Reduction in Database Search Space by Utilization of Amino Acid

Article pubs.acs.org/ac

Reduction in Database Search Space by Utilization of Amino Acid Composition Information from Electron Transfer Dissociation and Higher-Energy Collisional Dissociation Mass Spectra Thomas A. Hansen, Fedor Kryuchkov, and Frank Kjeldsen* Department of Biochemistry and Molecular Biology, University of Southern Denmark, Campusvej 55, DK-5230 Odense M, Denmark S Supporting Information *

ABSTRACT: With high-mass accuracy and consecutively obtained electron transfer dissociation (ETD) and higherenergy collisional dissociation (HCD) tandem mass spectrometry (MS/MS), reliable (≥97%) and sensitive fragment ions have been extracted for identification of specific amino acid residues in peptide sequences. The analytical benefit of these specific amino acid composition (AAC) ions is to restrict the database search space and provide identification of peptides with higher confidence and reduced false negative rates. The 6706 uniquely identified peptide sequences determined with a conservative Mascot score of >30 were used to characterize the AAC ions. The loss of amino acid side chains (small neutral losses, SNLs) from the charge reduced peptide radical cations was studied using ETD. Complementary AAC information from HCD spectra was provided by immonium ions. From the ETD/ HCD mass spectra, 5162 and 6720 reliable SNLs and immonium ions were successfully extracted, respectively. Automated application of the AAC information during database searching resulted in an average 3.5-fold higher confidence level of peptide identification. In addition, 4% and 28% more peptides were identified above the significance level in a standard and extended search space, respectively.

T

development of probability scoring algorithms, peptide matching is still challenging due to the typically minor fraction (10−50%) of the total MS/MS pool of spectra that are correctly assigned to a matching peptide sequence.5,6 The low success rate of peptide ID can be caused by the fragmentation of contaminant nonpeptide ions, low-quality MS/MS spectra, and the absence of the predicted peptide sequence in the database used for comparison with the fragmented peptide ions. These factors all increase false negative rates.7 Improvement in the reliability of peptide ID can therefore be achieved by improving MS/MS peptide matching. To date, several research groups have extended the specificity of the fragment ions submitted to search engines in order to improve peptide matching. Approaches such as complementary golden pairs,8,9 multiple MS/MS stages,10,11 deisotoping,12 as well as new scoring algorithms13−16 have significantly improved peptide ion scores and yielded more accurate peptide IDs in large-scale proteomics studies. However, improving the ion score of peptide matches is just one way of advancing database searching. In most search engines, the best matching peptide is selected from the total set of potential peptide candidates fulfilling the criteria of the

ypical mass spectrometry-based (MS-based) proteomics studies using current methodologies identify thousands of proteins that can potentially explain or generate biological hypotheses worthy of further investigation. Therefore, growing attention has been paid to the robustness and reliability of MSbased protein identification (ID) strategies.1 Reliable protein ID can prevent incorrect biological conclusions and reduces the time and resources spent on unnecessary biological follow-up studies. A widespread proteomics approach is bottom-up,2 involving the enzymatic digestion and separation of complex protein mixtures by liquid chromatography (LC) in conjunction with online MS. In the mass spectrometer, peptide ions are subjected to tandem mass spectrometry (MS/MS) using a range of different fragmentation techniques. The fragment ions produced from each individual peptide are then used to match predicted fragment masses from all peptide candidates within a predefined interval of mass tolerances in protein sequence databases.3 Most database search engines operate by differentiating falsepositive IDs from true results by assigning an ion score to each MS/MS spectrum. In this approach, the reliability of a correct match between a recorded MS/MS spectrum and the predicted spectrum is usually achieved through probability scoring algorithms. In the search engine Mascot,4 the ion score reflects by this principle the number of matching fragment ions to a hypothetical peptide MS/MS spectrum. Despite ongoing © 2012 American Chemical Society

Received: April 16, 2012 Accepted: June 29, 2012 Published: June 29, 2012 6638

dx.doi.org/10.1021/ac3010007 | Anal. Chem. 2012, 84, 6638−6645

Analytical Chemistry

Article

ETD MS/MS data or to spectra of both types obtained consecutively.

search parameters. This pool of potential peptides is referred to as the search space. The search space size is determined by the number of peptides generated by in silico digestion of the proteins from the species in question, with and without potential post-translational modifications (PTMs), and is affected by the peptide mass tolerance. Hence, the benefit of obtaining peptide mass measurements with high mass-accuracy is to reduce the search space size by restricting the number of peptide candidates against which the MS/MS spectrum has to be matched. A more restricted search space results in shorter search times and greater confidence in peptide ID. For this reason, high-mass accuracy mass spectrometers, such as FTICR MS and orbitrap MS, have been highly appreciated in the proteomics field. Since it is unlikely that PTMs in biological samples are known a priori, the main variable parameter for reducing the database search space is the accuracy of peptide mass measurements. However, other valuable information is contained in each MS/MS spectrum that could potentially lead to further exclusion of invalid peptides from the databases. For instance, a typical MS/MS spectrum of tryptic peptides contains on average 100−200 peaks of which only a fraction (approximately 10−20) are peptide sequence specific (backbone cleavages) fragment ions. Although many of the remaining peaks are due to chemical/electronic noise and unknown chemical decompositions, previous studies have focused on the information content of other unused amino acid specific information found in MS/MS spectra. Immonium ions generated via higher-energy collisional dissociation have been of particular interest.17 Another type of ion in ECD/ETD mass spectra that is amino acid specific is those ions formed from charge reduced ion species, termed small neutral losses (SNL).18−21 In ETD,22 polyprotonated peptide cations react with radical anions in the gas-phase, leading to electron transfer from the radical anions to protonated peptide cations. During the reaction, unstable radical intermediates are formed that can undergo radical initiated backbone bond cleavages producing specific c,z-type fragment ions.22 The alternative and also prominent reaction is the formation of SNLs from the charge reduced species. This reaction has been previously shown to provide specific AAC information for a number of amino acid residues.19−21 Statistical analysis of SNL has been undertaken by the group of Zubarev23 using the sibling technique of electron capture dissociation (ECD).24 Recently, Coon and coworkers21 investigated SNL using ETD of Lys-C peptides. Despite the sizable difference between tryptic and Lys-C peptides, as well as differences in the gas-phase reaction conditions, many similarities in SNL product formation were found in the two studies. Similarly, the analytical value of immonium ions in various de novo algorithms has been investigated, as well as its use in search space restriction.25,26 Currently, search engines do not routinely use AAC information in proteomics studies, and as a consequence, peptide spectra are matched against an unnecessarily large number of peptide candidates, resulting in fewer reliable peptide IDs. The purpose of the present study is to further restrict database search spaces by taking advantage of AAC information extracted from HCD and ETD spectra. Furthermore, this work presents a statistical analysis of the AAC information from ETD and HCD data, as well as demonstrates a program (CompositionAnalyzer) that automatically integrates the extracted AAC information into database searches. The presented strategy can be applied independently to HCD or

■

EXPERIMENTAL SECTION Sample Preparation. Hela 229 cells (American Type Culture Collection (ATCC), Manassas, VA) were cultured in 15 cm tissue culture plates in custom formulated, high glucose DMEM with L-glutamine and pyruvate, and supplemented with 10% dialyzed fetal calf serum. The cells were allowed to grow to 80% confluence before they were harvested by scraping from the plates. The cells were washed once in PBS buffer prior to storage at −80 °C. HeLa cells from one 15 cm plate were lysed in 1 mL 0.5% SDS, 0.1 M dithiothreitol (DTT), and 50 mM NH4HCO3 on ice. To cleave the nuclei, 1.5 μL MgCl2 and 1.5 μL benzonase were added to the lysis buffer. After 15 min, the Hela proteins were separated from the lysis buffer by centrifugation (14 000 rpm at 10 °C for 25 min) in a microcon device (YM-10, Amicon). After centrifugation, 50 μL (0.1 mM) of iodoacetamide (Sigma Aldrich) in 8 M urea was added to carboxamidomethylate the free thiol functionalities, and the solution was left for 20 min at RT. The protein mixture was then washed three times with 200 μL of 8 M urea, followed by centrifugation to retrieve the protein. Finally, the proteins were dissolved in 250 μL of 50 mM NH4HCO3 and subjected to proteolysis using 20 μg trypsin (Promega, Madison, WI) with incubation overnight at 37 °C. The digest was collected by centrifugation (14 000 rpm at 10 °C for 25 min), and the filter was rinsed with 100 μL of 0.1 M NH4HCO3. Formic acid (FA) was added to a final concentration of 2%, and the sample was centrifuged for 10 min at 14 000 rpm to remove any lipids left in the sample. Peptide Fractionation and LC-MS/MS. Tryptically digested Hela peptides were fractionated into 20 fractions using an in-house packed capillary column (0.32 mm × 350 μm) of fused silica (Polymicro Technologies). The capillary was capped with a PEEK in-line microfilter (Upchurch Scientific). The column was packed with 3 μm TSKGel Amide 80 resin (Tosoh Bioscience) HILIC column packing. A capillary flow HPLC system (Agilent 1200) was used to deliver a linear gradient of 100−60% mobile phase B (mobile phase A, 0.1% TFA in H2O; mobile phase B, 90% acetonitrile/10% H2O with 0.1% TFA) in 30 min at a flow rate of 6 μL min−1. Fractions were evaporated and resuspended in 6 μL of 0.1% aqueous formic acid. Sample aliquots (5 μL) were loaded via an EASY-nano LC system (Thermo Fisher, Denmark) directly onto a fused silica capillary column (200 mm × 100 μm) packed with ReproSil-Pur C18 AQ 3 μm reverse phase material (Dr. Maisch, Ammerbuch-Entringen, Germany). The HPLC linear gradient was 0−40% mobile phase B (mobile phase A, 0.1% formic acid in H2O; mobile phase B, 90% acetonitrile/ 10% H2O with 0.1% formic acid) in 120 min at a flow of 250 nL/min. Eluting peptides were electrosprayed (capillary voltage +2100 V) into an LTQ-Orbitrap XL mass spectrometer (Thermo Scientific, San Jose, CA). The LTQ-Orbitrap XL was operated in a data-independent mode, automatically switching between MS (30 000 fwhm at m/z 400, mass range m/z 300−1800) and MS/MS using both higher-energy collisional dissociation (HCD; normalized collision energy 40 arbitrary units; activation time 1 ms; resolution 7500 fwhm at m/z 400) and consecutively electron transfer dissociation (ETD; reaction time 85 ms for 2+ ions, charge state dependent activation time, 7500 fwhm at m/z 400, supplementary activation enabled). After each mass spectrometry survey scan 6639



Article

To find the number of peptides identified at a given false discovery rate (FDR) we calculated the FDR using the following formula:

(MS), three MS/MS events followed using a 2.5 Da isolation window. Data Processing. Peak picking was performed using Proteome Discoverer software (version 1.2, Thermo Fisher Scientific). Individual mgf-files were generated for HCD and ETD spectra, respectively. The signal-to-noise ratio was set to 2 with all other parameters at the default settings. Peptides were identified using Mascot version 2.3.02 and the ipi.Human.v3.43 database with a reversed database concatenated. The following search parameters were used: enzyme, trypsin; instrument, ESIFTICR (HCD) and ETD-TRAP (ETD); mass accuracy precursor, 5 ppm; mass accuracy fragments, 0.02 Da; fixed modifications, carbamidomethyl (C); variable modifications, oxidation (M); max missed-cleavages 1. This standard search space was extended to a larger search space by addition of the following parameters: enzyme, semitrypsin; variable modifications, acetyl (K), acetyl (protein N-term), amidated (C-term), deamidated (NQ), methyl (DE), oxidation (M), phosphorylation (STY). Data Analysis. For the AAC data analysis, we used a set of 6706 consecutive HCD and ETD MS/MS spectra with unique confident peptide sequence assignments (Mascot ion score >30). In cases where only one of the HCD or ETD spectra led to a confident match, this sequence was assigned to both MS/ MS spectra. All immonium ions and SNL ions were identified using a mass accuracy of 5 ppm. The sensitivity and positive prediction value (PPV) of both immonium ions and SNLs were calculated as follows:

FDR =

Retrieving AAC Information from HCD and ETD Spectra. Prior to database searching an in-house software program (CompositionAnalyzer, freely available at http:// composition.sdu.dk/) was applied for extraction of the ≥97% reliable AAC information from their corresponding peak lists (mgf files) of both HCD and ETD spectra (Scheme 1). The Scheme 1. Flowchart of the AAC Approach

positive predictive value =

CompositionAnalyzer software analyses within seconds an entire mgf file and creates a new search file containing only reliably determined (PPV ≥ 97%) AAC information. Briefly, for each HCD and ETD MS/MS spectrum, all specific immonium ions and SNL were determined and evaluated to ensure their reliability. Any redundant information was omitted, and the information was written into the mgf file(s) using the Mascot [Comp] command. This command makes Mascot disregard any peptide not containing the specified amino acid(s). All MS/ MS data were searched as HCD data, with and without the addition of the AAC information. Evaluation of the AAC approach was conducted by comparing unique peptides with a maximum 0.05 expectation value.

∑ true predictions of amino acid x ∑ predictions of amino acid x

sensitivity =

∑ true predictions of amino acid x ∑ peptides containing amino acid x

An intensity ranking was applied to determine the significance of immonium ions at different intensities among all peaks in the mass range m/z 70−160. For each immonium ion, the rank order was then established where only PPVs ≥ 97% were accepted. Since HCD spectra were acquired with the lower mass of m/z 70, it precludes the detection of expected immonium ions of Gly, Ala, and Ser at m/z 33.034, 44.049, and 60.044, respectively. As a formal rule for identifying SNL, we ignored all peaks that could potentially arise as isotope peaks of other fragment ions. Hence, we considered peak X a potential isotope peak if another peak Y existed which satisfied the following criteria (eqs 1 and 2): Ymass = X mass − 1.0028,

Yintensity >

within 5 ppm

■

RESULTS AND DISCUSSION Discrimination Power of Peptide Candidates Using AAC Information. To assess the analytical potential of the AAC approach to proteomics, we first evaluated the discriminative power of the AAC information in silico using a tryptic digestion of all human proteins. Masses of all tryptic peptides, including those with one missed-cleavage and methionine oxidation from the forward and reversed database, were distributed into bins of 10 Da intervals. For each peptide, the number of peptide candidates within a 5 ppm mass tolerance window was counted and the average numbers for all the peptides in each bin were plotted (Figure 1). Even with 5 ppm mass accuracy, the search space of human tryptic peptides was still substantial. For instance, in the mass range of typical tryptic peptides (m/z 700−1200), each peptide has an average of 188 alternative peptide candidates with precursor masses deviating by 5 ppm. The large density of tryptic peptides in that mass range increases the probability of encountering peptides in the database with identical elemental composition or near-

(1)

X intensity × 300 Da X mass

(∑ identified decoy peptides) × 2 ∑ identified peptides

(2)

The sensitivity and PPV were calculated separately for ETD spectra with precursor ions in charge states 2+ and >2+. For the ETD spectra with >2+ charged state precursors, the SNL PPV and sensitivity were determined for losses originating from singly charged charge reduced species and higher charged charge reduced species separately. 6640



Article

density of tryptic peptides in the database decreases, which results in a greater degree of specificity. This analysis indicates that the reliability of peptide ID in large-scale proteomics studies, even with high performance instrumentation, would significantly benefit from additional restriction of the search space, especially in the lower mass region. Simulating the discrimination power obtained by the knowledge of three different amino acids (Ala, Asn, and Trp) is shown in Figure 1 for comparison. These amino acids residues are common, uncommon, and rare in human proteins (frequencies are 7.0% for Ala, 3.6% for Asn, and 1.3% for Trp). Interestingly, even the frequently appearing amino acid residue, Ala, more than halves on average the number of peptide candidates for all peptides below m/z 1200. As expected, the discriminating power is stronger for the less common amino acid residues (Asn and Trp). For instance, the AAC information from Asn resulted in a more than 3-fold reduction in the same mass range. Such discrimination power exceeds that obtained with accurate mass measurements of 1 ppm up to m/z 1200 and 3 ppm up to m/z 2500 for peptides without utilization of AAC information (Supporting Information, Figure S1). Therefore, the AAC approach is particularly useful for tryptic peptides since it restricts the search space most for peptides with fewer amino acid residues. Consequently, the discriminating power of the AAC information diminishes as a function of peptide mass. This is expected since

Figure 1. Average number of candidate human peptides within 5 ppm mass tolerance window as a function of the peptide molecular mass: (×) no AAC restriction applied to the search space; (◇) Ala restriction; (△) Asn restriction; (○) Trp restriction. Solid lines () represent fitted functions of the ratio of peptides containing AA of the three different composition amino acid restrictors.

identical masses. In the worst case scenario, peptides with identical elemental composition obviously cannot be distinguished by their mass alone. At higher peptide masses, the

Figure 2. Example using AAC information from HCD and ETD spectra. One immonium ion and one SNL were confidently characterized for the peptide QCLPSLDLSCK. 6641



Article

Table 1. PPV and Sensitivity (in Percent) of SNLs Determined for Different Charge States of the Precursor Ions (Prec) and Charge States of the Charge Reduced Species (CRS) from Which the Loss is Observeda CRS = 1+, Prec = 2+

a

CRS = 1+, Prec > 2+

CRS > 1+, Prec > 2+

amino acid(s)

mass difference, Da

sequences with AA

sensitivity

PPV

sensitivity

PPV

sensitivity

PPV

D/E S/T N/Q D/E D H H E C C Y

27.995 35.037 45.022 46.005 60.021 82.053 83.061 89.048 90.001 107.028 108.058

5542 5145 4431 5542 3587 1004 1004 4354 1047 1047 1608

40 2 34 8 25 14 0 1 44 14 3

87 90 97 95 99 99 NAN 98 99 99 100

4 3 14 2 8 2 3 2+ precursor charge states. Additional charging of peptide ions can be further promoted by the addition of a supercharging agent to the electrospray solutions.27,28 These observations justify the need for a charge state dependent analysis of the SNL information. Further studies are necessary in order to explain the great diversity in PPV among the same SNLs originating from different charge states of reduced species and precursor ions. Frequency Analysis and Validation of Immonium Ions. The presence of interfering coisolated peptides prior to peptide dissociation significantly reduces the overall reliability of the immonium ions extracted from HCD spectra. Despite extensive prefractionation of the Hela peptide mixture, coisolated peptides in the MS/MS events were abundant (>30% of all peptide IDs). Similarly, Old and co-workers7 found that more than 50% of all MS/MS spectra contained contaminating peptides at a level greater than 20% of the precursor ions in an unfractionated human peptide sample. In the present study, the human peptide sample was prefractionated into 20 fractions and fragmented using a 2.5 Da isolation window. Despite this, only 4 out of 16 unique immonium ions of individual amino acid residues could be determined with

IDs of larger peptides are more specific due to reduced peptide density and, on average, a greater probability that the peptide contains a larger variety of the 20 natural occurring amino acids. Overall, these features make the AAC approach highly complementary to the conventional method of bottom-up proteomics and database search space restriction based on highmass accuracy measurements of peptides. Database Searching of a Single Peptide Using AAC Information. Figure 2 shows an example of AAC information extracted from HCD and ETD product mass spectra for the peptide QCLPSLDLSCK. In the HCD spectrum, the immonium ions corresponding to the amino acids Leu/Ile are observed at m/z 86.0967 as determined with a mass deviation of 3.2 ppm. In the ETD spectrum, SNL ions corresponding to Cys-(carboxymethylated) were identified at m/z [(M + 2H)+• − 89.9987] with a mass deviation of 1.9 ppm. For comparison, the peptide HCD MS/MS peak list was searched against the human protein database with and without addition of the AAC information as search space restrictors. As expected, the Mascot ion score did not change with the addition of the AAC information as it is a function of how well the MS/MS spectrum matches the theoretically suggested peptide MS/MS spectrum. At present, both the Mascot and Sequest search engines can match immonium ions but not SNL ion signals.3 The Mascot expectation value decreased from 0.019 (1 in 526) to 0.000 35 (1 in 2857) as a result of the reduction in the database search space from 294 to 51 peptide candidates after addition of the AAC information. The expectation value is an estimate of the number of matches with equal or better scores than are expected to occur by chance alone. Therefore, for the QCLPSLDLSCK peptide, the ID was >5-fold more confidently assigned using the AAC information. Frequency Analysis and Validation of SNL. The ETD fragmentation pathway giving rise to SNLs results in peaks in the mass range 0−135 Da below the peptide charge reduced species [M + nH](n‑1)+•. In order for the AAC information from SNLs to be useful in proteomics studies, the AAC information must have an acceptable reliability and the frequency of occurrence should be sufficiently high. Since AAC information is applied a priori to database searching, it is crucial to use stringent criteria for the reliability. In this study, we chose a conservative reliability threshold (PPV ≥ 97%), so that the presence of a particular SNL must coincide with the predicted amino acid residue in a minimum 97% of cases. In order to 6642



Article

Table 2. PPV and Sensitivity of Immonium Ionsa amino acid

mass, Da

(unfiltered) sensitivity

(unfiltered) PPV

intensity ranking

(after ranking) sensitivity

(after ranking) PPV

P V T I/L N D Q K E M H F R C Y W

70.065 72.081 74.060 86.096 87.055 88.039 101.071 101.107 102.055 104.053 110.071 120.081 129.113 133.043 136.076 159.092

38 42 17 84 19 11 52 3 36 25 69 74 0 52 80 70

83 97 98 95 88 79 81 100 97 88 70 79 NAN 92 72 68

3 4 7 2 3 0 0 none 10 2 1 1 0 2 0 1

14 23 8 50 5 0 0 3 30 9 17 23 0 16 0 15

99 100 99 99 98 NAN NAN 100 98 99 97 99 NAN 99 NAN 100

a

An intensity ranking was applied to ensure a PPV of 0.97 or higher for immonium ions with at least this rank (bold and italicized) (NAN = not a number).

PPV ≥ 97%. With the aim of ensuring a PPV ≥ 97% for all immonium ions, we introduced an immonium specific filtering criteria based on the signal abundance (rank 1−10) of the individual immonium ions (Table 2). The intensity ranking is constructed such that the PPV is determined as a function of its intensity rank in the mass range m/z 70−160 for each immonium ion. For instance, the total PPV of Asn (N) is only 88% if no ranking is applied. However, the analysis shows that if the immonium ions of Asn are present as the first (rank 1), second (rank 2), or third (rank 3) most intense signal in the specific mass range, its PPV is 100%, 98.1%, and 97.7%, respectively. The average PPV of rank 1−3 is 98%. However, the immonium ion of rank 4 for Asn is 5

0 1 2 3 4 5 >4

892 1646 586 58 1 0 0

547 1198 496 69 2 0 0

197 453 208 24 1 0 0

48 128 53 10 2 0 0

17 36 18 5 0 0 0

0 5 4 0 0 0 0

1 0 1 0 0 0 0

6706 unique peptide sequences.

In total, 87% of all peptides produce at least one informative AAC peak. On average, 1.8 AAC reliable peaks can be successfully extracted from a combination of both HCD and ETD. The presence of AAC containing peaks is almost equivalent in the two fragmentation techniques with 5162 SNL and 6720 immonium ions verified from the full analysis data set. Reduction in Search Space Using Integrated AAC Information. The general frequency of reliable AAC information of HCD and ETD spectra encouraged further application of AAC information to database searching. Using the entire data set and a custom-written software program, SNL and immonium specific fragment ions were automatically translated into corresponding AAC information and added to the search header of each individual peptide HCD MS/MS peak list. The AAC containing mgf files are applicable to any search engine that accepts this type of information for search space restriction, and in the present case we applied Mascot. In all instances, the Mascot search parameters were identical and considered to be reliably identified with an expectation rate equal to or less than 0.05. An immediate benefit of the AAC approach is an observed increase in the peptide identification rate as a result of the reduced search space. Using the AAC 6643



Article

no AAC information. Such a spectrum is typically recorded at a different retention time where the contaminating peptide is no longer present. As a result, the poorer quality spectrum would receive a lower Mascot ion score, and the comparison would be unfavorable. Importantly, addition of false AAC information rarely (82%) of all identified peptides in the standard search space were achieved with higher confidence using the AAC information (average 4.1-fold). The confidence of peptide ID for about 18% (1213 peptides) was not significantly improved with the AAC approach. For this group of peptides, no AAC informative peaks were successfully extracted. Finally, the IDs of only 53 out of 6846 peptides (