JUMPlib: Integrative Search Tool Combining Fragment Ion Indexing with Comprehensive TMT Spectral Libraries

Abstract

The identification of peptides is a cornerstone of mass spectrometry-based proteomics. Spectral library-based algorithms are well-established methods to enhance the identification efficiency of peptides during database search in proteomics. However, these algorithms are not specifically tailored for tandem mass tag (TMT)-based proteomics due to the lack of high-quality TMT spectral libraries. Here we introduce JUMPlib, a computational tool for TMT-based spectral library search. JUMPlib comprises components for generating spectral libraries, conducting library searches, filtering peptide identifications, and quantifying peptides and proteins. Fragment ion indexing in the libraries increases the search speed, and utilizing experimental retention time of precursor ions improves peptide identification. We found that methionine oxidation is a major factor contributing to large shifts of peptide retention time. To test the JUMPlib program, we curated two comprehensive human libraries for the labeling of TMT6/10/11 and TMT16/18 reagents, with ~286,000 precursor ions and ~304,000 precursor ions, respectively. Our analyses demonstrate that JUMPlib, employing the fragment ion index strategy, enhances search speed and exhibits high sensitivity and specificity, achieving approximately a 25% increase in peptide-spectrum matches compared to other search tools. Overall, JUMPlib serves as a streamlined computational platform designed to enhance peptide identification in TMT-based proteomics. Both the JUMPlib source code and libraries are publicly available.

Graphical Abstract

1. Introduction

Tandem Mass spectrometry (MS)-based shotgun proteomics has emerged as a powerful analytical tool for identifying and quantifying the large-scale proteome, post-translational modifications, and higher-order structures and networks ^{1, 2}. Peptide identification in shotgun proteomics is a critical component of the workflow. In the past two decades, three primary methods have been used for identifying peptides, namely de novo sequencing, reference database search, and spectral library search. De novo sequencing algorithms directly infer sequence tags from MS/MS spectra, considering possible combinations of amino acids ^{3, 4}. In contrast, reference database search generates theoretical mass spectra by in-silico digestion of a protein reference database, that is then compared with experimental tandem mass spectra to identify peptides ^{5, 6}. While reference database search is widely used and capable of identifying sample-specific new peptides compared to spectral library search, it can be time-consuming, as it repeatedly searches for peptides that are not present in samples ^7-9.

Spectral library search has emerged as an alternative approach that catalogs high-confidence peptide spectra from the reference database search into a searchable reference library, enabling new identifications through spectral matching ^{7, 10}. This approach offers high sensitivity, rapid, and reproducible identification of peptides ¹¹. Over the past decade, several spectral library search tools, such as SpectraST, COSS and X!Hunter, have been developed to analyze data-dependent acquisition (DDA) MS data^{7, 12-18}, while spectral library-based data-independent acquisition (DIA) pipelines, such as Skyline ¹⁹, EncyclopeDIA ²⁰, DIA-NN ²¹, are available to process MS data. These tools have significantly improved the sensitivity of peptide identification. Spectral library search tools have also been designed for identifying post-translationally modified peptides ^22-25 and cross-linked peptides ²⁶, and sample-specific libraries containing spectra specific to the sample being analyzed have also been constructed ^27-29.

Tandem mass tag (TMT)-based quantitative proteomics has gained considerable popularity owing to its ability to multiplex samples and improve throughput. While TMT offers numerous advantages over label-free approaches, such as increased sample throughput and reduced technical variation, a major challenge lies in the absence of TMT-specific libraries. To overcome this limitation, we previously introduced 10-plex TMT spectral libraries and reported that TMT labeling can alter the fragmentation patterns of peptides ²⁹. Recently, predicted TMT-labeled peptides, such as the Prosit-TMT library ³⁰ was also introduced to the field. However, as the multiplex capacity (e.g., 11-, 16-, and 18-plex) has advanced further, empirical libraries supporting these higher plexes have remained largely unavailable. Furthermore, spectral library search tools designed for these higher plex TMT libraries are lacking. Therefore, there is potential to further enhance peptide identification sensitivity and search speed by curating higher plex TMT libraries and creating advanced tools.

In this study, we introduce JUMPlib, a specialized tool designed for searching TMT-based proteomics data. The JUMPlib program enables TMT library generation, library search, identification filtering, and protein quantification (as shown in Figure 1). To evaluate the performance of the JUMPlib program, we conducted an assessment using a large-scale TMT data set. In addition, the JUMPlib program can readily be adapted for label-free library generation and library search. The source code for JUMPlib program is freely accessible to the research community. Moreover, we curated comprehensive 11-plex and 18-plex TMT libraries from human brain samples, providing valuable resources to the research community.

Figure 1. Schematic diagram illustrating the implementation of the JUMPlib program.

(A) Spectral library generation. The spectral library is generated from high-quality identifications of TMT-MS datasets. (B) Spectral library search. TMT MS/MS spectra are searched against the spectral library, and candidate matches are ranked using the normalized dot product (JDscore). (C) Filtering and quantification. The identified candidates are filtered based on the JDscore and retention time (RT) and quantified using reporter ions.

2. Methods

2.1. Implementation of JUMPlib program

2.1.1. Spectrum preprocessing

The spectrum preprocessing procedure implemented in JUMPlib aims to improve the quality of spectral data. The preprocessing takes mzXML or mzML file format as an input, followed by two analyses: deisotoping and mass correction. The deisotoping involves identifying MS1 isotopic windows using MS2 precursor $m∕z$ values and selecting peaks within those windows. The algorithm assumes charges ranging from 2 to 5 to identify potential isotopic clusters. For each charge, a reference distribution is generated using an averagine model and the precursor $m∕z$. A theoretical isotopic distribution is then computed based on this reference distribution. Next, dot products value is calculated between the experimental peaks and the theoretical peaks of the isotopic distribution. These dot products measure the similarity between the experimental and theoretical peaks, aiding in the identification of the most representative isotopic peaks. The top two precursor $m∕z$ values are selected as the output of the deisotope process. Finally, the mass correction step is applied to the preprocessed MS1 and MS2 spectra. This process involves adjusting the mass values of the identified peaks based on the $m∕z$ of the TMT-126 reporter. Briefly, the experimental $m∕z$ values are extracted from the raw mass spectrometry data followed by the calculation of the mass shift by comparing the observed $m∕z$ of the TMT-126 reporter ion to its theoretical value. Thus, calculated mass shift (ppm) is applied as a correction factor to adjust the $m∕z$ values of both precursor and all product ions. This aligning step significantly enhances the mass accuracy, which is essential for reducing mass measurement errors and improving the reliability of the proteomic analysis. The ultimate output of the preprocessing is an .ms2 data file containing the processed spectra.

2.1.2. Construction of TMT spectral library

TMT spectral library construction in JUMPlib involves four main components: (i) library generation and quality control, (ii) retention time (RT) assignment and alignment, (iii) library merging, and (iv) decoy library generation. To generate a high-quality spectral library, a rigorous quality control (QC) process was established for selecting precursors (Figure 2A; Figure S1). We utilized the JUMP tool to identify peptides as an input to generate spectral library. JUMP is a comprehensive proteomics software suite designed for the identification of peptides and proteins from mass spectrometry data ⁹. The combined score in JUMP is the Jscore, a search score used to evaluate the confidence of peptide identifications. Initially, Jscore threshold of greater than 30 was set to identify potential precursor candidates. These candidates were then consolidated into representative library precursors (Lprecursors) based on the "top N precursors" user-defined parameter. For Lprecursors with numerous precursors, only the top N were selected for consolidation. Conversely, if an Lprecursor had fewer than N precursors, all were included in the consolidation process. An additional layer of QC was then applied to the pre-consensus library, where each Lprecursor underwent similarity checks using the scaled dot product score (DP) which ranges from 0 to 1, with a cutoff value of greater than 0.8 to re-select the candidate precursors when an Lprecursor had multiple associated precursors. If an Lprecursor had only one precursor and a DP less than 0.8, that precursor was retained, and this was noted. Using these re-selected precursor candidates, a final consolidation was carried out to create the consensus library, which comprises unique Lprecursors for each batch. During this consolidation step, the intensities of the precursor candidates are summed, and the retention time (RT) information is stored for each Lprecursor in the library. To assign precursor RT, RT values associated with each scan of identified PSMs were extracted and grouped based on their specific charge and sorted RT (Figure 2B). Each unique combination of a peptide and its charge was considered a precursor, and each precursor encompassed a variable range of RT values, spanning from a minimum of one to multiple values. To assign the final RT value for each unique precursor, a user-defined RT sliding window was set. The RT list was clustered based on this sliding window, resulting in one to multiple clusters. The weighted average of retention times (RTs), taking into account both RT and intensity, was calculated for each cluster. We considered four scenarios: 1) a single cluster, 2) multiple clusters each containing two or more RTs, 3) multiple clusters with a mixed population (either a single RT or multiple RTs), and 4) multiple clusters each with a single RT. In scenario 1, only one weighted RT was calculated. In scenario 2, the earliest cluster was used as the representative peak. In scenario 3, any single RTs were removed, leaving only multiple RTs, which were treated as either scenario 1 or 2. In scenario 4, the RT corresponding to the peak with the highest intensity was chosen. This process of RT assignment was repeated for all experimental runs, aligning each target run with the adjacent reference run. During the alignment process, we utilized shared precursors between the reference and target runs to generate a Lowess fitting curve (Figure 2B). Lowess fitting is a non-parametric technique that is independent of specific distribution ³¹. The RTs from the reference, including unique and shared precursors, were retained, and the calibrated RTs were integrated into the reference along with the precursor information to create a consensus run (i.e., new reference run). This entire procedure was repeated for the entire dataset, and a sliding cluster-based RT inference was employed to define the final RT for each unique precursor. To merge the libraries, we initially identified three types of Lprecursors in the reference and target libraries: (i) Lprecursors unique to the reference library, (ii) Lprecursors shared between the reference and target libraries, and (iii) Lprecursors unique to the target library (Figure 2C). We used the shared Lprecursors to generate a fitting curve using Lowess. The curve generated by Lowess fitting was then used to calibrate the retention time (RT) of all precursors in the target library. In addition, we selected the shared Lprecursors based on the best-scoring entry in the library. This allowed us to use all three populations of Lprecursors (unique to reference, unique to target, and shared) to create a merged library. The merged library was served as the new reference, and another library from the remaining experimental runs was selected as the target library for the subsequent round of merging. This process was repeated until all libraries were merged into a single library (Figure 2C).

Figure 2. Spectral library generation by the JUMPlib program.

(A) Flowchart illustrates the process of library generation, starting from peptide-spectrum matches (PSMs) to consensus precursors. (B) Retention time (RT) extraction is performed individually in each run, followed by RT alignment across multiple runs. Examples of Pearson correlations (R) between reference (run #1) and target (run #2) runs. (C) The libraries generated from reference and target sources are merged to create a comprehensive spectral library.

JUMPlib offers two methods for generating decoys: precursor swap and fixed precursor mass shift ³². For the precursor swap, several modifications were made to the original technique ³³. Modifications to the original precursor swap technique include sorting target precursors by mass ($m∕z$) and charge ($z$) to prevent random overlaps between targets and decoys. Each target is compared against decoy candidates with the same charge state and located outside a user-defined swapping distance (+/−d in $m∕z$). Candidates beyond this distance are sorted by their absolute $m∕z$ difference from the target; the one with the smallest difference is chosen as the decoy. This swap retains the original product ions, creating a target-decoy pair, such as precursor X and Y swapping roles. This results in two target-decoy pairs per match. Matched pairs are then added to an exclusion list to prevent further decoy assignments. If a target lacks a corresponding decoy, it is placed on a rescue list. Targets on the rescue list have decoys generated by adding the swapping distance to their m/z, ensuring an equal number of targets and decoys in the database.

The second method involves a fixed mass shift (+/−d $m∕z$) to generate decoys. For odd-numbered entries in the library (1, 3, 5, etc.), the decoy $m∕z$ is calculated by subtracting d from the Lprecursor $m∕z$. For even-numbered entries (2, 4, 6, etc.), the decoy $m∕z$ is calculated by adding d. This systematic approach ensures a consistent generation of decoy precursors throughout the library.

2.1.3. Library searching

The library search in JUMPlib accepts three types of input: .mzXML, mzML and .ms2 files. When mzXML/mzML files are provided, the program performs a presearch process that involves deisotoping and mass calibration (see the preprocessing step) and produces ms2 files.

For each query spectrum, the precursor and charge are initially matched against the library based on the user-specified peptide tolerance (in parts per million, ppm). Candidate peptide-spectrum matches (PSMs) are then compiled into a table. The search program identifies product ions in the library with the highest intensity, determined by a user-defined parameter (top_ions). This parameter is used to filter the candidates by using one of the top ions in a spectrum. For example, if top_ions is set to 1,3, at least one of the top three ions must be present in the library. If this condition is satisfied, the query spectrum is adjusted based on the presence or absence of the library product ions. Library product ions that are absent from the query spectrum are assigned an intensity of zero. This modification ensures that both the query and reference spectra have the same length, which is necessary for calculating DP similarity score, which ranges from 0 to 1.

2.1.4. Fragment ion indexing implemented in JUMPlib

The JUMPlib program utilizes the fragment ion indexing method ^{34, 35} alongside DP similarity scoring. Briefly, all library and query MS/MS spectra are preprocessed by selecting the top n ions per 100 $m∕z$ (e.g., n = 6), with four decimals of the $m∕z$ maintained for each ion, allowing a tolerance of 0.2 ppm at 500 $m∕z$. Relative intensities of each spectrum are calculated by normalizing to the highest peak, setting the intensity of the most intense peak at 100. For each query spectrum, the precursor ion $m∕z$ and a mass tolerance (e.g., 10 ppm) were used to identify library candidates. A fragment ion index is then created for these library candidates by storing all fragment ions in a dictionary format. Ions from the query spectrum are compared against this dictionary. The number of matching ions for each candidate is recorded. Candidates with matches below a specified cutoff (e.g., 3) are excluded to optimize scoring efficiency, especially where few matching pairs exist between the library and query spectra. For each remaining candidate, matching ion pairs are compared in terms of their intensities. All matching fragment ions are used to calculate DP (default setting), entropy score, and hypergeometric score. In addition to the matching score, a coverage score is also calculated by calculating the matched ion numbers to the number of ions in the library. The final score (JDscore) is generated by multiplying the matching and coverage scores. The scoring method is different from the dot product used in SpectraST (Figure S2)

The respective equation for dot product (DP) and JDscore are given below:

$$DP=\frac{\sum I_{expX}{I}_{lib}}{\sqrt{I_{exp}^{2}X{I}_{lib}^2}}$$ eq. 1

where,

$$I_{exp=intensityarrayof{}^m∕_{z}valuesofms2spectra}$$

$$I_{lib=intensityarrayofcorresponding{}^m∕_{z}valuesinms2specta}$$

$$coverage=\frac{\#ofmatchedions}{\#oftotalions}$$ eq. 2

$$JDscore=DPXcoverage$$ eq. 3

2.1.5. Library filtering

PSMs are filtered using a target/decoy approach, as previously described ^{36, 37}. The program provides two options: score only and a combination of score and RT (Figure S1). In the score only method, JDscore is used to compute a q-value for each PSM, and a user-defined cutoff is applied to select accepted PSMs. The final false discovery rate (FDR) is then calculated as the ratio of the total number of decoys to the total number of targets among the accepted ones. The advanced method incorporates delta RT (library RT – experimental RT) along with JDscore (Figure S3). A user-defined JDscore cutoff (e.g., 0.6) is applied to the entire search data, and a 2-dimensional matrix (y-axis = delta RT bins, x-axis = JDscore bins) is generated. The intersection between a delta RT bin and a JDscore bin forms a box that contains candidate PSMs. To compute FDR, the program starts with the best JDscore bins. For example, if the bin size is 0.1, the program begins with the JDscore bin that ranges from 0.9 – 1. The program computes the local FDR (lFDR) for each bin, along with the total number of targets and decoys. The starting point for cumulative FDR (cFDR) is the highest target-containing box with an lFDR < the user-defined value. From this box, the program computes cFDR upwards until cFDR > the user-defined value. It then computes cFDR downwards until cFDR > the user-defined value. This completes the first JDscore bin computation. The program then iteratively computes cFDR for the remaining JDscore bins using the previously determined cFDR. This produces a roughly triangular pattern of passed FDR, with the initiation box as the center of the base.

2.2. TMT data sets used for generating spectral libraries

The TMT data sets were from 70 human brain TMT profiling projects. The samples were processed with a previously optimized protocol ^{38, 39}. The generated MS data were analyzed by JUMP, a hybrid database search algorithm ⁹. The human protein reference database was generated by combining downloaded Swiss-Prot, TrEMBL, and UCSC databases and removing redundancies (human: 170,122 entries; mouse: 96,193 entries), followed by the concatenation of the decoy database. JUMP is a tag-based database search tool for peptide identification. JUMP combines pattern scoring with de novo tag scoring, displaying a significant improvement in PSM scoring. Major parameters included precursor and product ion mass tolerance (±15 ppm), full trypticity, static mass shift for the TMT tag (+229.16293 for 10- and 11-plex, and +304.2071 for 16- and 18-plex) and carbamidomethylation (+57.02146) of cysteine, dynamic mass shift for methionine oxidation (+15.99491), maximal missed cleavage (n = 2), and maximal modification sites (n = 3). Putative PSMs were filtered by mass accuracy and then grouped by precursor ion charge state and filtered by JUMP-based matching scores (Jscore and dJn) to reduce protein FDR below 1%. If one peptide could be generated from multiple homologous proteins, the peptide was assigned to the protein with the highest PSM score based on the rule of parsimony. The identification results were kept in the format of .pepXML for subsequent analysis.

2.3. Search settings for SpectraST, COSS and Comet

A data set was searched against the spectra library using SpectraST, COSS and Comet using most of the default parameters. For SpectraST, the following parameters were changed: (1) precursor isolation window = 1.6 Th; (2) disabling binning and performing peak-to-peak matching; (3) excluding TMT reporter ions; and (4) precursor mass tolerance of 0.02 Da. Precursor mass tolerance of 10 ppm was used for COSS and Comet. The FDR was estimated by FDR = nd/nt, where nt and nd were the numbers of target and decoy hits, respectively. The search results were then filtered by the fval (SpectraST score), Score (COSS score), and Xcorr (Comet score) to reduce the FDR to a certain level.

2.4. Data and code availability

JUMPlib is written in python. All source codes, manuals, example data, and two spectral libraries are available at https://github.com/JUMPSuite/JUMPlib. Raw MS data used for spectral library generation are available via the Synapse repository with identifiers syn5759470 (https://www.synapse.org/Synapse:syn21347564), syn32136022.1 (https://www.synapse.org/Synapse:syn32136022.1) and syn51180043.2 (https://doi.org/10.7303/syn51180043.2), as well as on ProteomeXchange with accession numbers PXD007985 (https://www.ebi.ac.uk/pride/archive/projects/PXD007985), PXD011482 (https://www.ebi.ac.uk/pride/archive/projects/PXD011482), PXD025261 (https://www.ebi.ac.uk/pride/archive/projects/PXD025261), and PXD057366 (https://www.ebi.ac.uk/pride/archive/projects/PXD057366).

3. Results and Discussion

3.1. JUMPlib assigns and aligns experimental RT for all individual ions

The construction of a reliable spectral library is paramount for ensuring confidence in peptide identification. In addition to the accurate mass of precursor and product ions, reliable RT values are essential for a high-quality library. The JUMPlib program incorporates stringent quality controls for RT assignment to ensure the reliability of TMT spectral libraries. These controls include RT extraction, RT alignment, and RT exclusion. In addition, the JUMPlib program implements RT-assisted peptide filtering (Figure S3; see methods).

During the library alignment process, we observed that a vast majority (~98%) of precursor RTs exhibited significant overlap between two fractions as expected. However, it was noted that 1.3% (62/4,940) of the precursors exhibited a deviation in RT > ∣10∣ minutes (Figure S4A). Among these 62 precursors, the vast majority (95%) are those containing oxidized methionine (Figure S4B). These modified peptides exhibited consistent RT shifts across multiple runs (Figure S4C). For example, a peptide, HM@MAFEQEANEK, showed a 20-minute RT shift between fractions in the data (Figure S4D). This observation is consistent with previous studies suggesting that methionine oxidation may alter the conformation of the peptide and affect its affinity for the stationary phase in reverse-phase (RP) chromatography ⁴⁰. Therefore, the JUMPlib provides an option for RT exclusion for peptides containing oxidized methionine during the filtering process.

3.2. JUMPlib provides a comprehensive TMT spectral library of high quality

The TMT spectral libraries curated by the JUMPlib program comprise high-quality peptides obtained from diverse TMT experiments. To build a library, rigorous quality control (QC) measures were applied: (i) only PSMs that pass the stringent peptide FDR < 0.5% were considered; (ii) PSMs that have Jscore >= 30 were used; (iii) a stringent DP > 0.8 was used further to remove less similar PSMs; (iv) the experimental $m∕z$ of product ions and precursor $m∕z$ were replaced with theoretical $m∕z$; (iv) experimental RT of all PSMs for each precursor were used.

In our previous study ²⁹, we introduced 10-plex TMT spectral libraries, which comprised a collection of 392,537 human and 363,138 mouse consensus spectra. These spectra were obtained from a total of 531 and 1,162 LC-MS/MS runs for humans and mice, respectively. Owing to the advancements in the multiplicity of the TMT strategy, this study aims to present a comprehensive TMT spectral library generated by 11-, 16-/18-plex strategies. The updated libraries now consist of approximately 9.9 and 13.8 million PSMs, corresponding to 221,976 and 233,356 peptides, and 16,583 and 18,026 proteins, respectively (Figure 3A). These PSMs were derived from 38 MS runs with 11-plex and 32 MS runs with 18-plex experiments, respectively. Most spectra (264,709/286,205; 92% in 11-plex; and 280,123/304,334; 92% in 18-plex) originated from charges 2 and 3 (Figure 3B). The protein numbers distribution suggests that our data have almost achieved the detection limits (Figure 3C).

Figure 3. Summary of spectral libraries.

(A) A summary table showing the numbers of features in 11- and 18-plex libraries. (B) Histogram showing the percentage of peptide charges in 11- and 18-plex libraries. (C) Numbers of proteins identified with an increase in TMT experiments. The x-axis represents the number of experiments (from 1 to 38) and the y-axis represents the number of proteins. (D) Up-set plot showing the numbers of precursors and peptides identified by 11- and 18-plex. (E) Histogram showing the distribution of dot product (DP) score.

By comparing 11- and 18-plex libraries, a substantial majority of precursors (246,327/344,212; 72%) and peptides (200,354/254,978; 79%) were found to be shared between the two libraries, indicating a significant level of similarity in precursor identification (Figure 3D). To further assess the quality of precursor identification, a similarity score (i.e., DP) was calculated for cases with multiple PSMs per precursor. The analysis revealed that the vast majority of precursor (99%) showed a DP > 0.80 (Figure 3E), indicating a high degree of similarity between the identified precursors for a given peptide. These findings suggest the potential interchangeability of the 11- and 18-plex methods in precursor identification.

While JUMPlib performs well with high-resolution MS2 data, it may not yet be the optimal solution for ion trap (SPS-MS3) data. SPS-MS3 typically relies on low-resolution MS2 scans generated by collision-induced dissociation (CID) fragmentation, whereas our TMT spectral library contains ions generated by higher-energy collisional dissociation (HCD) fragmentation. These two fragmentation methods can produce significantly different fragmentation patterns. Moving forward, we plan to enhance our TMT libraries and the JUMPlib tool to better support SPS-MS3 data.

3.3. JUMPlib Improves the speed and sensitivity of peptide identification

The integration of JUMPlib with fragment indexing has substantially improved peptide identification speed. When analyzing a TMT dataset consisting of 30,735 MS/MS scans, the use of fragment indexing reduced the total search time by approximately 66% (from around 9 minutes to about 3 minutes) compared to the dot product scoring strategy. Furthermore, it is 50% faster than the Comet tool when using the human protein database (168,305 entries). This enhancement is especially significant in large-scale proteomics studies dealing with extensive data volumes.

We next compare the performance of the JUMPlib program with other tools: (1) SpectraST, a widely used spectral library-based search tool, (2) COSS, a recently developed spectral library searching tool, and (3) Comet, a sequence database search tool using one LC-MS/MS run. The JUMPlib with the fragment index strategy identified 11,768 PSMs at the PSM FDR < 1%, which is 4% and 25% more than COSS and SpectraST with the same library configuration, and 18% more than Comet (Figure 4A), with the comprehensive human proteome database. The JDscore from JUMPlib, score from COSS, fval from SpectraST, and Xcorr from Comet displayed distinct score distributions for both targets and decoys (Figure 4B-E). Shorter peptide sequences generally produce simpler spectra with fewer peaks, resulting in shorter vector lengths. When the length of the vector is smaller, this has the potential to make the dot product relatively larger compared to longer peptides. Therefore, the dot product tends to favor identified peptides with shorter lengths. JUMPlib with the fragment index normalizes the dot product score by the length of peptide, which slightly improves the sensitivity.

Figure 4. Performance evaluation of JUMPlib, COSS, Comet, and SpectraST using a TMT dataset.

(A) The numbers of accepted PSMs at 1% PSM FDR were compared between the four tools. The fragment indexing strategy in JUMPlib was used. (B-D) Distribution of scores for targets and decoys detected by each search tool.

We subsequently sought to evaluate the performance with a large TMT proteomics dataset. This dataset consists of 40 LC-MS/MS runs generated by the 18-plex TMT strategy from the brain cortical tissue of frontotemporal lobar degeneration (FTLD) patients, which were not utilized for the library generation. JUMPlib identified 454,300 PSMs at a 1% FDR (Figure 5A), representing 26% and 22% more PSMs compared to SpectraST (337,530) and Comet (355,827), respectively. The peptide numbers detected by JUMPlib, SpectraST, and Comet are 124,874, 96,496, 105,846, respectively. Notably, JUMPlib covered 91% (306,699/337,530) of the PSMs detected by SpectraST, and ~98% (347,704/355,827) of the PSMs detected by Comet (Figure 5B). For the PSMs missed by the spectral library approach, this was primarily due to the stringent search criterion used in JUMPlib (requiring 1 out of the top 3 ions to be identified in the library).

Figure 5. Application of JUMPlib to a large-scale TMT dataset.

(A) Numbers of accepted PSMs by the three tools. (B) Up-set plot showing the number of precursors identified by the three tools. Set Size is the total number of precursors in each group. Interaction Size represents the number of precursors in the intersections of the groups as indicated by the black dots below the bars.

4. Conclusion

In summary, we introduce JUMPlib, a computational tool tailored for TMT-based proteomics alongside a comprehensive spectral library. By leveraging experimental RT data in the spectral library, JUMPlib improves the sensitivity and specificity of peptide identification when compared to other established tools, such as SpectraST, COSS and Comet. Overall, JUMPlib serves as a spectral library search tool for TMT-based proteomics data. Moreover, we provide comprehensive 11-plex and 18-plex TMT libraries, making it a valuable resource for the proteomics research community.

Supplementary Material

Supplementary Figures

Figure S1. Detailed workflow of the JUMPlib pipeline.

Figure S2. Comparison of the workflows and scoring methods used in JUMPlib and SpectraST.

Figure S3. Diagram showing the filtering strategy used in the JUMPlib program.

Figure S4. Impact of methionine oxidation modification on retention time shift.