12):1367–1372. 10.1038/nbt.1511View ArticlePubMedGoogle Scholar
- Geiger T, Cox J, Ostasiewicz P, Wisniewski J, Mann M: Super-SILAC mix for quantitative proteomics of human tumor tissue. Nature Methods 2010.Google Scholar
- Duncan D, Craig R, Andrew J: Parallel tandem: a program for parallel processing of tandem mass spectra using PVM or MPI and X! Tandem. Journal of proteome research 2005, 4(5):1842–1847. 10.1021/pr050058iView ArticlePubMedGoogle Scholar
- Deutsch E, Mendoza L, Shteynberg D, Farrah T, Lam H, Tasman N, Sun Z, Nilsson E, Pratt B, Prazen B, et al.: A guided tour of the Trans-Proteomic Pipeline. Proteomics 2010, 10(6):1150–1159. 10.1002/pmic.200900375PubMed CentralView ArticlePubMedGoogle Scholar
- Keller A, Nesvizhskii A, Kolker E, Aebersold R: Empirical statistical model to estimate the accuracy of peptide identifications made by MS/MS and database search. Analytical chemistry 2002, 74(20):5383–5392. 10.1021/ac025747hView ArticlePubMedGoogle Scholar
- Cox J, Neuhauser N, Michalski A, Scheltema R, Olsen J, Mann M: Andromeda a peptide search engine integrated into the MaxQuant environment. Journal of proteome research 2011.Google Scholar
- Palmblad M, Mills D, Bindschedler L, Cramer R: Chromatographic alignment of LC-MS and LC-MS/MS datasets by genetic algorithm feature extraction. Journal of the American Society for Mass Spectrometry 2007, 18(10):1835–1843. 10.1016/j.jasms.2007.07.018View ArticlePubMedGoogle Scholar
- Gelman A: Bayesian data analysis. CRC press; 2004.Google Scholar
- Steel R, Torrie J: Principles and procedures of statistics: a biometrical approach. 1980.Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.