ISSN 10619348, Journal of Analytical Chemistry, 2010, Vol. 65, No. 14, pp. 1462–1468. © Pleiades Publishing, Ltd., 2010. Original Russian Text © A.A. Goloborod’ko, C. Mayerhofer, A.R. Zubarev, I.A. Tarasova, A.V. Gorshkov, R.A. Zubarev, M.V. Gorshkov, 2010, published in Zhurnal Massspek trometria, 7(1), pp. 46–52.

ARTICLES

Alternative Methods for Verifying the Results of the Mass Spectrometric Identification of Peptides in Shotgun Proteomics A. A. Goloborod’koa, C. Mayerhoferb, A. R. Zubarevb, I. A. Tarasovaa, A. V. Gorshkovc, R. A. Zubarevb, and M. V. Gorshkova aInstitute

of Energetic Problems of Chemical Physics, Russian Academy of Sciences, Leninskii pr. 38, bld. 2, Moscow, 119334 Russia email: [email protected] b Biological and Medical Center, Uppsala University, Uppsala, S75123 Sweden email: [email protected] c Semenov Institute of Chemical Physics, ul. Kosygina 4, Moscow, 119991 Russia Received October 7, 2009; in final form, December 2, 2009

Abstract—Database search is the most popular approach used for the identification of peptides in contem porary shotgun proteomics; it utilizes only mass spectrometric data. In this work, we introduce three criteria for the verification of peptide identification; these are based on the analysis of data orthogonal to tandem mass spectra. The first one utilizes chromatographic retention times of peptides. The development of such approaches has been hindered by the relatively low accuracy of retention time prediction algorithms. In this work, we suggest the use of two independent models of the liquid chromatography of peptides, which increase the reliability of the results. The second criterion utilizes the mean number of missed tryptic cleavages per peptide. The third one results from the analysis of the difference between theoretical and experimentally mea sured peptide masses. The proposed criteria were applied to the tandem mass spectra of tryptic peptides from rat kidney tissue, which were processed by the Mascot search engine. All the criteria showed that Mascot sig nificantly overestimated the reliability of an identification. This conclusion was supported by the PeptidePro phet algorithm. Keywords: proteomics, Mascot, mass spectrometry, liquid chromatography, retention time prediction. DOI: 10.1134/S1061934810140042

INTRODUCTION Shotgun proteomics is one of the most widely used methods for the identification of proteins [1, 2]. In this method, a tested protein mixture is cleaved into pep tides using a protease with a high specificity, for exam ple, trypsin. The obtained mixture of proteolytic pep tides is analyzed by tandem mass spectrometry com bined with liquid chromatography (HPLC– MS/MS). Tandem mass spectra are processed by search engines, such as Mascot [3], SEQUEST [4], and others [5–7]. The aim of these programs is to compare the experimental fragmentation spectrum with the theoretical spectra calculated for proteolytic peptides contained in the database. A translated genome of a source organism of the studied mixture is usually used as the database. The degree of matching is characterized by a confidence index of a peptide. The confidence of the peptide–spectrum match, for example, in the Mascot model is characterized by the difference between the confidence index and the individual threshold value of the index. The latter, in its turn, depends on the amount of peptides with a suitable mass in the database. Thus, the size of the

database determines if one or another index value is enough for the peptide–spectrum match to be consid ered confident. However, if the true sequence for the analyzed spectrum is absent in the database, it is prob able that the search engine will suggest a sequence with a relatively high index. This can be caused by two rea sons: the homology of sequences and uncertainties in the interpretation of MS/MS spectra. The former is due to the fact that the sequences of two different pro teins can overlap up to several amino acids (according to some estimates, up to 20% of proteins contain at least one common tryptic peptide [8]). This can be explained by both evolution and the limitations posed on the amino acid sequences of proteins with a certain tertiary structure. Certainly, homologous peptides do not always have the same or similar masses, which leads to an error in the MS/MS identification. How ever, the presence and wide diversity of posttranslation modifications (according to the Unimod databse [9], more than 600 modifications are known) makes this limitation less severe. Uncertainty in the interpreta tion of MS/MS spectra is a nonbiological problem; the point is that two nonhomologous peptides can

1462

ALTERNATIVE METHODS FOR VERIFYING THE RESULTS

often yield similar fragmentation spectra. In particu lar, each peptide forms such a pair with its inverted copy, because the series of b fragments of the direct peptide is identical to the series of (yH2O) fragments of the reversed peptide. Moreover, the most important reasons for the uncertainity of MS/MS spectra is the replacement of the amino acid residues of isoleucine by leucine, asparagin by two glycines, and glutamine by alanine and glycine, amino acid rearrangements, and other isomeric effects. Nonisomeric effects also matter such as the replacement of amino acid residues with similar masses, for example, glutamine by lysine. Both the homology of peptides and uncertainity in MS/MS spectra often result in several peptides with close confidence indices exceeding the threshold value being attributed to one spectrum [10]. The probability of this misidentification, among other issues, also depends on the quality of the fragmentation spectra: lowresolution spectra, obtained, for example, using ion traps, cause more uncertainties than the data with a high resolution and accurate mass determination. With an increase in the confidence index, the contri bution of the mentioned uncertainties in the peptide identification decreases; however, there is no unam biguous answer to the question regarding the index value at which the identification can be considered absolutely confident. Usually, the threshold confidence index is selected so that the probability of a wrong identification (false discovery rate, FDR) is less than 5%. The most popular approach used for the determination of FDR is the method of a decoy database search [11]. In spite of its wide use, this method has a number of drawbacks, some of which were discussed earlier [8]. However, the most serious argument against this approach is that any model must be examined by a method compli mentary to those used in its development. In the con text of shotgun proteomics, this means that it is advis able that the results of proteomic search engines are checked by a method based on another principle than the database search or, even better, not on mass spec trometric methods at all. According to this principle, the method for determining FDR by the false database search, which is currently widespread, is only a test of the search results for consistency. In the analysis of peptide identification in shotgun proteomics, we have earlier proposed the use of theo retical retention times predicted by two algorithms developed independently [12, 13]. The application of this approach for examining a limited amount of data has shown that, for the majority of identified peptides, significant deviations of the experimental retention times from the theoretical retention times were observed. On one hand, such deviations can be explained by the presence of serious problems in the chromatographic algorithms. On the other hand, this can be a result of uncertainity in MS/MS spectra. In the present work, we tried to assess which of the mentioned effects is really definitive by processing a JOURNAL OF ANALYTICAL CHEMISTRY

Vol. 65

1463

statistically significant array of peptide identifications using more complex methods for estimating the confi dence of identifications. EXPERIMENTAL Retention time prediction algorithms. In this work, we used two relatively new models for predicting retention times: BioLCCC (Liquid Chromatography of Biomolecules under Critical Conditions) [14–16] and SSRCalc [17–20]. BioLCCC is a physical model that considers peptides as macromolecules and analyt ically describes their separation in thermodynamic terms. The latter model, SSRCalc, descends from early additive retention models and takes into account many additional factors of peptide separation using empirically determined coefficients. Both models are available on the Internet (http://proteomics.fizteh. ru/biolccc and http://hs2.proteome.ca/SSRCalc/ SSRCalc.html, respectively) and are applicable over a rather wide range of experimental separation condi tions most widely used in proteomics [16, 18, 19]. The BioLCCC model contains a comparatively small number of phenomenological parameters, namely, 20 experimentally found values of the energy of the interaction of amino acids with a reversed phase, 4 values of energy for terminal groups, and 2 values of energy characterizing the interaction of solvent mole cules with a surface. Because of this, BioLCCC can easily be adapted to any types of stationary phase and solvents. Rather good correlation was demonstrated in practice between the theoretical (RTBioLCCC) and experimental (RTexp) retention times (determination coefficient R2 > 0.9), as well as the capability of recog nizing small differences in a peptide sequence [14]. The SSRCalc model was developed at the Mani toba Center for Proteomics (Canada). The main dif ference of SSRCalc from the additive approach con sists in the introduction of additional nonlinear coeffi cients for terminal amino acids, physicochemical properties of a peptide, and others. As a result, the SSRCalc model calculates relative hydrophobicity RHSSRCalc of a peptide, which later can be converted into the retention time using a linear equation. The coefficients of this equation depend only on the slope of the used gradient and the dead volumes of the chro matographic system. This model demonstrates a good correlation between the predicted values of RHSSRCalc and the experimental retention times RTexp: R2 reaches ~0.98 for the training set of peptides and 0.95 for real peptide mixtures obtained from the human cancer cell lysates [17, 18]. Sample preparation. A peptide mixture was obtained from rat kidney tissues. The tissue (25 μg) was twice washed in a phosphate buffer solution with pH 7.4 (PBS1, Sigma–Aldrich, St. Louis, United States) and stored in an ultrasonic bath for 45 min in a 10 mM Tris buffer solution containing 6 M urea, 2 M thiourea, and 1% of octylβglucopyranoside at No. 14

2010

1464

GOLOBOROD’KO et al.

pH 8.0. After that, the mixture was centrifuged for 10 min at 13000 rpm. Then, the buffer solution in the supernatant was replaced by 50 mM NH4HCO3 using PD SpinTrap G25 columns (GE Healthcare, Swe den) according to the producer’s instruction. Next, dithiotreitol was added to the mixture up to a concen tration of 15 mM, and the mixture was incubated for 10 min at 95°C. The resulting mixture was stored in the presence of 30 mM iodacetamide for 30 min at room temperature in a dark place. Then, 1 μg of trypsin for sequence analysis (11418025001, Roche, Germany) was added to 50 μg of the sample, and the obtained mixture was stored overnight at 37°C. After that, the sample was filtered through Microcon cen trifugal filters (10 kDa, Millipore, United States), dried on a SpeedVac system, and purified using ZipTip pipette tips with a reversedphase adsorbent. HPLC/MS. Chromatographic separation was per formed on an Agilent 1100 nanoflow system (Agilent, United States). The analytical column was a glass cap illary 15 cm in length with inner and outer diameters of 75 and 375 μm, respectively (Proxeon Biosystems, Denmark) with a ReprosilPur C18AQ 3μm silica adsorbent. The column temperature was not con trolled. The liquid phase consisted of component A (0.5% of acetic acid and 99.5% of water) and compo nent B (0.5% of acetic acid, 10% of water, and 89.5% of acetonitrile). The sample (5 μL) was automatically applied to the column. For the first 10 min, the col umn was washed in an isocratic mode with the liquid phase containing 2% of component B at the flow rate of the binary solvent mixture of 500 nL/min. Then, the column was washed for 90 min with the liquid phase with a gradient of 4 to 50% of component B at a flow rate of 200 nL/min. An LTQ FT hybrid 7 T mass spec trometer (ThermoFisher Scientific, Germany) with a nanospray ion source (Proxeon Biosystems, Den mark) was set at the outlet of the column. The mass spectrometric analysis was performed in an automatic mode, when a common highresolution spectrum (resolution power to 100000) and the fragmentation spectra of five most intensive detected peptides (reso lution power to 25000) were measured in turn. Data processing. The peptide sequence was deter mined by tandem mass spectra using Mascot search program, version 2.1 (Matrix Science, Great Britain). The search was performed using International Protein Index IPI_RAT database dated July 9, 2009, com bined with its inverted copy. The following search parameters were set: up to three missed sites of pro teolytic cleavage per peptide; fixed carboxyamidome thylation of cysteine and probable oxidation of methionine; the accuracy of the determination of the masses of protonated ions 10 ppm, and for fragmented ions, 1.5 Da. The instrument type was selected as “ESI–FTICR”. The obtained file was treated by the PeptideProphet software [21]. The retention times were predicted using the algorithms BioLCCC and SSRCalc available on line. Because none of the pre

sented algorithms can consider the oxidation of methionine, the corresponding peptides were excluded from further analysis. In the case of repetitive identification of peptide, we selected the peptide– spectrum match with the maximum confidence index. RESULTS AND DISCUSSION Relative confidence index of identification. For each peptide–spectrum match, Mascot calculates confi dence index Mion, which reflects the degree of corre spondence between the theoretical and experimental MS/MS spectra. However, the confidence of the iden tification also depends on the number of candidates in the database, which is recalculated into the individual threshold of the confidence index Tidentity. Using these two values, the mathematical expectation Evalue can be estimated for the number of the peptide–spectrum matches with the confidence index larger or equal to Mion, providing that the candidate sequences from the database are random: Evalue = 0.05 × 10

– M rel /10

,

(1)

where Mrel = Mion – Tidentity. (2) The value Mrel will further be used for the assess ment of the confidence of identification. With the use of the described procedure, we obtained 2770 peptide–spectrum matches, 1463 of which had Mrel > 0. Then, the entire list of peptide– spectrum matches was numbered in the order of increasing values of Mrel and was divided in this order into 17 overlapping sets of 300 pairs in each. Thus, the first set contained peptides numbered 1 to 300, the second set, 151 to 450, etc. The next study was performed to find independent (orthogonal) methods for checking FDR values obtained using conventional approaches, such as Mas cot, PeptideProphet, and the search against a false database. Chromatographic criterion. The dependences between the predicted and experimental retention times of peptides are shown in Figs. 1a and 1b. To plot the dependences, only peptides with a high confidence level (Mrel > 50) were used. Both models demonstrated good results; SSRCalc being more accurate (R2 = 0.94) than BioLCCC (R2 = 0.86). The sequences of all the found peptides were used to compare the models (Fig. 1c). The total correlation of the predicted reten tion times was R2 = 0.86; however, it decreased in the region of low retention times. This can be explained by the features of the models themselves. On one hand, Fig. 1b indicates that the model SSRCalc underesti mates the predicted values for peptides with low reten tion times. On the other hand, in the BioLCCC model, peptide is considered as a randomly folded flexible chain, which is better suited for long chains.

JOURNAL OF ANALYTICAL CHEMISTRY

Vol. 65

No. 14

2010

ALTERNATIVE METHODS FOR VERIFYING THE RESULTS

RHSSRCalc

RTBioLCCC, min

60 40 20

(b)

50

y = 1.07x – 11.33

RTBioLCCC, min

(a)

80

40 30 20 y = 0.66x – 7.25

10

30 40 50 60 70 80 90 RTexp, min

30 40 50 60 70 80 90 RTexp, min

1465

(c)

140 120 100 80 60 40 20 0

y = 1.59x + 1.23

0

20 40 60 RHSSRCalc

Fig. 1. Correlation between the theoretical and experimental (RTexp) retention times: (a) values predicted by the BioLCCC model (RTBioLCCC) for peptides with Mrel > 50; (b) values predicted by the SSRCalc model (RHSSRCalc) for peptides with Mrel > 50; and (c) correlations of the predicted values of both models.

ΔBioLCCC, min

50 0 –50

–100 –100

0 –50 ΔSSRCalc, min

50

Fig. 2. Correlation of the deviations of the theoretical retention times, predicted by the algorithms BioLCCC and SSRCalc, from the experimental retention times for all identified peptides.

The distribution of the difference between the theoretical * (ΔBioLCCC and ΔSSRCalc, retention times and R T exp respectively) for all found peptides is demonstrated in Fig. 2. A dependence is observed within this distribu tion, described by the equation ΔBioLCCC = ΔSSRCalc. Although this correlation can be quantitatively char acterized by R2, this coefficient was too sensitive to the position of some points at low values. Therefore, we introduced a new method for measuring the correla tion. It is based on the fact that the distribution of the correlated values is rather asymmetric. For example, the distribution presented in Fig. 2 was stretched along with the ΔBioLCCC = ΔSSRCalc axis and had a signif icantly lower dispersion along with the perpendicular axis ΔBioLCCC = –ΔSSRCalc. This stretching can be quan titatively characterized using standard deviations along with the corresponding axes:

Note that the BioLCCC model yields retention times RTBioLccc close to the experimental values RTexp to within a fixed shift (generally determined by the dead volumes of the chromatographic system), whereas the SSRCalc model counts the relative hydro phobicity values RHSSRCalc, which are then converted into the time scale. To operate with the time values in a unified scale, we converted all RTexp and RHSSRCalc into scale RTBioLCCC and marked the obtained values as R T *exp and R T *SSRCalc , respectively. This conversion was performed using the following equations: R T *exp = AexpRTexp + Bexp, (3) * * R T SSRCalc = ASSRCalcR H SSRCalc + BSSRCalc, (4) which describe the linear dependences of RTexp on RTBioLCCC for peptides with Mrel > 50 and RHSSRCalc on RTBioLCCC for all peptides, respectively. μC =

1  N

N

∑

i=1

i ( ( Δ BioLCCC

+

i Δ SSRCalc )

1 – ( Δ BioLCCC + Δ SSRCalc ) ) –  N

JOURNAL OF ANALYTICAL CHEMISTRY

N

2

Vol. 65

∑

2 (5) i i ( ( Δ BioLCCC – Δ SSRCalc ) – ( Δ BioLCCC – Δ SSRCalc ) ) ,

i=1

No. 14

2010

1466

GOLOBOROD’KO et al. (a) 1.8

30

1.6

20

1.4

10

1.2 0

50 Mrel (b)

100

0

50 Mrel

100

6

〈nRK〉

μC

40

σ(Δm)

5 4 3 2

Fig. 3. Dependence of (a) the quantitative characteristics of the criteria μC and 具nRK典 of the reliability of peptide identifications, orthogonal to mass spectrometry, and (b) the mass spectrometric criteria σ(Δm) on the relative confidence index Mrel. In part (a), the values of the coeffi cient μC are indicated by circles, while 具nRK典, by hatched diamonds.

where ΔBioLCCC = RTBioLCCC – R T *exp ,

(6)

ΔSSRCalc = R T *SSRCalc – R T *exp . (7) The coefficient μC enables the quantitative charac teristic of the correlation of two values connected by a known ratio (in our case, ΔBioLCCC = ΔSSRCalc). In each peptide group, the coefficient μC deter mined as described above gradually decreases with increasing index Mrel averaged over the group (Fig. 3a). However, it is unexpected that the coefficient μC does not stop decreasing at Mrel = 0. It reaches its minimum value only at Mrel ~ 20 and then remains almost con stant for all peptide–spectrum matches with Mrel > 20. It is unlikely that such a curve shape can be caused by either simultaneous errors of both models or some chromatographic effects, because these phenomena should not depend on the reliability of identification. The only explanation remains that the correlation of the errors in the predicted values of both models is caused by the presence of false peptide–spectrum matches in the group. If the error was introduced in the determination of peptide sequence, both models will predict relatively close retention times for this sequence, which, however, will be different from the experimental value. Thus, the introduced chromato

graphic criterion raises a question on the validity of the estimate for the threshold confidence index by the Mascot search engine. Proteolytic criterion. The peptide sequences in the false peptide–spectrum matches can have an average amino acid composition significantly different from the expected one. Lysine (K) and arginine (R) in tryp tic peptides can be mentioned as examples. Trypsin is widely used in proteomics, because it cleaves proteins specifically and efficiently at the Cterminal peptide bond of these amino acid residues. Therefore, in an ideal case, the identified peptides should not contain more than one of these residues. The average numbers of lysine and arginine groups per peptide 具nRK典 for the groups of the peptide–spectrum matches, determined earlier, are shown in Fig. 3a. The value 具nRK典 for the group of the peptide–spectrum matches with the con fidence index equal to its threshold value is ~1.6. According to Eq. (2), this group should have FDR ≈ 5%. However, even if we assume that all peptide sequences in the false peptide–spectrum matches contain three lysine or arginine residues and the sequences in the true pairs contain, on the average, 1.1 such residues (the value 具nRK典 for the peptide–spec trum matches with high Mrel), the obtained maximum estimate for the group of the peptide–spectrum matches with FDR = 5% is ~1.2. The obtained devia tion can be explained by the fact that the Mascot pro gram contains an incorrect algorithm for evaluating the threshold confidence index. This conclusion fol lows from the analysis of the dependence of 具nRK典 on Mrel (Fig. 3a). As seen from the figure, the shape of this dependence is similar to those obtained for the coeffi cient μC calculated from the chromatographic data. Both dependences, μC and 具nRK典 on Mrel, converge at the values corresponding to the true identifications found at the values of the relative confidence index Mrel that is considerably higher than the threshold index. Mass spectrometric criterion. The threshold of the confidence index of Mascot can be evaluated by con sidering the masses of identified peptides. The depen dence of the standard deviation of the difference σ(Δm) between the theoretical and experimental pep tide mass on the mean confidence index Mrel for each group of the peptide–spectrum matches is shown in Fig. 3b. The shape of the obtained curve matches the shapes of the curves in Fig. 3a. The upper estimate of σ(Δm) for the group with 5% of the false peptide–spec trum matches (the accuracy of the mass determination for the true and false pairs is 2 and 10 ppm, respec tively) yields a value of 3 ppm, which is considerably lower than ~5 ppm for the group with the mean Mrel equal to zero. All these arguments prove again that the determination of the threshold value of the Mascot index is incorrect. Peptide prophet. This algorithm is based on the assumption that the true and false peptide–spectrum matches form two subsets, which can be resolved by

JOURNAL OF ANALYTICAL CHEMISTRY

Vol. 65

No. 14

2010

Probability of true identification using PeptideProphet, %

ALTERNATIVE METHODS FOR VERIFYING THE RESULTS

1467

The second considers the mean number of missed pro teolysis sites. Finally, the third criterion is based on the analysis of the difference between the theoretical and experimental masses of peptides. All three criteria demonstrated a similar dependence on the Mascot confidence index. The estimates found for two of them raise the assumption that the threshold value of the Mascot confidence index is determined incorrectly. Further examination by the PeptideProphet algorithm has proved this hypothesis.

100 80 60 40 20 0 0

50 Mrel

100

Fig. 4. Results of analysis of all peptide identifications using Mascot and PeptideProphet algorithms.

constructing a distribution of a certain quantitative characteristic. A function depending on the Mascot index, the number of detected fragments, the differ ence between the theoretical and experimental masses, etc., can serve as such a characteristic. The value of this function, assigned to each pair of pep tide–spectrum, is called a discrimination index. The goal of PeptideProphet is to find the parameters of a function at which the peaks of the pairs in a bimodal distribution by the discrimination index will be maxi mally resolved. As a result, the PeptideProphet algo rithm assigns to each peptide–spectrum match the probability that it pertains to a set of true identifica tions. In other words, one can determine the FDR of the groups of peptide–spectrum pairs using this algo rithm. The results of the application of this algorithm to the studied peptide identifications are presented in Fig. 4. As seen from the figure, the peptide–spectrum matches with the confidence index equal to the threshold index value (Mrel = 0) received poor esti mates from the PeptideProphet algorithm: the proba bility of their confidence was ~20%. Moreover, only the peptide–spectrum matches with Mrel > 20 got the estimate FDR < 5%. Both these results are in contra diction with the threshold value of the confidence index, evaluated by the Mascot program, and in com plete agreement with the proposed criteria and with the estimate of the true threshold index value (corre sponding to FDR < 5%), made on the basis of these criteria. CONCLUSIONS Three alternative orthogonal criteria were pro posed in the work for estimating the reliability of pep tide identification. The first is based on the correlation of the errors of the theoretical retention times of pep tides, calculated using two independent algorithms. JOURNAL OF ANALYTICAL CHEMISTRY

Vol. 65

ACKNOWLEDGEMENTS The work was performed under the Program of Basic Research of the Presidium of the Russian Acad emy of Sciences “Development and Sophistication of the Methods of Chemical Analysis and the Studies of the Structure of Substances and Materials” with the financial support from the American Civil Research and Development Foundation, project no. RUBI2909MO07. The work was also sup ported by the Russian Foundation for Basic Research, projects nos. 080401339, 080491121, and 0904 00633; and by the INTAS, project Genomics no. 05 100000047759. REFERENCES 1. Link, A.J., Eng, J., Schieltz, D.M., Carmack, E., Mize, G.J., Morris, D.R., Garvik, B.M., and Yates, J.R., III, Nat. Biotech., 1999, vol. 17, no. 7, p. 676. 2. Silva, J.C., Denny, R., Dorschel, C., Gorenstein, M.V., Li, G., Richardson, K., Wall, D., and Geromanos, S.J., Mol. Cell. Proteomics, 2006, vol. 5, no. 4, p. 589. 3. Perkins, D.N., Pappin, D.J.C., Creasy, D.M., and Cot trell, J.S., Electrophoresis, 1999, vol. 20, no. 18, p. 3551. 4. Eng, J.K., McCormack, A.L., and Yates, J.R., III, J. Am. Soc. Mass Spectrometry, 1994, vol. 5, no. 11, p. 976. 5. Geer, L.Y., Markey, S.P., Kowalak, J.A., Wagner, L., Xu, M., Maynard, D.M., Shi, W., and Bryant, S.H., J. Proteome Res., 2004, vol. 3, no. 5, p. 958. 6. Fenyö, D. and Beavis, R.C., Anal. Chem., 2003, vol. 75, no. 4, p. 768. 7. Colinge, J., Masselot, A., Cusin, I., Mahe, E., Nikne jad, A., ArgoudPuy, G., Reffas, S., Bederr, N., Gle izes, A., Rey, P.A., and Bougueleret, L., Proteomics, 2004, vol. 4, no. 7, p. 1977. 8. Zubarev, R.A., Zubarev, A.R., and Savitski, M.M., J. Am. Soc. Mass Spectrometry, 2008, vol. 19, no. 6, p. 753. 9. Creasy, D.M. and Cottrell, J.S., Proteomics, 2004, vol. 4, no. 6, p. 1534. 10. Frank, A.M., Savitski, M.M., Nielsen, M.L., Zubarev, R.A., and Pevzner, P.A., J. Proteome Res., 2007, vol. 6, no. 1, p. 114. 11. Moore, R.E., Young, M.K., and Lee, T.D., J. Am. Soc. Mass Spectrometry, 2002, vol. 13, no. 4, p. 378. No. 14

2010

1468

GOLOBOROD’KO et al.

12. Tarasova, I.A., Zubarev, R.A., Goloborod’ko, A.A., Gorshkov, A.V., and Gorshkov, M.V., Massspektrome tria, 2008, vol. 5, no. 1, p. 7. 13. Tarasova, I.A., Goloborodko, A.A., Zubarev, R.A., and Gorshkov, M.V., Proc. 56th ASMS Conference on Mass Spectrometry and Allied Topic, Denver, 2008. 14. Gorshkov, A.V., Evreinov, V.V., Tarasova, I.A., and Gorshkov, M.V., Polymer Sci. B, 2007, vol. 49, no. 3, p. 93. 15. Tarasova, I.A., Gorshkov, A.V., Evreinov, V.V., Adams, K., Zubarev, R.A., and Gorshkov, M.V., Poly mer Sci. A, 2008, vol. 50, no. 3, p. 309. 16. Gorshkov, A.V., Tarasova, I.A., Evreinov, V.V., Savitski, M.M., Nielsen, M.L., Zubarev, R.A., and Gorshkov, M.V., Anal. Chem., 2006, vol. 78, no. 22, p. 7770.

17. Dwivedi, R.C., Spicer, V.L., Harder, M., Anto novici, M., Ens, W., Standing, K.G., Wilkins, J.A., and Krokhin, O.V., Anal. Chem., 2008, vol. 80, no. 18, p. 7036. 18. Krokhin, O.V., Craig, R., Spicer, V., Ens, W., Standing, K.G., Beavis, R.C., and Wilkins, J.A., Mol. Cell. Proteomics, 2004, vol. 3, no. 9, p. 908. 19. Krokhin, O.V., Anal. Chem., 2006, vol. 78, no. 22, p. 7785. 20. Krokhin, O.V., Ying, S., Cortens, J.P., Ghosh, D., Spicer, V., Ens, W., Standing, K.G., Beavis, R.C., and Wilkins, J.A., Anal. Chem., 2006, vol. 78, no. 17, p. 6265. 21. Nesvizhskii, A.I., Keller, A.A., Kolker, E., and Aeber sold, R., Anal. Chem., 2003, vol. 75, no. 17, p. 4646.

JOURNAL OF ANALYTICAL CHEMISTRY

Vol. 65

No. 14

2010