Search tips
Search criteria 


Logo of nihpaAbout Author manuscriptsSubmit a manuscriptHHS Public Access; Author Manuscript; Accepted for publication in peer reviewed journal;
Anal Chem. Author manuscript; available in PMC 2010 May 6.
Published in final edited form as:
PMCID: PMC2865572

Charge Prediction Machine: A tool for inferring precursor charge states of Electron Transfer Dissociation tandem mass spectra


Electron Transfer Dissociation (ETD) can dissociate highly charged ions. Efficient analysis of ions dissociated with ETD requires accurate determination of charge states for calculation of molecular weight. We created an algorithm to assign the charge state of ions often used for ETD. The program, Charge Prediction Machine (CPM), uses Bayesian decision theory to account for different charge reduction processes encountered in ETD, and can also handle multiplex spectra. CPM correctly assigned charge states to 98% of the 13,097 MS2 spectra from a combined dataset of four experiments. In a comparison between CPM and a competing program, Charger (ThermoFisher), CPM produced half the mistakes.

Keywords: ETD, pattern recognition, MudPIT, charge prediction, multiplex spectra


An important element of proteomics is the process of protein identification. A common approach to protein identification involves the use of tandem mass spectrometry. In this process peptide ions are isolated, fragmented using Collision Activated Dissociation (CAD), and the fragment ions are analyzed to determine their mass to charge ratios (m/z). Tandem mass spectra are analyzed through protein sequence database searching, sequence tag analysis, or de novo interpretation. All of these processes require calculation of an accurate molecular weight for the peptide1-3.

Peptides generated by trypsin digestion and ionized by electrospray ionization produce ions of predominately +2 and +3 charge states. Determining the charge state of peptide ions using CAD spectra and other features has been performed by several groups 4-7. When the charge state cannot be determined it is common for a database search to be performed with a tandem mass spectrum for which a molecular weight has been calculated with both charge states. The correct answer and therefore the correct molecular weight is then determined by examining the sequence matches. This situation increases the computational burden for database searching and can increase the difficultly of correctly assigning tandem mass spectra to sequences, but in general this approach has worked reasonably well.

Electron Capture Dissociation (ECD) and Electron Transfer Dissociation (ETD) are new methods to dissociate ions in mass spectrometers. A feature common to both of these methods is a preference for highly charged ions, which generally means larger polypeptides or even proteins can be dissociated with these methods. Both methods cleave randomly along the polypeptide backbone and work well when trying to pinpoint the location of posttranslational modifications. When ETD or ECD is used, the charge states of polypeptides observed can run from +2 to +7 and up; thus, the challenge of calculating molecular weight is increased and existing software for CAD cannot be applied. The alternative to accurate calculation of molecular weight of testing all possible calculations for each potential charge state (“all hypothesis search”) greatly increases the computational overhead associated with the analysis of ETD/ECD spectra and thus is not desirable. Sadygov et al. developed a method (Charger) using linear discriminant analysis with autocorrelation to decipher charge states associated with ETD8. In comparison to analysis with the all hypothesis search, we observed 15% fewer identifications suggesting the method was not fully optimized. We developed a new algorithm, Charge Prediction Machine (CPM), to accurately classify the charge states for ETD spectra. This algorithm is based on Bayesian decision theory and introduces a Relaxation Parameter to solve cases with low confidence, and can deal with multiplexed spectra (different co-fragmented ion species). CPM is shown to efficiently classify charge states up to +7 while greatly reducing the number of searches as compared to the all hypothesis searches.

Methods and algorithm

Preparation of the testing (unlabeled) dataset

The yeast ETD MS2 testing dataset consists of 4 LC-MS/MS experiments. Cells were harvested, lysed, and digested with endoproteinase Lys-C and trypsin as previously described9. (A) Chromatography. Capillary HPLC was performed with an Agilent, Inc model 1200 Quaternary HPLC (Palo Alto, CA). Fused-silica capillary columns (50μm i.d.) with a 2-3μm opening and packed with reversed-phase C18 resin were prepared10. HPLC buffer solutions were: water/acetonitrile/formic acid (95:5:0.1, v/v/v) as buffer A, and water/acetonitrile/formic acid (20:80:0.1, v/v/v) as buffer B. The capillary columns were conditioned with buffer A; then digested proteins were pressure-loaded (~5 to 10μg) on to the column and eluted with a three-hour linear gradient of buffer B from 10-100% into the ion source. (B) Mass Spectrometry. LC-MS/MS experiments were preformed on a linear ion-trap LTQXL-ETD mass spectrometer (ThermoFisher, San Jose, CA) having a chemical ionization source that generates Fluoranthene anions11 for electron transfer dissociation. Mass spectra were acquired using a data-dependent approach where each survey scan (300-2000 m/z) was followed by five ETD MS2 scans of the most intense precursor ions. The automatic gain control (AGC) target for the Fluoranthene anion (m/z 202) was set at a value of 100,000. Ion/ion reaction duration was set at 50ms. Supplemental activation function was applied for all the experiments in this study set. The mass spectrometer scan functions and HPLC solvent gradients were controlled by the Xcalibur data system (ThermoFisher, San Jose, CA).

Preparation of the training (labeled) dataset

The yeast ETD MS2 training dataset consists of 12 LC-MS/MS experiments and was acquired in the Coon laboratory; a detailed description is provided elsewhere12;13. These ETD tandem mass spectra were acquired on a hybrid linear ion trap-Orbitrap mass spectrometer (ThermoFisher, San Jose, CA). An Orbitrap MS scan was used to provide accurate precursor ion m/z assignment and high resolution to help assign charge state. Data-dependent acquisition of ETD was performed on 6 precursor ions which were then analyzed in the linear trap.

Precursor charge assignment in the testing (unlabeled) dataset

MS2 spectra from the testing dataset were extracted using RawExtract14 and assigned charge states by using an all hypothesis search for charge states +2 through +9 for each spectrum. Since the testing dataset was acquired on a low-resolution ion trap instrument, the only way to independently assign precursor charge states was to perform a protein database search in order to match the tandem mass spectra to yeast peptides. Once the peptide spectrum matches (PSMs) were determined, the testing dataset was limited to contain only those peptides identified with extremely high confidence (false discovery rate of less than 0.1%).

The tandem mass spectra were searched against a Saccharomyces cerevisiae protein database containing 5,873 protein sequences, containing the translations of all systematically named ORFs, downloaded as FASTA-formatted sequences from the Saccharomyces Genome Database (database released on December 16, 2005), and 123 common contaminant proteins, for a total of 5,996 target database sequences. In order to calculate confidence levels and false discovery rates, a decoy database containing the reverse sequences of the 5,996 proteins was appended to the target database15, and the SEQUEST algorithm was used to find the best matching sequences from the combined database.

SEQUEST searches were done on an Intel Xeon 80-processor cluster running under the Linux operating system. The peptide mass search tolerance was set to 3 Da. Average masses were used for predicted (M+H)+ values in the search, and monoisotopic masses were used for the predicted fragment ions. The mass of the amino acid Cysteine was statically modified by +57.0 Da, to take into account the carboxyamidomethylation of the sample. No enzymatic cleavage conditions were imposed on the database search, so the search space included all candidate peptides whose theoretical mass fell within the mass tolerance window, regardless of their tryptic status.

The validity of peptide/spectrum matches was assessed in DTASelect16;17 using two SEQUEST-defined parameters, the cross-correlation score (XCorr) and normalized difference in cross-correlation scores (DeltaCN). The search results were grouped by charge state (+2 to +9) and tryptic status (fully tryptic, half-tryptic, and non-tryptic), resulting in 24 distinct sub-groups. In each one of these sub-groups, the distribution of Xcorr and DeltaCN values for (a) direct and (b) decoy database hits was obtained, then the direct and decoy subsets were separated by quadratic discriminant analysis. Outlier points in the two distributions (for example, matches with very low Xcorr but very high DeltCN) were discarded. Full separation of the direct and decoy subsets is not generally possible; therefore, the discriminant score was set such that a false discovery rate of 0.1% was determined based on the number of accepted decoy database peptides. This procedure was independently performed on each data subset, resulting in a false discovery rate independent of tryptic status or charge state.

In addition, a minimum sequence length of 7 amino acid residues was required, and each protein on the list was supported by at least two peptide identifications, with a minimum sequence coverage of 5%. These additional requirements resulted in the elimination of most decoy database and false positive hits, as these tended to be overwhelmingly present as proteins identified by single peptide matches, or with very low sequence coverage. After this last filtering step, the false discovery rate was estimated to have been reduced to below 0.1%.

Precursor charge assignment in the training (labeled) dataset

MS2 spectra from the training dataset were extracted using RawExtract14 and assigned charge states using isotopic information present in the high-resolution full MS scans performed in the Orbitrap analyzer. The Xcalibur software (ThermoFisher, San Jose, CA) was used to assign precursor ion charge states to all spectra. The spectra for which XCalibur did not make an unambiguous charge assignment were removed from the training set. We thus obtained a total of 53,027 charge-assigned tandem mass spectra for the training dataset.

The Charge Prediction Machine

CPM uses three MS2 attributes to predict a spectrum's charge state: the complementary fragment ions, the charge reduced precursors, and the neutral losses. The classification problem is solved in an 18-dimensional feature space having each attribute unfold into 6 dimensions as described below.

Before describing each attribute, the notation used throughout is introduced here. Let S be a mass spectrum, understood as a set of spectral “peaks”, the ith of which characterized by Ii and m/zi, respectively its ion current and its measured mass to charge ratio. The total ion current of S (TIC) is defined as


Let PPM (x, y) be a parts per million indicator between two quantities x and y,


and UPPM a user specifiable parts per million tolerance; such is required for computational purposes. Let H be the mass of a Hydrogen atom.

(A) The complementary ion feature, CIF, is derived from the expectation that fragmenting a +2 precursor ion in CID generates pairs of singly charged complementary product ions whose masses sum up to the mass of the precursor plus that of two hydrogens4;18. ETD favors the production of c- and z*-ions, and therefore the above expectation translates into za(m/zi) + zb(m/zj) = zprecursor(m/zprecursor) − H, constrained by za + zb = zprecursor, i, j [set membership] {1,2,…,|S|}, and za, zb [set membership] {2,…,7}. In practice, accounting for all za and zb combinations that add up to zprecursor increases computational time without substantially increasing discriminatory power. In this regard, CPM focuses on an ordered subset (ς) of combinations that has proven to be effective during the cross-validation tests (data not shown). For the +2 precursor, this subset is ς2 = {+1,+1}; accordingly, ς3 = {+l,+2}, ς4 = {+2 +2}, ς5 = {+2,+3}, ς6 = {+3,+3}, and ς7 = {+3,+4}. In our notation, ςk, the indexer (k) stands for a hypothesized precursor charge state and ςk [j] stands for the jth member of the set.

CPM's CIF is computed for six dimensions, resulting in the features CIF2, CIF3,…,CIF7. For the dimension k, let its expected complementary ion sum, similar to the above, be


Let RSk be the subset {l, m} of peaks from S for which PPM (ςk[1](m/zl) + ςk [2](m/zm), ECSk) ≤ UPPM. Then


where IjRSk=Ij if j [set membership] RSk ( IjRSk=0 otherwise). The expectation is that the precursor's charge will equal the dimension indexer k of the dimension achieving the highest score.

(B) The charge reduced precursor feature, CRPF, is based on evidence that intact charge reduced precursors (CRPs) are frequently observed in MS2 spectra and could be effectively used for charge determination4. Figure 1 shows a typical ETD spectrum of a +5 peptide: the +2, +3, and +4 CRPs are present as dominant peaks in the spectrum, together with the +5 precursor ion itself. We present a methodology to identify CRPs, one that considers a subset (Ψ) of possible CRP charges for a given precursor charge state. Considering all possibilities increases the CRPF computation time without significantly increasing the model's discriminatory power (data not shown). Moreover, some CRPs could lie beyond the detectable m/z bounds. For the +2 precursor, CPM considers the +1 CRPs (Ψ2 = {+1}); accordingly, Ψ3 = {+l,+2}, Ψ4 = {+l,+3}, Ψ5 = {+3,+4}, Ψ6 = {+3,+4,+5}, and Ψ7 = {+4,+5,+6}.

Figure 1
ETD spectrum for +5 charged peptide

The CRP sets were chosen to minimize the overlap among the expected CRP m/z values, denoted by CRP, for which an estimate is


for a precursor charge k reduced to Ψk[j]. For a precursor of expected charge k, let RSk [subset or is implied by] S be such that PPM (m/zl, CRPj) ≤ UPPM for all j [set membership] Ψk and all l [set membership] S. Then CRPFk is given by


(C) The neutral loss feature, NLF, follows from the fact that, in ETD, CRPs frequently lose an uncharged or neutral fragment. CPM's algorithm searches for two common neutral losses, water (H2O, ~18.02 amu) and ammonia (NH3 ~17.03 amu), amounting to the value we call neutralLossMass, in subsets (ξ) of CRP's that expectedly lie within the m/z detection bounds. The chosen CRP subsets are: ξ2 = ξ3 = {+l,+2}, ξ4 = {+l,+3}, ξ5 = {+3,+4}, ξ6 = {+3,+4,+5}, and ξ7 = {+4,+5,+6}. For a given k, let RSk [subset or is implied by] S be such that PPM(m/zl+neutralLossMassξk[j],CR^Pj)UPPM for all j [set membership] ξk and all l [set membership] S, with CRPj defined as above (but on ξk[j]). Then


(D) Formalization of the charge prediction machine. CPM is a semi-supervised learning strategy. The learning process begins by computing an input vector (x) for each mass spectrum in the training dataset following the schema: [left angle bracket]ωi[right angle bracket] [left angle bracket]feature1:value1[right angle bracket][left angle bracket]feature18:value18[right angle bracket]. In the latter, ωi [set membership] {2,3,…7} stands for the class label (a precursor charge state assigned by a specialist; in our case, a combination of high resolution Orbitrap MS1 and software); value1 through value6, value7 through value12, and value13 through value18 correspond to the computed scores for CIF2CIF7, CRPF2CRPF7, and NLF2NLF7, respectively. Feature1 through feature18 correspond to the numbers 1 through 18; this is done to comply with the sparse matrix representation schema that is widely adopted in the pattern recognition community and in the PatternLab for proteomics project, of which this tool is part19.

CPM adopts, for each class ωi, the Bayesian discriminant function


where P(ωi) is the empirically obtained prior probability of class ωi derived from the respective charge state frequency in the training dataset, mu is the mean vector, Σ is the covariance matrix, |Σ| its determinant, and Σ−1 its inverse. CPM stores these variables to disk for quick retrieval and classification of future unseen examples. Classification is performed as follows: for each spectrum, the Bayesian scores are computed, and then re-mapped according to the following procedure:


All results are associated with their respective spectra and saved in a data structure referred to as the candidate solution array. For example, if there were 100 spectra in the classification dataset, and 6 charge states being accounted for, the candidate solution array would have 600 elements. This array is then ordered in a nonincreasing order. Clearly, the most favorable solution for every input vector will obtain a score of 0; solutions with higher scores are associated with a lower expectation of class membership. CPM allows more than one output label per input vector on a case basis by making use of a user-specified Relaxation Parameter (RP). For example, for an RP of 1.5, the first 150 solutions in our example candidate solution array will be assigned to their corresponding input vectors. Clearly, every input vector will hold their highest expectation solution; the following 50 solutions that are also associated with a high degree of truth will be assigned to their respective spectra / input vector.

(E) Accounting for multiplexed spectra. Previous work shows that in truly complex samples, closely situated m/z precursors can jointly be selected for fragmentation20. This event's frequency can greatly vary from sample to sample (e.g., ~2 to ~10% or more in a single phase analysis of a digested yeast cell lysate; multiplexed spectra were observed in both the training and the testing dataset).

Fragmentation usually prevails in one ion species, making the other leave more noticeable charge determination features in the mass spectrum (e.g., CRPs). While the search engine most likely identifies the more abundant peptide ion, CPM assigns a charge state according to the most evident charge features, possibly biased towards the more poorly fragmented ion species. To account for multiplexed spectra, CPM applies a post-processing correction to potentially multiplexed spectra. The correction searches for +2 precursors that co-fragmented with a +3 or +4 precursor (2+3or4) and for +3 precursors that co-fragmented with a +2 or +4 precursor (3+2or4), and includes the extra charge state hypothesis.

The 2+3or4 procedure first selects spectra whose precursor ions are less than 850 m/z and have +2 as their most confident charge state. For each of these spectra, it sums the ion current of an expected +1 CRP peak, assuming the precursor charge is +2 (CRP2to1), and accordingly, CRP3to2, and CRP4to3. The CRP values are obtained as described in the charge reduced precursor feature section. Afterwards, CPM stores the multiplex charge state hypothesis by selecting between the CRP3to2 and CRP4to3 of highest value. Only the hypotheses of highest expectation will be included in the final output as described later.

In example, suppose the CRP3to2 was selected for a given spectrum. CPM then checks whether the +3 charge state hypothesis has not been included during the relaxation procedure and that CRP3to2 is greater than 5% of CRP2to1's value. If both hold true, CPM generates a data structure containing the spectrum's scan number, the +3 charge state hypothesis, the CRP3to2:CRP2to1 ratio, and stores the structure in an array.

If the CRP4to3 had a higher value, an analogous procedure would be performed, but considering the CRP4to3 instead of the CRP3to2 and the +4 charge state hypothesis; the result would be stored in the same array. After accounting for all spectra, the array is sorted in a nonincreasing order, according to the ratio score. Finally, CPM selects from the resulting array the first x hypotheses and includes them in the final charge state prediction output, where x equals is the product of the number of spectra and a user-specified variable named 2+3or4Correction. By default, 2+3or4Correction equals to 0.075. Clearly, this is analogous to allowing CPM to “relax” an extra 7.5% to account for multiplex spectra of this class.

For the 3+2or4 correction, CPM selects spectra derived from ions with an initial hypothesized charge state of +3, and having precursors of less than 1300 m/z. The procedure follows similarly as above, by selecting spectra that pass a minimum cut-off and storing the CRP2to1:CRP3to2 or CRP4to3:CRP2to1 ratios together with the new charge state hypothesis and scan number in an array. The results of highest expectation are selected and included accordingly. The 3+2or4Correction has a default value of 0.075.

D) Computation. All computational procedures were carried out on an HP DV-5 notebook with a T9400 2.5 GHz microprocessor, 3GB RAM and Windows Vista. We recommend at least 1.5GB RAM and using a microprocessor with two or more cores. This is because CPM was programmed using parallel computing libraries to take advantage of the latest multi-core technology by distributing the jobs.

Results and discussion

The benchmarking was designed to account for lab to lab variability and tailored to reflect real operating conditions. In this regard, the training (12 LC-MS/MS experiments) and testing (4 LC-MS/MS experiments) datasets were acquired in different labs (the Coon laboratory and the Yates laboratory, respectively). Their “true” charge state distribution and number of spectra are presented in Table I. The charge states of peptides in the training dataset were assigned using high-resolution Orbitrap MS1 data in the Coon laboratory while the testing dataset is a subset of spectra identified only with high confidence in the Yates laboratory. The use of high confidence peptide identifications to assign charge states has been previously done elsewhere5;8.

Table I
Charge state distribution of the training and testing datasets

Cross-validation results using the training dataset

The cross-validation (CV) was performed by excluding one of the 12 LC-MSMS datasets, and using the remaining eleven for training. CPM then predicts the precursor charge states in the spectra from the excluded dataset to evaluate the empirical error. This process is repeated for all datasets of the set. Two CV's were performed; one accounting for charges +2 through +7 during the learning phase, and the other excluding +7. No multiplex spectra correction was applied. The results from these approaches, tested with various RP setting, are presented in Tables II and III, respectively. Assignment of charge states for spectra with charge states of +2 through +7 performed slightly better when only considering charges +2 through +6, and that the charge state prediction efficiency quickly drops as RP increases. As can be noted in both CV's, CPM achieved an error rate close to 1% while working with a RP of 1.75. The RP efficiency was measured by dividing how many newly and correctly assigned charge states were included by the number of charge state assignments resulting from the RP. Assigned charge states were taken as correct if they matched the charge state assigned using the high resolution Orbitrap data.

Table II
Cross-validation in training dataset considering charges 2 through 7 during the learning phase
Table III
Cross-validation results using the training dataset and considering charges 2 through 6 during the learning phase

CPM's benchmarks on the testing dataset and selection of default operating parameters

CPM was benchmarked against the testing dataset for various RP values and using two models obtained from the training dataset; one trained with charge states +2 through +7, and the other without +7. These results are presented in Tables III and andIVIV respectively. In contrast to the previous analysis described above, the model accounting for charges +2 through +6 achieved a slightly better overall performance. This happens because there are no +7 spectra within the testing dataset; therefore, there is one less false hypothesis to account for and this makes the relaxation procedure seem more efficient. In light of the results from Tables III and andIV,IV, we empirically determined the default CPM RP to be 1.75 (with a 3+2or4Correction of 0.075, and a 2+3or4Correction of 0.075); however, we recall that the user can change these values and alter the compromise between search engine time and loss. With these settings, CPM achieved an error rate of 2.2% in the testing dataset. The later can be claimed to be conservative as the DTASelect software that was used to filter SEQUEST results was set to allow a 1% false positive rate in the dataset.

Table IV
CPM benchmarks in the testing dataset using the classification model that considers charges +2 through +7

The results from the training and testing dataset suggest that CPM can adequately generalize between labs when operating with an RP of 1.5 or higher. Even though the classification efficiency quickly declines as the RP is increased, in our view it is better to allow a few more false charge state assignments rather than sacrifice the number of identified protein peptides (this is why we set the default RP to 1.75). The gist of CPM's relaxation procedure is to minimize the error rate through global decisions by ordering all charge state hypotheses in the solution array according to scores, so as to choose the most favorable ones within a user-specified relaxation bound.

The multiplex corrections can be interpreted as complementary relaxations that work on orthogonal premises to the RP procedure. An efficiency plot, evaluated similarly to the RP efficiency, is presented in Figure 2 for both multiplex corrections. Indeed, by manually verifying the spectra selected by these filters, we observed characteristics of multiplexed spectra (e.g., 3to2CRPs together with 2to1CRPs, etc.). However, these claims are bound to a specialist's interpretation; as far as we know, there is no reference software to account for such. Both correction efficiencies quickly declined as its user-specified value increased. These corrections were applied after the relaxation procedure; therefore, higher RP's yield even lower multiplex correction efficiency (not shown). The default parameters for both multiplex corrections were set to 0.075 so as to still be effective (~0.1) when operating in conjunction with suggested RP's (1.75 or 2.0).

Figure 2
Multiplex Correction Efficiency for RP = 0

Comparison between CPM and Charger

Charger was executed on the same testing set as CPM and assigned 14,500 charge states to 13,058 spectra; 1,926 spectra did not match the high confidence SEQUEST results, yielding a 14.7% error. Since Charger assigned an extra ~11% charge states, we set CPM's relaxation and multiplex corrections to allow up to an 11% global relaxation for a fair comparison. The RP was set to 1.06, the 3+2or4Correction to 0.0025, and the 2+3or4Correction to 0.0025. CPM produced a 7.2% mismatch for the same dataset, thus, showing a better performance when operating under equivalent conditions for this dataset. In our view, even CPM's 7.2% error was still high and this is why we suggest using a higher relaxation parameter of 1.75 as justified above, as to better handle datasets from different labs.

The CPM software

The Charge Prediction Machine is a pattern recognition software programmed in C#. It can be installed with one single click of a mouse in a Windows (XP or Vista) based PC and only requires the freely available .NET 3.5 or later. In case the user does not have the .NET, CPM can automatically update the computer by interfacing with Microsoft's website. Our software can also run on a Linux or Mac, thanks to the Mono project ( CPM can be executed in the command prompt to provide seamless integration in computational proteomic pipelines; however, it is also user-friendly because it can be executed using a graphic user interface (GUI) as shown in Figure 3. The GUI provides extra functionality such as creation and benchmarking of new classification models. For example, a lab might wish to contribute by publishing a new model if they use different enzymes other than trypsin and Lys-C during sample preparation.

Figure 3

The CPM Windows version carries an automatic update to ensure the use of the latest version. Differently than web-based software, if there was an unpleasing software change, a roll-back option which allows CPM to return to its previous state. Taken together, CPM is an easy-to-use, powerful, and flexible tool for assigning charge states to precursors of ETD MS2 spectra.


CPM, together with the two classification models (charges +2 through +7 and +2 through +6) can be downloaded at the PatternLab for proteomics19 project website or at the Yates Lab website (; the license is free for academic use. Both classification models were generated using a trypsin and Lys-C digestion protocol followed by analysis in an ETD mass spectrometer. To ensure optimal performance, applying CPM to data using different enzymes or an ECD instrument requires that a model be generated accordingly.

Table V
CPM benchmarks in the testing dataset using the classification model that considers charges +2 through +6


The authors acknowledge CAPES, CNPq, a FAPERJ BBP grant, NIH 5R01 MH067880, NIH P41 RR01, and the Genesis molecular biology laboratory for financial support. The authors thank Dr. Joshua Coon and Dr. Danielle Swaney at the Department of Biomolecular Chemistry, University of Wisconsin, Madison, for contributing with ETD data.

Reference List

1. MacCoss MJ, Wu CC, Yates JR., III Anal Chem. 2002;74:5593–99. [PubMed]
2. Perkins DN, Pappin DJ, Creasy DM, Cottrell JS. Electrophoresis. 1999;20:3551–67. [PubMed]
3. Tabb DL, Saraf A, Yates JR., III Anal Chem. 2003;75:6415–21. [PMC free article] [PubMed]
4. Sadygov RG, Eng J, Durr E, Saraf A, McDonald H, MacCoss MJ, Yates JR., III J Proteome Res. 2002;1:211–15. [PubMed]
5. Na S, Paek E, Lee C. Anal Chem. 2008;80:1520–28. [PubMed]
6. Magnin J, Masselot A, Menzel C, Colinge J. J Proteome Res. 2004;3:55–60. [PubMed]
7. Klammer AA, Wu CC, MacCoss MJ, Noble WS. Proc IEEE Comput Syst Bioinform Conf. 2005:175–85. [PubMed]
8. Sadygov RG, Hao Z, Huhmer AF. Anal Chem. 2008;80:376–86. [PubMed]
9. Washburn MP, Wolters D, Yates JR., III Nat Biotechnol. 2001;19:242–47. [PubMed]
10. Gatlin CL, Kleemann GR, Hays LG, Link AJ, Yates JR., III Anal Biochem. 1998;263:93–101. [PubMed]
11. Schroeder MJ, Webb DJ, Shabanowitz J, Horwitz AF, Hunt DF. J Proteome Res. 2005;4:1832–41. [PubMed]
12. Hubler SL, Jue A, Keith J, McAlister GC, Craciun G, Coon JJ. J Am Chem Soc. 2008;130:6388–94. [PMC free article] [PubMed]
13. McAlister GC, Berggren WT, Griep-Raming J, Horning S, Makarov A, Phanstiel D, Stafford G, Swaney DL, Syka JE, Zabrouskov V, Coon JJ. J Proteome Res. 2008;7:3127–36. [PMC free article] [PubMed]
14. McDonald WH, Tabb DL, Sadygov RG, MacCoss MJ, Venable J, Graumann J, Johnson JR, Cociorva D, Yates JR., III Rapid Commun Mass Spectrom. 2004;18:2162–68. [PubMed]
15. Peng J, Elias JE, Thoreen CC, Licklider LJ, Gygi SP. J Proteome Res. 2003;2:43–50. [PubMed]
16. Cociorva D, Tabb L, Yates JR. Curr Protoc Bioinformatics. 2007;Chapter 13(Unit) [PubMed]
17. Tabb DL, McDonald WH, Yates JR., III J Proteome Res. 2002;1:21–26. [PMC free article] [PubMed]
18. Dancik V, Addona TA, Clauser KR, Vath JE, Pevzner PA. J Comput Biol. 1999;6:327–42. [PubMed]
19. Carvalho PC, Fischer JS, Chen EI, Yates JR, III, Barbosa VC. BMC Bioinformatics. 2008;9:316. [PMC free article] [PubMed]
20. Hu J, Qian J, Borisov O, Pan S, Li Y, Liu T, Deng L, Wannemacher K, Kurnellas M, Patterson C, Elkabes S, Li H. Proteomics. 2006;6:4321–34. [PMC free article] [PubMed]