DtaRefinery, a Software Tool for Elimination of Systematic Errors from Parent Ion Mass Measurements in Tandem Mass Spectra Data Sets

Abstract

Hybrid two-stage mass spectrometers capable of both highly accurate mass measurement and high throughput MS/MS fragmentation have become widely available in recent years, allowing for significantly better discrimination between true and false MS/MS peptide identifications by the application of a relatively narrow window for maximum allowable deviations of measured parent ion masses. To fully gain the advantage of highly accurate parent ion mass measurements, it is important to limit systematic mass measurement errors. Based on our previous studies of systematic biases in mass measurement errors, here, we have designed an algorithm and software tool that eliminates the systematic errors from the peptide ion masses in MS/MS data. We demonstrate that the elimination of the systematic mass measurement errors allows for the use of tighter criteria on the deviation of measured mass from theoretical monoisotopic peptide mass, resulting in a reduction of both false discovery and false negative rates of peptide identification. A software implementation of this algorithm called DtaRefinery reads a set of fragmentation spectra, searches for MS/MS peptide identifications using a FASTA file containing expected protein sequences, fits a regression model that can estimate systematic errors, and then corrects the parent ion mass entries by removing the estimated systematic error components. The output is a new file with fragmentation spectra with updated parent ion masses. The software is freely available.

A key component in modern proteomics research is peptide identification through LC coupled to tandem MS where a selected parent or precursor ion from an MS scan undergoes fragmentation by collisionally activated/induced dissociation or any other methods (1). Identification of the putative peptides corresponding to the parent ions selected for fragmentation is performed by matching the observed to the theoretical MS/MS fragmentation patterns. The first step in the data analysis process is to create a set of input files representing the fragmentation spectra. For example, for the data sets from LTQ¹ FT and LTQ Orbitrap instruments, software tools such as extract_msn (part of BioWorks software package, Thermo Electron, San Jose, CA) or DeconMSn (2) are often used for this step, creating files in “.dta” or other formats for the fragmentation spectra. These files contain the mass and charge of the parent ion and observed fragmentation pattern in the form of a list of m/z and intensity pairs. Once created, database search tools such as SEQUEST (3), X!Tandem (4), OMSSA (5), InsPect (6), MASCOT (7), Spectrum Mill (8), RAId_DbS (9), and others are used to analyze the .dta files to associate each MS/MS fragmentation pattern with a corresponding putative peptide sequence. Therefore, MS/MS fragmentation pattern information plays a primary role and ultimately can be used as essentially the only type of information for peptide identification in LC-MS/MS experiments (4, 10). However, in this case, the lack of constraint on parent ion mass measurement error (MME) results in a high rate of incorrect peptide identifications. Conversely, improved mass accuracy helps to achieve a better discrimination between true and false peptide identifications (8, 11).

To fully utilize the high mass measurement accuracy of modern instruments, it is advantageous to eliminate systematic mass measurement errors. Eliminating the systematic MME component results in a more coherent distribution of anecdotal MME and helps to reduce the maximum allowable deviation (of the measured mass from the theoretical peptide mass) for true peptide identifications (12). Multiple sources of variation can cause systematic errors in mass measurements; for example, power supply voltage drift over time, space charge effects, differing ion compositions within the cell, ion intensity variation, and outdated calibration coefficients (for a review, see Ref. 13).

The use of internal calibrants or standards co-injected with the sample into the mass spectrometer (14–18) can help reduce such systematic errors to a certain extent but may have some practical limitations. Internal calibrants well capture scan to scan variations and correct for time and/or total ion current (TIC)-dependent systematic errors, which are associated with the entire MS scan. Technically, it is also possible to correct for intensity-related dependence of MME, which is quite prominent on a certain type of instrument (13). However, it will require an increase in the number of calibrants to cover the entire dynamic range of the mass spectrometer. In addition, in certain cases, the calibration function (MME dependence on m/z parameter) shows evidence of non-linear behavior and may not be corrected by one or even a few calibrants (for example, see Fig. 7 in Ref. 13 and Fig. 5 in Ref. 19). Thus, there still potentially can be residual systematic biases even if internal calibration is applied.

Fig. 7.

Mass error distribution histograms for all peptide to spectrum assignments with 2+ charge produced by SEQUEST searches for fully tryptic peptides within 136 LC-MS/MS data sets. No XCorr or ΔCn filtering criteria have been applied. The bin width is 0.5 ppm. The DtaRefinery tool has noticeably reduced the width of the MME distribution histogram and thus maximum allowable deviation of the parent ion mass for true peptide identifications from about 10 (left, blue) to about 3 (right, green) ppm.

Fig. 5.

Scatter plots showing parent ion MME before (blue) and after (green) additive regression refinement for different parameters: scan number, m/z, log₁₀ of ion intensity, and total ion current of trapped ions.

A number of alternative approaches based on knowledge about the sample content have been introduced recently (13, 20–25). Instead of spiked or co-injected calibrants, they make use of putative peptide identifications as internal calibrants. Initially, such recalibration approaches have been limited mostly to peptide identifications based on either high accuracy measurements of peptide masses alone (21) or in combination with LC retention times (13, 20). Recently, we and others have proposed that partial sample knowledge can also be utilized for recalibrating parent ion masses in MS/MS data sets obtained on hybrid instrumentation (12, 13, 23–26). In one implementation, described as “postexperiment monoisotopic mass filtering and refinement” (26), the parent ion masses in the .dta files were replaced with the mass of the ion averaged over all scans in which it was observed followed by a simple recalibration. That recalibration assumes one constant value of the systematic error for the entire data set, which can be estimated by zero centering the parent ion MME distribution. There is another implementation that resembles this concept (23, 25) that is based on using putative peptide identifications as calibrants from neighboring MS/MS scans, that is scans either immediately before or after the MS scan chosen for recalibration. We reviewed this approach earlier (13) and suggested that it has a potential benefit, although it does have a few potential limitations that need to be addressed, e.g. disregarding individual ion intensity information, use of only linear calibration functions, and lack of control of the m/z range covered by putative calibrants. Recently, another MS/MS data set recalibration tool has been developed (24) that incorporates the time component into recalibration equation. However, in this case, the authors assumed linear relationships of MME and time, which is rarely the case (13) and may serve only as a rough approximation. In line with our previous report (13), we recommended using a multidimensional non-parametric recalibration, an approach that is not limited by the disadvantages mentioned above.

To derive practical benefits from our previous study on systematic MME behavior for the proteomics community, here, we developed an algorithm for eliminating systematic biases in the parent ion MME for the MS/MS data sets and implemented it into a software tool. This tool, DtaRefinery, is designed to work in tandem with either extract_msn or DeconMSn (Fig. 1). DtaRefinery first reads a set of fragmentation spectra from a concatenated .dta file (supplemental Fig. 1) produced by either extract_msn or DeconMSn. Next, it internally calls an MS/MS search engine to identify putative peptides based on matching MS/MS fragmentation patterns against an appropriate, user-specified FASTA file containing sequences of proteins expected to be present in the sample. Conceptually, it does not matter which MS/MS search engine is used, although we prefer X!Tandem as it is free, open source, and relatively lightweight. X!Tandem is included in the package: there is no need for installation of a search engine. All the database searching is done behind the scene, and the generated files with MS/MS search results exist only temporarily and are deleted after the parsing step. DtaRefinery then computes the parent ion MME based on observed masses and theoretical monoisotopic masses derived from peptide sequences. In the next step, it examines the parent ion MME of the peptides for dependences on scan number, m/z, log₁₀ of ion intensity, and TIC. If dependences are found, DtaRefinery trains a regression-based prediction model for the systematic components of the MME (Fig. 2). If an estimated prediction error of the regression model indicates an improvement of MME, then the model is applied to correct the observed parent ion masses within the entire MS/MS data set. This process is applied iteratively until no systematic MME dependences are detected for all of the considered explanatory variables (e.g. m/z, scan number, log₁₀ of ion intensity, and TIC). At this final point, a new concatenated .dta file is created with corrected parent ion masses. It also produces quality control images, allowing the researcher to visually explore the behavior of the MME and the log file with the records on all the processing steps and potential errors.

Fig. 1.

Flowchart showing position of DtaRefinery in MS/MS data processing pipeline. The first step is extraction of MS/MS data from a binary file with DeconMSn or extract_msn. Next, the extracted MS/MS data are processed with DtaRefinery or alternatively can be directly used for searching the peptide identifications. The format of the refined data produced by DtaRefinery is the same as originally extracted by DeconMSn or extract_msn. Finally, the refined MS/MS data can be searched using the MS/MS search engine of choice. Note that DtaRefinery uses the X!Tandem MS/MS search engine. It is incorporated into the tool and independent of the search engine of choice used in the pipeline.

Fig. 2.

Example showing correction of highly pronounced systematic parent ion MME along dimension of scan number parameter. The example is an actual LC-MS/MS analysis on an LTQ Orbitrap instrument that is out of calibration with significant sample overloading. Because of sample overloading, the automatic gain control system was not able to properly modulate the ion population within the Orbitrap cell, resulting in space charge effects causing noticeable systematic MME. However, after applying the DtaRefinery and subtracting the systematic MME components predicted by the regression models trained in the space of all four parameters (scan number, m/z, log₁₀ of ion intensity, and TIC), the mean of the MME distribution shifts from −16 ppm to approximately 0 ppm, and the standard deviation contracts from 4.3 to 0.8 ppm (data not shown). A, the individual parent ion MME plotted as a function of scan number (blue circles). B, smoothing the MME residuals with Tukey's running median (yellow circles). C, fitting a spline function into smoothed data to have a continuous function for prediction of systematic MME (red line). D, corrected parent ion MME by subtracting the systematic MME predicted by the model trained using only the scan number parameter.

MATERIALS AND METHODS

Software Implementation

DtaRefinery is implemented using the Python programming language and depends on the wxPython, Matplotlib, and NumPy open source Python libraries for graphical user interface (GUI), plotting images and numerical computations, respectively. For peptide identifications, the program uses X!Tandem (4), which is also freely available and open source. However, there is no need to install X!Tandem because it is shipped within DtaRefinery. Optionally, the R statistical environment and R(D)COM server application can be used for smooth spline fitting, which minimizes the sum of squared errors plus weighted squared secondary derivative (smooth.spline() function). DtaRefinery has a GUI but also can be run at the command prompt for batch automation. The software tool is available in two formats, both as a stand-alone executable version for the Microsoft Windows 32-bit platform or as a platform-independent collection of Python scripts. If installed as the stand-alone executable version, there is no need to install Python or any prerequisite libraries.

Features and Algorithm

The concept behind the parent ion MME refinement process is shown in Fig. 3. The input is a concatenated .dta file that contains MS/MS data and a FASTA file with anticipated protein sequences. Optionally, the DtaRefinery will utilize the _DeconMSn_log.txt and _profile.txt files along with a concatenated .dta file created by DeconMSn while processing a Thermo raw file. This additional information includes the MS scan at which each parent ion was detected, parent ion intensity, TIC in the Orbitrap or FTICR cell, and the automatic gain control accumulation time for the MS scan. These extra parameters can be utilized as optional explanatory variables for regression analysis to find and correct systematic error dependences. Thus, the use of DeconMSn instead of extract_ msn is highly recommended because it not only accurately determines the monoisotopic masses of the parent ions but also provides additional useful information about parent ion measurement.

Fig. 3.

DtaRefinery process flowchart. The data set is analyzed by an MS/MS search engine against an expected list of protein sequences. Spectra with identified peptides go into Table 1, which will be further used for training a regression model predicting systematic MME. Table 2, which is used to store the predicted systematic MME, contains all spectra irregardless of assigned peptides. After the model is trained, the parent ion masses in the original data set are corrected based on the predicted systematic MME values that are stored in Table 2 and written into an updated MS/MS data file (“*_FIXED_dta.txt”).

DtaRefinery analysis settings can be customized by editing the XML file containing processing options or using the GUI of the program. The concatenated .dta file can be produced by either DeconMSn (-XCDTA option) or concatenation of individual .dta files produced by extract_msn (using the Peptide File Extractor). Concatenated .dta files (*-dta.txt) contain each fragmentation spectrum in the .dta format separated by a header line describing the source .dta file (supplemental Fig. 1).

Once spectra are loaded, DtaRefinery internally uses X!Tandem to briefly search the MS/MS data entries from the input file to identify peptides. Note that the MS/MS search does not have to be exhaustive. Capturing the majority of the peptides should be sufficient for recalibration. For example, we suggest searching global tryptic digest samples only for fully tryptic peptides and ignoring partially tryptic and non-tryptic peptides. This non-exhaustive search, restricted to the space of fully tryptic peptides, is ∼1–2 orders faster than searching for all peptide types yet still captures close to 90% of all identifiable peptides. For the purpose of sample recalibration, those ∼10% of semi- and non-tryptic peptides can be safely ignored. Another way of speeding MS/MS searches can be achieved by narrowing the parent ion MME tolerance. For example, we use a parent ion MME tolerance for preliminary peptide identification searches in the range of ±20–100 ppm for data sets obtained using high resolution hybrid mass spectrometers. This tolerance range allows for a faster search than the ±3-Da tolerance commonly used for data sets from ion trap instruments. To assess the gain in speed, we timed X!Tandem searches for a mouse brain tryptic digest peptide sample that was analyzed using a 30-min-gradient LC separation and an LTQ Orbitrap mass spectrometer. With a ±3-Da parent ion MME tolerance and no enzyme rule setting, searching the test data set took ∼9 h or ∼250 ms/spectrum/megabyte of database (on a single thread). With the parent ion MME setting adjusted to ±100 ppm and a requirement that peptides be fully tryptic, the search took ∼1 min or 0.5 ms/spectrum/megabyte of database. This comparison illustrates a ∼500-fold increase in processing speed and indicates that the MS/MS search is not a “bottleneck” or time-consuming part of the DtaRefinery work flow compared with the regression analysis.

After the MS/MS search is completed, the MS/MS scan entries that identify peptide sequences with X!Tandem E-values higher than a specified threshold are compiled into a table (further referred as “Table 1”). Our default E-value threshold is 0.01; i.e. the chance that the spectrum-peptide assignment is wrong is at most 1 in a 100. The table contains the MS/MS scan number, charge state, MS scan number, theoretical parent ion m/z value, and the difference (in ppm) between the observed and theoretical m/z value. If the _DeconMSn_log.txt and _profile.txt files from DeconMSn output are available, then the table will also contain the parent ion intensity and TIC for the corresponding MS scan. The information in Table 1 is used to examine potential MME trends versus specified parameters as well as to train an additive regression model that explains such trends. DtaRefinery also compiles a second similar table, “Table 2”; however, unlike Table 1, this table includes all MS/MS scan entries without exception, including entries that are not identified by a peptide sequence. The systematic MME residuals in Table 2 were initially set to 0 ppm for all entries; i.e. the starting assumption is that there are no systematic errors. The purpose of Table 2 is to store the predicted systematic MME for all of the MS/MS scan entries during iterative training of the regression model.

The rationale behind the prediction algorithm has been reported in detail in an earlier publication (13). Briefly, the MME residuals are plotted as a function of elution time, m/z, log₁₀ of ion intensity, or other explanatory variables. After visualizing the scatter plots, it is usually apparent whether a systematic error is present or not (Fig. 2A). Although such systematic error trends can be readily modeled with non-parametric regression or scatter plot smoothing techniques, a few concerns must be taken into account. These concerns include the presence of false peptide identifications, the multidimensionality of the problem, overfitting, and finally the computation cost.

Earlier, we suggested modeling the systematic MME using a multidimensional projection pursuit regression (27, 28) (Equation 1).

where β denotes the coefficients of the optimal projection and g denotes the non-parametric regression functions. Error residuals are modeled as a hypersurface in space with user selected parameters and MME as dimensions. The regression is iterative; at each step, the model finds one-dimensional projections as linear combinations of the space parameters that best explain the observed data and then subtracts the residuals. Because the MS/MS searching results may contain false peptide identifications, Tukey's running median (29) (Fig. 2B) smoothing is initially applied to mitigate the effect of outliers. Next, a non-parametric regression technique, such as smoothing splines (30) or LOWESS (31) (Fig. 2C) is applied for modeling the trend and predicting the systematic error values. Finally, the predicted systematic error values are subtracted from the actual error residuals (Fig. 2D). Finding the best one-dimensional projections involves optimization techniques, which are computationally expensive. Hence, in the DtaRefinery tool, for reasons of practicality and the speed of computation, we used a simplified additive model (Equation 2) instead of a more sophisticated projection pursuit regression (Equation 1). This approach avoids having to search for optimal projections and instead performs the regressions along the parameter dimensions. Consequently, it is much less computationally expensive, although it may not capture potential complicated interparameter dependences.

For each iteration, DtaRefinery loops through and estimates the prediction error of the regression models for each of the parameters (supplemental Fig. 2). The prediction error is estimated using one of the standard machine learning approaches, namely K-fold cross-validation as root mean squared error (RMSE) or approximately standard deviation (supplemental Fig. 3). As such, the data are split into K parts. The regression model is trained on K − 1 parts, and the systematic MME is predicted for the data points in the reserved part. The procedure is repeated K times to ensure that data points from every part contribute toward the estimated prediction error of the model. After completion of all K rounds, the RMSE is computed from the MME residual left after subtraction of the predicted systematic MME components. By default, the K-value is set to 10 but can be changed by the user. If the prediction error is higher than the original RMSE prior to the regression, then the regression fit is not considered successful because the model is distorting the data either by overfitting or some other way. When the iteration is successful, i.e. at least one of the used parameters proved to be a successful explanatory variable, a parameter providing the lowest prediction error is selected, and the corresponding predicted MME residuals are subtracted from the mass error values in both tables. The prediction error not only serves as a criteria for selecting the best explanatory variable but also as an indicator of whether any systematic trend is present that regression can remove. The iterations stop when none of the parameters provide a successful regression model for the round. Thus, although there is no limit to the number of iterations, in our experience it usually takes ∼5–20 iterations to converge.

Depending on the user-specified options, the MME residuals for a single iteration can be modeled in one of two ways: either as is or using the overfitting proof mode. Use of the overfitting proof mode may be especially important for sparse data sets or sparse areas of the LC-MS/MS data sets where the danger of interpolation exists. Examples of sparse areas include the flanks of chromatograms and the high and low extremes of m/z or ion intensity values. Interpolation of the data points artificially removes not only systematic but also random errors, which can lead to undesirable effects. Overfitting can be avoided by making sure that the MS/MS scan entries used to train the regression are not in the subset for which the error is predicted. This can be achieved by using an approach quite similar to K-fold cross-validation. In particular, we split the entries for which the predictions need to be made into N + 1 partitions. The first partition contains the entries of non-identified spectra that are exclusively present in Table 2 and absent from Table 1. At this stage, overfitting is not an issue because the points used for training and prediction do not overlap. However, overfitting potentially may be an issue for scans that are present in both Table 1 and Table 2. In this case, the identified spectra entries in both tables are split into N equal partitions. To predict the systematic MME for the entries in each of the N partitions in Table 2, entries from the corresponding remaining N − 1 parts from Table 1 are used to train the regression model (Fig. 4). By default, DtaRefinery uses the overfitting proof mode with N set to 10, which splits the identified MS/MS entries into 10 parts.

Fig. 4.

Outline of overfitting proof regression option. The data set is randomly split into N equal parts. For each part, the regression is trained on the remaining N − 1 parts, and the learned function is evaluated on the reserved part followed by subtraction of the predicted values. In such an approach, the MME residuals for which the systematic component is predicted are never used for model training.

As mentioned earlier, the iterations are stopped when further regression using any of the parameters does not improve the MME residuals distribution, i.e. the RMSE after trial regressions is not smaller than the RMSE before the regression attempt. At this point, the estimated contributions of the systematic MME in Table 2 are considered final. DtaRefinery uses these final estimates of the systematic MME to correct the parent ion masses and output the refined MS/MS data to form a new concatenated .dta file. Additionally, DtaRefinery outputs the mass error distribution histograms before and after parent ion mass refinement plus scatter plots that show original and final dependences of MME on parameters selected for building the additive regression model. The mean and S.D. for the MME distribution are estimated in two alternative ways. The first estimate is computed using the expectation maximization approach, which models errors as a mixture of a normal distribution of true identifications and a uniform distribution of false peptide identifications. This approach is more appropriate for data that contain a significant amount of false positive identifications because they can be explicitly modeled as a uniform distribution. Note that the uniform distribution approximates the density of false identifications well enough only within an MME window of 20–30 ppm or less, although this is sufficient for high resolution instruments like the LTQ FT and LTQ Orbitrap. The second approach treats false peptide identifications as outliers and estimates the mean as the median value and the standard deviation using the median absolute deviation. This approach tends to be more robust for cases in which the histogram of true identifications noticeably deviates from a normal distribution. Also note that DtaRefinery can optionally output scatter plots corresponding to regression fits along each of the parameters for each of the iterations.

Sample Preparation and LC-MS/MS Analysis

The tryptic peptides from four different mouse brain regions (cortex, striatum, cerebellum, and the rest of the brain) were prepared and fractionated by strong cation exchange chromatography as described elsewhere (32). Each of 32 fractions and unfractionated samples were analyzed by a liquid chromatography system with ∼200 full-width half-maximum peak capacity. The 75-μm-inner diameter × 15-cm-long columns were packed with 3-μm Jupiter C₁₈ particles (Phenomenex, Torrance, CA). The mobile phase solvents consisted of 0.1% formic acid in water (A) and 0.1% formic acid in 90% acetonitrile (B). An exponential 35-min gradient was used for the separation, starting with 100% A and gradually increasing to 60% B over 35 min at a constant pressure of 5,000 p.s.i. The column was coupled with an LTQ Orbitrap (Thermo Fisher Scientific Inc., Waltham, MA) mass spectrometer. The instrument method was as follows: MS survey scan having 100,000 full-width half-maximum resolution and 1 × 10⁶ automatic gain control target followed by five MS/MS scans analyzed by the ion trap (∼1,000 resolution). All 136 LC-MS/MS data sets are available upon request.

RESULTS AND DISCUSSION

Here, we evaluated the effect of improving parent ion mass measurement accuracy in the context of a “cataloguing” type of proteomics study wherein the final product or result is a non-redundant list of the peptides and proteins observed in a given biological sample. To assess the performance of such proteomics analyses, it is important to estimate two quality metrics: false discovery rate (FDR) and false negative rate (FNR) both for peptide and protein identifications. For the sake of simplicity, we will limit ourselves here with estimation of FDR and FNR for non-redundant peptide identifications only. For this type of study, FDR should be defined as the proportion of non-redundant false peptide entries within the non-redundant peptide identification entries, whereas FNR is the ratio of the number of non-redundant true peptides not passing the threshold criteria required for confident identification to the total number of non-redundant true peptides.

As a demonstration, we used the results from 136 LC-MS/MS analyses of a tryptic digest mouse brain sample prefractionated with strong cation exchange chromatography. To refine these data sets, we selected a non-linear, LOWESS-based, additive regression model with the overfitting proof option enabled. Fig. 5 demonstrates an example of a quality control output of the DtaRefinery tool, showing MME residuals as a function of multiple parameters before and after refinement for a typical data set. For this particular LC-MS/MS data set, prediction followed by subtraction of the systematic errors effectively shifted the overall systematic bias from −2.9 to practically 0 ppm and reduced the standard deviation ∼1.5-fold from ∼1.0 to ∼0.65 ppm as evident from another refinement quality control plot (Fig. 6). Note (supplemental Figs. 4 and 5) that linear recalibration is not as efficient as one that is LOWESS-based because there are still some obvious residual trends in the scan number and m/z domains. The resulting decrease in the standard deviation, although also being quite significant, is not as large. For example, according to Expectation Maximization estimates, the standard deviation decreases from 0.96 to 0.85 ppm in linear recalibration mode and decreases to 0.69 ppm in the case of LOWESS.

Fig. 6.

Example of typical parent ion mass error distribution histograms before (left, blue) and after (right, green) applying additive regression refining procedure. The bin width is 0.5 ppm. The red dashed line shows the Expectation Maximization (EM) fit of the mixture distribution, modeled as normal for true identifications and uniform for false identifications. The parameters of the mixed model distribution were estimated by an expectation maximization procedure. The light blue dashed line shows the normal distribution with robust estimates of mean and S.D. The maximum allowable deviation improves from about ±5 to ±2 ppm. est., estimate; stdev, standard deviation.

As noted earlier, a stricter requirement for the maximum allowable parent ion mass deviation serves to improve discrimination between true and false peptide identifications (8, 11, 12, 18). Thus, we expect that preprocessing the MS/MS data sets by removing the systematic component of the MME of the precursor ions should appreciably improve the quality of peptide identifications. For example, in the demonstrated case (Fig. 6) on the original data set, one needs to use no less than a ±6-ppm threshold to retain most of the true peptide identifications. However, after refinement, the maximum allowable deviation may be decreased to ±2 ppm, providing the potential for an ∼3-fold decrease in the FDR of peptide identifications with almost no loss of true identifications. The other benefit of reducing the maximum allowable tolerance for parent ion MME may be a reduction in the FNR by including poorly fragmenting peptides that receive MS/MS fragment match scores below a specified score threshold. These peptides have the potential to be matched by loosening MS/MS match score thresholds and simultaneously applying more stringent requirements for parent ion mass deviation.

To assess the effect of elimination of the systematic MME component, we searched all the above mentioned 136 LC-MS/MS data sets against mouse International Protein Index database version 3.51 combined with sequences of human keratin proteins and pig trypsin peptides. For assessing the number of true and false identifications after the search is completed, we created a decoy database by concatenating the combined FASTA file with itself but with reversed protein sequences (17). All of the identifications from reversed sequences were considered false, whereas identifications from forward sequences could be either true or false. False identifications were assumed to be distributed equally between matches to forward and reversed sequences. The data sets were preprocessed by DeconMSn to generate the concatenated .dta files. For peptide identification analysis, we used both the original concatenated .dta files as well as files processed by DtaRefinery. The peptide identification was accomplished using the SEQUEST search engine. However, qualitatively, the results and conclusions should stay the same regardless of the MS/MS search engine utilized. The SEQUEST searches were performed both ways: restricting peptide sequences only to the ones that satisfy trypsin cleavage specificity and with no restriction to cleavage specificity. Fig. 7 shows the parent ion MME distribution histogram for all the spectra with assigned peptides for all 136 data sets collated together. Note that this is a rather typical data analysis arrangement, i.e. where all the data sets are processed at once with a fixed maximum allowable parent mass deviation. However, sequentially “one-by-one” processing all the data sets and applying individual tolerance criteria would provide better results but would require specialized software (such as DtaRefinery). Without such correction of the systematic errors, one would have to allow up to ±10-ppm mass measurement error for identified peptides to retain the majority of the true identifications. Removal of the systematic error can be as simple as zero centering the entire histogram that is shifting by about −2.5 ppm with corresponding recalculation of parent ion masses, shifting and recalculating the parent ion masses for the individual LC-MS/MS data sets as suggested before (26), or as sophisticated as applying multidimensional non-parametric regression models to the individual data sets (Fig. 7B). In the latter case, the maximum allowable deviation of mass measurement error can be reduced to as low as 3 ppm. Such a reduction provides leverage for improvement of peptide identification results, which can be quantified by a decrease both in FDR and FNR of the non-redundant peptide identifications. We should also note that despite noticeable shift of partially and non-tryptic peptide observations toward higher m/z values the number of fully tryptic peptides was sufficient to capture the trend, correct for systematic errors, and reduce the overall spread for the two other types of peptides (supplemental Fig. 6).

Fig. 8 shows the estimate of the maximum number of non-redundant true peptide identifications with 2+ charge for all 136 data sets that can be achieved with a given allowed maximum FDR. The maximum number of non-redundant peptide identifications within the given FDR was selected by searching for the best combination of ΔMass, XCorr threshold, and ΔCn within the ranges from 1 to 10 ppm ΔMass (with a 0.1-ppm step size), from 0 to 8 XCorr (0.1 steps), and from 0 to 0.4 ΔCn (0.01 steps), giving 288,000 combinations in total. Clearly, in most of the cases and especially in the range of reasonably low FDRs (up to 5%), searches of the refined data sets deliver better results. In other words, for any given number of true identifications, it is possible to achieve significantly lower FDRs. Conversely, for any given FDR limit, improvements in mass measurement accuracy allow us to obtain more peptide identifications. The MS/MS fragmentation spectra of those rescued peptides probably did not contain enough information to meet the elevated MS/MS fragment matching criteria because of physicochemical properties unfavorable for fragmentation, low intensity, or various other reasons. Table I lists several specific values of true and false peptide identifications shown on Fig. 8. As shown, the average reduction of FDR is about 2-fold. The actual FDR decrease varies from 1.3- to 3-fold and is based on search type and the absolute FDR values. For example, in the case of non-restricted SEQUEST searches, after applying 8.9 ppm, 2.3, and 0.17 for ΔM, XCorr, and ΔCn criteria, respectively, it is possible to achieve the maximum number of 13,380 true peptide identifications within the 1% FDR limit. Notably, after refining the data sets and tightening the ΔM criteria to 2.9 ppm with simultaneous slight relaxation of ΔCn to 0.16, the number of false peptide identifications goes down almost 3-fold from 132 to 48 while maintaining the same or even larger number of true identifications.

Fig. 8.

Estimates of maximum number of true positive unique peptides with 2+ charge that can be identified by SEQUEST within allowed FDR values. The green curve represents the results of SEQUEST searches of 136 LC-MS/MS data sets preprocessed by DtaRefinery. The results of non-preprocessed data sets are shown in blue. The maximum number (max. num.) of peptide identifications (ids) was obtained by searching the space of ΔM, XCorr, and ΔCn parameters within the ranges from 1 to 10 ppm ΔMass (with a 0.1-ppm step size), from 0 to 8 XCorr (0.1 steps), and from 0 to 0.4 ΔCn (0.01 steps), giving 288,000 combinations in total. Note that for any given FDR value the results from refined data set searches provide more true peptide identifications. It is also true that for a given number of true peptide identifications it is always possible to achieve noticeably lower FDR by preprocessing the LC-MS/MS data sets with DtaRefinery.

Table I. The effect of preprocessing of LC-MS/MS data sets by DtaRefinery on FDRs and FNRs of unique peptide identifications (IDs) from SEQUEST analyses.

The results are presented for the 2+ charge state. A, searches only for fully tryptic peptides; B, searches with no enzyme rules applied.

Putatively true unique peptide IDs	Putatively false unique peptide IDs	FDR	FNR	Optimized threshold criteria
				ΔM	XCorr	ΔCn
		%	%	ppm
A
Original data sets
18,933	2,080	10.0	7	9.4	1.7	0.07
17,897	934	5.0	12	8.5	1.8	0.12
15,275	154	1.0	25	8.5	2.2	0.19
Refined data sets (reduced FDR, constant FNR)
18,957	1,548	7.5	7	4.1	1.6	0.05
17,900	616	3.3	12	3.4	1.8	0.08
15,295	74	0.48	25	3.1	2.1	0.19
Refined data sets (constant FDR, reduced FNR)
19,280	2,132	10	5	4.3	1.6	0.02
18,495	968	5.0	9	3.3	1.7	0.05
16,192	162	1.0	21	2.8	1.9	0.17
B
Original data sets
16,603	1,844	10.0	15	8.8	2.0	0.08
15,480	804	5.0	21	8.6	2.0	0.12
13,349	134	1.0	32	8.6	2.1	0.18
Refined data sets (reduced FDR, constant FNR)
16,614	862	4.9	15	2.5	2.0	0.06
15,504	356	2.2	21	2.9	2.0	0.11
13,368	46	0.34	31	2.7	2.3	0.16
Refined data sets (constant FDR, reduced FNR)
17,663	1,950	10.0	9	2.8	1.8	0.04
16,585	872	5.0	15	2.8	2.1	0.05
14,448	138	1.0	26	3.0	2.3	0.12

The DtaRefinery software has been extensively evaluated at our laboratory for processing LC-MS/MS data sets where parent ion mass has been acquired with high accuracy, including the LTQ Orbitrap and LTQ FT instruments, and has consistently reduced overall MME deviations. However, the extent of the mass accuracy improvement varies due to several factors, e.g. the time between instrument calibrations, the amount of sample loaded for analysis, the length of the LC gradient, the automated gain control settings, and complexity of the sample itself. Regardless, the final values for maximum allowable deviation are typically about ±2 ppm for individual data sets obtained using either LTQ Orbitrap or LTQ FT instruments.

To date, we have used DtaRefinery in our proteomics data processing pipeline (Fig. 1) in tandem with DeconMSn prior to X!Tandem, InsPect, or SEQUEST MS/MS search engines. The fact that DtaRefinery utilized a concatenated .dta file as an input and outputs the concatenated .dta file in the same format allows its ready incorporation into various MS/MS data processing pipelines. The concatenated .dta file can be used for input into a search engine without modification or can be split into individual .dta files or converted into MASCOT generic format with currently available tools. Such files can be further converted into a number of other formats if necessary (33). In the future, we foresee other preprocessing approaches, e.g. more sophisticated peak picking in MS/MS spectra, MS/MS spectra recalibration (as implemented in VEMS (34) or in MS-Dictionary (26)), and others put together as additional components in the MS/MS data processing pipeline.

Based on our literature survey of studies utilizing LTQ Orbitrap or LTQ FT instruments for shotgun proteomics to date, the maximum allowable MME deviation is typically set between ±5 and ±10 ppm (e.g. Refs. 35 and 36). DtaRefinery clearly could be useful for these studies as it can generally help reduce the parent ion MME tolerance severalfold (typically to ±2 ppm for these instruments) and thus reduce the number of both false positive and false negative peptide identifications. Given the increasing use of such hybrid MS instrumentation, we anticipate that the DtaRefinery software tool can be widely used in tandem with DeconMSn, in particular for those proteomics applications in which peptide identification confidence remains a challenging issue because of a significantly increased search space. Applications where DtaRefinery can benefit the most include identification of peptides with post-translational modifications, identification of peptides resulting from nonspecific proteolysis, and searches using exhaustively translated genomes in all six reading frames from stop-to-stop codons as a set of putative protein sequences.