Proteoform-predictor: Increasing the Phylogenetic Reach of Top-Down Proteomics

Abstract

Proteoforms are distinct molecular forms of proteins that act as building blocks of organisms, with post-translational modifications (PTMs) being one of the key changes that generate these variations. Top-down proteomics (TDP) is the leading technology for identifying proteoforms, and searching mass spectrometry (MS) data against a proteoform database is a crucial step. To extend the reach of TDP to organisms with limited PTM annotations, we developed Proteoform-predictor, an open-source tool that integrates homology-based PTM site prediction into proteoform database creation. The new tool creates databases of proteoform candidates after registration of homologous sequences, transferring PTM sites from well-characterized species to those with less comprehensive proteomic data. Our tool features a user-friendly interface and intuitive workflow, making it accessible to a wide range of researchers. We demonstrate that Proteoform-predictor expands proteoform databases with tens of thousands proteoforms for three bacterial strains by comparing them to the reference proteome of Escherichia coli (E. coli) K12. Subsequent TDP analysis for Serratia marcescens (S. marcescens) demonstrates significant improvement in protein and proteoform identification, even for proteins with variant sequences. As TDP technology advances, Proteoform-predictor will become an important tool for expanding the applicability of proteoform discovery and PTM biology to more and diverse species across the phylogenetic tree of life.

Graphic abstract

INTRODUCTION

Post-translational modifications (PTMs) involve covalent modifications to protein molecules, introducing chemical groups such as acetylations, phosphorylations and methylations.^1,2 This process generates distinct forms of a protein, known as proteoforms,^3,4 that fulfill different biological functions within living organisms. To understand the biological roles of proteoforms and their PTMs, it is essential to accurately characterize them within biological contexts. Top-down proteomics (TDP) uses mass spectrometry (MS) to measure intact proteoforms extracted from biological samples, providing accurate PTM information on protein molecules. Important steps in the workflow of TDP are database creation and searching, where MS data are matched against a database of candidates to generate protein- and proteoform-level identifications.^5,6

While some organisms are well-studied and have high-quality proteoform databases, many non-model organisms have databases with poorly annotated PTM information. Generally, proteoform databases are available from the UniProt Knowledgebase (UniProtKB),^7,8 a resource with over 200 million protein entries from more than a million species. However, many of these protein entries are unreviewed in UniProtKB, indicating that they usually lack detailed PTM annotations (Table S1). This results in a limited number of proteoform candidates, eventually leading to fewer identified proteoforms in TDP studies.

Introducing potential PTM sites into a proteoform database is an effective method to expand the pool of proteoform candidates. The Basic Local Alignment Search Tool (BLAST)⁹ is a powerful bioinformatics tool that compares an input sequence against a list of sequences to identify regions of local similarity. This method has been used for protein functional studies, including PTM site prediction,¹⁰ Here, we present Proteoform-predictor, an open-source Python-based tool, that integrates this BLAST-based PTM site prediction into the workflow of TDP database generation. Our tool transfers PTM information from well-annotated proteins to those with incomplete PTM annotations, thereby increasing the total number of proteoform candidates in the database.

One challenge in database creation is that the total number of proteoform candidates should remain manageable for TDP database search strategies. During TDP database searches, the search algorithm generally matches each precursor ion mass in the MS¹ data to the masses of proteoforms in the database.¹¹ If the database contains an excessive number of mass candidates, this process would be unmanageable by the current TDP search tools. Some methods that introduce potential PTM sites to a protein are not suitable for TDP. For example, in bottom-up proteomics (BUP), a common method to detect PTMs on peptides during its database search involves matching an observed precursor mass to a list of masses for all possible modified peptides (e.g., assigning a mass shift of phosphorylation to all serine, threonine, and tyrosine residues).¹² This method is effective for BUP because the typical size of a tryptic peptide is less than 30 amino acids, leading to a limited number of potential PTM sites per peptide. However, this approach is not viable for TDP database creation because it creates an astronomical number of mass candidates for intact proteins. In contrast, BLAST-based PTM site prediction in BUP has been successfully used for cross-species mapping of PTM sites, allowing the identification of new PTM sites without excessively increasing the number of potential candidates in the database. Building on this idea, we extended this PTM site prediction approach to TDP and implemented it in Proteoform-predictor.^13–15 According to current UniProtKB statistics, there are about 57,000 reviewed protein entries, and the average number of PTM sites per reviewed entry is about 0.11; thus, the number of potential PTM sites to be considered by our tool is not excessive. Additionally, Proteoform-predictor allows users to monitor the number of predicted PTM sites in the updated database, providing users with the ability to control the size and search space of a proteoform database.

Traditionally, predicting PTM sites has been accomplished using stand-alone tools that analyze amino acid (AA) sequences and provide predictive outcomes. However, this approach presents a barrier to the application of PTM site prediction in the TDP community due to the lack of tools to bulk edit multiple PTM sites in proteoform databases. Current TDP search tools such as ProSightPD^16,17 and MetaMorpheus¹⁸ require the input of proteoform database in the UniProt-XML format.⁷ This format is essential because it integrates comprehensive proteoform information, including PTM types and sites, and AA variations, but it also results in a more complex data structure and requires more random-access memory for storage compared to regular FASTA files. Consequently, it is impractical for users to manually edit PTM information in large UniProt-XML files using standard text editors, especially when multiple PTM site candidates need to be added. Proteoform-predictor addresses these limitations by integrating BLAST-based PTM site prediction with the automated editing of UniProt-XML files, resulting in a ready-to-search database. This output database can be browsed and edited using ProSight Annotator¹⁹ and serves as a database for TDP search engines.

Finally, while hundreds of PTM types have been discovered, most PTM prediction tools generate PTM sites for around 20 common PTM types.^10,20,21 Admittedly, predicting a small list of PTM types shortens the computation time, including the training of the predictive models and the generation of PTM site candidates. However, many biologically meaningful but less abundant PTM types, such as histidine phosphorylation²² and α-N-terminal methylation,²³ are overlooked. In contrast, Proteoform-predictor does not pre-select PTM types based on their rate of occurrence, instead, it references all PTM types recorded in UniProtKB (https://www.uniprot.org/docs/ptmlist). Here, we show the utility of including diverse PTM types for cross-family and cross-genus identification of bacterial proteoforms.

EXPERIMENTAL SECTION

Proteoform-predictor is a stand-alone open-source software available on GitHub (https://github.com/Tao-su/Proteoform-predictor) written in Python 3. It was designed to mine overlooked proteoforms within TD-MS datasets by supplying database search tools with a protein database integrated with predicted PTM sites (Figure 1). Users of Proteoform-predictor need to prepare two protein databases of two different organisms in UniProt-XML format and designate them as reference and target databases. Proteoform-predictor will then map as many PTM sites as possible from the reference database to the target database based on sequence homology. The output of Proteoform-predictor is a new protein database containing predicted PTM site information, which can be served as an input for trending TD-MS analysis tools.

Figure 1.

The flowchart of TD-PTM-Predictor. A) TD-PTM-Predictor is a pipeline that accomplishes PTM prediction after cross-species sequence registration to construct proteoform databases for searching with data from top-down proteomics. B) The input databases can be downloaded in UniProt XML format obtained from the UniProt Knowledgebase. The output databases are compatible with the current TDP search tools for proteoform identification, such as TDPortal and others.

Installing Proteoform-predictor and Dependencies

To facilitate ease of use, we have made the scripts of Proteoform-predictor publicly available on GitHub. A comprehensive user guide is also provided, helping users to download required scripts, input files, and additional tools, and to install Proteoform-predictor dependencies using Conda. Since the input file must be in UniProt-XML format, our guideline also walks users through the process of downloading example input files from UniProtKB, introducing them to the proper input file format and the workflow of Proteoform-predictor. Additionally, the guide includes instructions for downloading, installing, and using the BLAST+ (version 2.9.1)²⁴ from the National Center for Biotechnology Information (NCBI), which is required to enable local BLAST searches. Users can install all required Python modules such as Biopython²⁵ and ElementTree from a .yml file provided in the guide via Conda on Windows. The Conda function can be installed following the regular installation instructions on the Conda official website. Once the inputs and dependencies are prepared, running the Proteoform-predictor needs only a single line of code in the Command Prompt on a Windows system. This design ensures that our tool is intuitive, user-friendly, and independent of any specific Python development.

Downloading UniProt-XML files

Proteoform databases in UniProt-XML format for four species, Escherichia coli (E. coli) K12, E. coli strain B, Salmonella typhimurium (S. typhimurium), and Serratia marcescens (S. marcescens), were downloaded from UniProtKB. The number of reviewed and unreviewed protein entries in each database is shown in Figure S1. All protein entries in the database of E. coli K12 have been reviewed, while most entries in the other three databases remain unreviewed, indicating a need for improved annotations. The total number of proteoforms was calculated using ProSight Annotator. The Protein sequence databases in FASTA format for the first three species were also downloaded from UniProtKB. The databases in UniProt-XML format served as input for Proteoform-predictor while the protein sequence databases in FASTA format were used for the preparation of the local BLAST.

Running a local BLAST

Proteoform-predictor utilizes the local BLAST to locate the region with sequence similarity. Prior to running a local BLAST, BLAST+ was used to create a BLAST database using the downloaded database in FASTA format. BLAST+ can be executed in Command Prompt on a Windows system and generates phr, pin, and psq files that in combined as one BLAST database.

Proteoform-predictor uses Biopython module to run a local BLAST. The BLAST setting was tuned for short-sequence search, in which a stringent substitution matrix, PAM30 log-odds, was chosen to score the sequence similarity.^26,27 To ensure the quality of the BLAST search results, the E-value of 1e⁻⁴ was set as a cutoff to filter out low-quality matches.²⁸ When multiple matches are found in each BLAST search, the match with the highest BLAST score will be selected as the most possible candidate for the downstream analysis. The other optional parameters for the local BLAST were set as default values, and the output of each BLAST hit, including its similarity score and matched sequence, was exported as a .txt file.

As noted above, Proteoform-predictor creates a library of subsequences based on the protein sequences and PTM sites in the reference database. Each protein sequence in the reference database is partitioned in silico into subsequences of the same length, which is user-defined, with a default length of 21 amino acids. Each subsequence contains a single PTM site positioned at the center unless the site is near the protein terminus. In the second step, Proteoform-predictor determines which PTM sites in the library are transferrable to the target database by running a local BLAST search with settings tuned for the short-sequence search. The BLAST+ tool scores the sequence similarity between each subsequence in the library and the protein sequences in the target database. If a subsequence’s score exceeds the E-value threshold defined by the BLAST+ tool, the PTM site is considered a confident prediction, and the location and the PTM type associated with that site are stored in an output file. In the final step, Proteoform-predictor inserts predicted PTM sites from the output file into the target database based on the following simple rule: If the predicted site is not recorded in the original target database, an XML element containing the site location and PTM type will be created as a new modified residue and added to the database. This informs TDP search tools to generate new proteoforms associated with the new PTM for the database search.

Top-Down Mass Spectrometry

Three biological replicates of E. coli K12 and S. marcescens were grown in LB media and nutrient broth and incubated at 37°C overnight, respectively. Cells were isolated from media by centrifuging and thoroughly washed with MS level water to remove excessive media. The isolated cells were lysed in a buffer containing 0.1% Triton-X100, 50 mM Tris-HCl pH 7.5, and protease inhibitor (Thermo Fisher Scientific) and sonicated through 10 cycles (Thermo Fisher Scientific) on ice. Cell debris were removed by centrifugation and protein concentration was quantified by a BCA assay (Thermo Fisher Scientific). Proteins were fractionated by molecular weight and analyzed using the passively eluting proteins from polyacrylamide gels as intact species for MS (PEPPI-MS) workflow.^29,30 In brief, 80 μg of protein was resolved on a precast NuPAGE 4–12% Bis−Tris gradient gels (Thermo Fisher Scientific) with NuPAGE MES running buffer. The samples were run to the bottom of the gel, and the region containing proteins <30 kDa was divided into four fractions based on molecular weight: <10 kDa, 10–20 kDa, 20–25 kDa, and 25–30 kDa. Proteins in each fraction were recovered by passive extraction using 300 μL of 0.1% (w/v) SDS in freshly prepared 100 mM ammonium bicarbonate. Proteins in the solution were precipitated using methanol, chloroform, and water at a volume ratio of 4:1:3 to remove SDS and other contaminants.

Following the precipitation, the protein pellet in each fraction was immediately dissolved in 20 μL buffer A (94.8% water, 5% acetonitrile, 0.2% formic acid). For each sample, two technical replicates were analyzed using reversed-phase liquid chromatography using an Ultimate 3000 LC system (Thermo Fisher Scientific). For each injection, 8 μL protein sample was loaded onto an in-house fabricated 350 μm O.D. × 150 μm I.D. trap column packed with 3-cm PLRP-S resin (5 μm particle, 1000 Å pore, Agilent Technologies). To separate proteoforms, an in-house fabricated 350 μm O.D. × 75 μm I.D. analytical column packed with 18-cm PLRP-S resin was used. Both trap and analytical columns were heated at 55°C during protein separation. The total LC run time was 105 min using a gradient of mobile phase A (94.8% water, 5% acetonitrile, 0.2% formic acid) and mobile phase B (4.8% water, 95% acetonitrile, 0.2% formic acid). The flow rate was set at 0.3 μL/min, and the gradient used to resolve proteins was: 5–25% B in 20 min, 25–50% B in 80 min, and 50–95% in 5 min. The outlet of the column was coupled to a PicoTip spray emitter (New Objective), packed with 0.2-cm PLRP-S resin.

On-line with separation, proteoforms were directed into an Orbitrap Eclipse (Thermo Fisher Scientific) through the electrospray emitter held at a potential of ~3 kV. For top-down experiments, the mass spectrometer was operated in data-dependent mode, with full MS¹ data collected from 750 to 2000 m/z at a resolution of 120,000 (at 200 m/z) using an automatic gain control (AGC) set at 4 × 10⁶ charges and a maximum injection time of 100 ms. For MS/MS, the top two most abundant precursor ion populations were acquired and fragmented using HCD with an isolation window of 5 m/z at a normalized collisional energy of 23. MS/MS scans were collected from 400 to 2000 m/z at a resolution of 60, 000 (at 200 m/z) and a maximum ion injection time of 800 ms. Dynamic exclusion duration was set to 60 s after a repeat count of 1.

TDP Data Analysis

Top-down proteomic .RAW files were processed using a publicly available TDPortal v4.0.0 (https://portal.nrtdp.northwestern.edu/) workflow based on the Galaxy Project³¹ that generated a report with results validated at 1% false discovery rate (FDR) assignment at the protein, isoform, and proteoform levels.³² In brief, the raw files were grouped using a 2-minute retention time tolerance and 0.1 m/z tolerance for deconvolution of precursors and fragment ions into “targets” using the Crawler algorithm. Both isotopically resolved and unresolved MS¹ and MS² spectra were searched with an absolute mass search with a precursor mass tolerance of 2.2 Da and a 10-ppm fragment tolerance. The raw files were also searched using a biomarker search with a 10-ppm precursor and a 10-ppm fragment tolerance. All the searches were performed using the databases downloaded from UniProtKB and the databases generated by Proteoform-predictor.

To evaluate the performance of Proteoform-predictor, the number of newly added PTM sites and proteoforms in the updated database were counted, and the composition of the predicted PTM sites was visualized in TDViewer. The MS² spectra of proteoform of interest was manually examined in TDValidator. Both tools are available at https://www.proteinaceous.net/.

RESULTS AND DISCUSSION

Workflow of Database Augmentation

Proteoform-predictor maps PTM sites from the reference to the target species based on the protein sequence similarity by completing three consecutive steps: 1) generation of short sequences, 2) local BLAST search, and 3) proteoform database updating, which are illustrated in the two flowcharts of Figure 1. Before using Proteoform-predictor, users must select two species, designating one as the reference and the other as the target, and download protein databases for both species from UniProtKB as UniProt-XML files. Files for both species contain protein sequences and known feature keys with known information for that organism, with the reference database typically containing more extensive PTM information than the target. For example, in our example use case, we demonstrate the process using the E. coli K12 as the reference and the E. coli strain B proteoform database as the target, as E. coli K12 has more annotated PTM sites.

Expansion of Proteoform Candidates Through Sequence Homology

We validated that the first set of results from the BLAST-based PTM prediction reflected the evolutionary distance between the reference and target species. For the validation, we selected E. coli K12 strain as the reference species and S. marcescens, Salmonella typhimurium (S. typhimurium), and E. coli strain B as three different target species, since they have distinct evolutionary distances from E. coli K12.^33,34 As shown in Figure 2A, Proteoform-predictor predicted 390, 457, and 532 PTM sites for the S. marcescens, S. typhimurium, and E. coli strain B, respectively. Moreover, 19%, 77.1%, and 99.2% of the PTM sites in the original databases of S. marcescens, S. typhimurium, and E. coli strain B are aligned with the Proteoform-predictor results. This alignment proportion inversely correlated with the evolutionary distance of the three target species to the reference species. This is expected because E. coli strain B and E. coli K12 are closely related, whereas S. typhimurium and S. marcescens are more distantly related to E. coli K12. These findings indicate that Proteoform-predictor introduces new PTM sites to the target species based on the sequence homology, reflecting their evolutionary relationships to the reference species.

Figure 2.

Evaluation of TD-PTM-Predictor results across three different species across bacterial families: S. marcenscens, S. typhimurium, and E. coli strain B. (A) Comparison of the number of PTM sites derived from the prediction and the original database. The common area represents PTM sites found in both sources. (B) Composition of predicted PTM types, with the major types highlighted. (C) Comparison of the total number of proteoform candidates between the original and database created using Proteoform-predictor. (D) Distribution of matched sequence tags from the BLAST search. Tags are categorized into three classes: exact matches and matches with either fewer than or more than three amino acid mutations.

Upon analyzing the types of all predicted PTM sites, we confirmed that our tool introduces biochemically relevant modifications to proteoform databases. Overall, about half of the predicted sites have modifications on lysine residues, with acetylation, succinylation, and pyridoxal phosphorylation being the most prevalent across all three target species (Figure 2B).^35–37 This is consistent with the widespread occurrence of lysine modifications in regulating a variety of biological activities such as gene expression and protein degradation.

These predicted PTM sites contribute to additional proteoform candidates, expanding the search scope while maintaining a manageable number of candidates for modern database search tools. In total, more than 4 × 10⁴, 1.6 × 10⁵, and 1.5 × 10⁵ new proteoform candidates were added to S. marcescens, S. typhimurium, and E. coli strain B proteoform databases, respectively (Figure 2C). To emphasize the tool's ability to enable TDP studies, we specifically quantified the number of new proteoform candidates with molecular weights under 30 kDa (given the cutoff of the protein fractionation method used here) and found that 4688, 5398, and 4576 new proteoforms have masses fall in this range.³⁸ These results suggest that the number of proteoform candidates generated by Proteoform-predictor is easily manageable for current TDP search tools working on bacterial systems.

PTM Site Prediction on Sequences with Multiple Amino Acid Variations

BLAST-based PTM site prediction does not require a 100% match in AA sequence to detect sequence similarity, allowing for predictions on sequences with AA variations to some extent. The number of exact and fuzzy matches varied across species, but it follows a clear trend: the greater the evolutionary distance between the reference and target species, the more AA mismatches appeared in the BLAST results. We found that over 80% of matched subsequences are exact matches in E. coli strain B, indicating its close relation to E. coli strain K12. In contrast, about half of the matched tags contained AA variations in S. typhimurium, and more than 60% of matches had AA variations in S. marcenscens (Figure 2D). These results can guide users to adjust the flexibility in matching during BLAST searches and select the appropriate reference species. For example, a larger E-value threshold may be necessary to obtain more matches when mapping two distantly related species.

However, this matching mechanism may incorrectly assign modifications to certain amino acids. For example, when a serine residue mutates to threonine, phosphorylation at that site remains possible; thus, it is reasonable to predict phosphorylation on the mutated residue. Conversely, if the serine residue mutates to an AA other than threonine, tyrosine, and histidine, phosphorylation is no longer feasible. To solve this issue, we provide a list containing the possible amino acids for common PTM types summarized from the PTM information on UniProtKB (Table S2). We then include a checkpoint in our tool to remove predicted PTM sites on the subsequences containing mutations that disenable the corresponding modification, ensuring accurate PTM assignment.

Improved Identifications in Database Searches

Utilizing the Proteoform-predictor updated proteoform databases in the TDP search tool led to enhanced identifications for both proteins and proteoforms. We generated TDP data for both E. coli K12 and S. marcescens through PEPPI-MS. The E. coli K12 proteomics data served as quality control for the TDP workflow, while the data for S. marcescens was used to compare the search results between the original and Proteoform-predictor-updated proteoform database. In short, more than 2000 proteins and about 4500 proteoforms under 30 kDa are recorded in the E. coli K12, original, and updated S. marcescens databases, respectively (Figure 3A). We identified 212 proteins and 529 proteoforms with a 1% global FDR using the E. coli K12 TDP data. This established a benchmark, leading us to anticipate that the number of identifications for S. marcescens would be lower, as its database contains fewer proteoforms candidates than the one of E. coli K12. By analyzing the TDP data of S. marcescens, we identified 161 and 167 proteins, and 327 and 381 proteoforms using the original and updated databases, respectively (Figure 3B). We believe the quality of the search result using the updated database is comparable to the original one since the two results have similar p- and q-value distributions (Figures S2 and S3). This data indicates that Proteoform-predictor leads to a 12% and a 16% increase in protein and proteoform identification, respectively, for the same S. marcescens TDP data by introducing predicted PTM site candidates into the database. Meanwhile, the protein identification is slightly enhanced due to the expansion of the pool of proteoform candidates.

Figure 3.

Assessment of the identifications of proteins and proteoforms with molecular weights under 30 kDa, using the database updated by TD-PTM-Predictor. (A) The number of protein and proteoform candidates in E. coli K12 and two S. marcenscens databases. (B) The number of identified proteins and proteoforms using the three databases. O stands for the original database; U stands for the database updated with Proteoform-predictor.

Enhanced Proteoforms Identifications in Bacterial Ribosomal Protein

Among the 54 newly identified proteoforms, more than half were derived from bacterial ribosomal protein (Table S3–S6). This suggests that Proteoform-predictor could be a valuable tool for the functional studies of bacterial ribosomes,³⁹ particularly in areas such as ribosomal maturation⁴⁰ and antibiotic resistance mechanisms.⁴¹ To demonstrate the benefits of using Proteoform-predictor, we showcase the MS spectra of two bacterial ribosomal proteoforms. Figure 4 shows that RpIL is a ribosomal protein that was identified in both E. coli K12 and S. marcescens. The latter possesses three AA variants: I24V, V46A, and A110S relative to the former. The RpIL proteoform, possessing N-terminal acetylation and K82 methylation, was identified in E. coli K12, indicating that its proteoform database records K82 methylation as a potential PTM site (Figure 4A). When searching the S. marcescens proteomics data against its original database, only the canonical form of RpIL was identified. However, when searching against the updated database, a new RpIL proteoform with K82 methylation was identified, alongside the canonical form (Figure 4B). This result indicates that the K82 methylation PTM site was successfully transferred from the E. coli K12 database to the S. marcescens database via Proteoform-predictor and further became a confident hit in the database search.

Figure 4.

MS spectra of rpIL proteoforms in E. coli K12 and S. marcescens. (A) Spectra of identified acetylated (gold) and methylated (green) rpIL in E. coli K12. (B) Spectrum of identified canonical (blue) and N-terminally methylated (red) rpIL in S. marcescens.

A broader range of proteoform identifications in TDP studies can also be achieved by using Proteoform-predictor because it includes less common PTM types. As illustrated in Figure 5, we identified an N-terminal methylated RpmG proteoform due to the inclusion of the predicted N-terminal methylation site in the S. marcescens database. Typically, N-terminal methylation is mediated by methyltransferases, which recognize specific consensus motifs at the N-termini of substrate proteins. Different methyltransferases have distinct motif requirements in terms of their amino acid composition and length. However, accounting for these diverse motifs increases the computational burden during the model training, thus, many PTM prediction tools do not include N-terminal methylation.⁴² In contrast, Proteoform-predictor focuses on PTM types recorded in the one user-provided database downloaded from UniProtKB, reducing the computational burden while enabling broader PTM site predictions. We are confident in the identification of N-terminal methylation on S. marcescens RpmG, as it contains the motif recognized by protein L11 methyltransferase.

Figure 5.

MS spectra of identified canonical (blue) and N-terminally methylated (red) rpmG in S. marcescens. Amino acid highlighted in green represents an identification of methylation.

CONCLUSION

The motivation to develop Proteoform-predictor stemmed from the gap in existing tools for integrating PTM prediction into TDP workflows for organisms with annotated genomes, but poorly studied proteomes. This limitation has hindered the ability of researchers to explore the proteoform biology of such species. Proteoform-predictor addresses this gap by providing a workflow that integrates BLAST-based PTM prediction that enables the TDP pipeline. It offers a user-friendly interface that greatly simplifies the complex process of PTM site prediction and database updating. Additionally, by making the tool freely available, we aim to broaden the range of researchers to enhance their proteomics studies with high-confidence PTM site annotations.

Demonstrating the effectiveness of Proteoform-predictor ~300 new PTM sites were predicted for each of three species, resulting in the introduction of tens of thousands of proteoforms in the Proteoform-predictor-updated proteoform databases. This led to a 12% increase in identified proteins and a 16% increase in identified proteoforms <30 kDa. This led to first-observation of proteoforms with PTMs like N-methylalanine on RpmG in S. marcescens. These findings not only expand our understanding of bacterial ribosomal protein modifications but also highlight the broader application of Proteoform-predictor in proteomics, where accurate and comprehensive proteoform databases are essential for high-confidence identifications.

Proteoform-predictor will help TDP technology grow in its utility, particularly in achieving better coverage for proteins and proteoforms with larger masses exceeding 30 kDa and those from archaeal, eukaryotic, and especially mammalian organisms. Ongoing use and refinement will allow Proteoform-predictor to transfer more PTM information from well-studied organisms to non-model organisms, thereby enhancing our understanding of proteomic landscapes across different species with convergent proteoform biology. Proteoform-predictor is positioned to be an essential tool for future TDP research across species, facilitating deeper insights into the proteoform program of life that has evolved on our planet.

Supplementary Material

Figure S1, total number of reviewed and unreviewed protein entries for the four species analyzed in this study; Figure S2, distributions of p-values derived from the search results of S. marcescens TDP data using the original and database updated with Proteoform-predictor; Figure S3, distributions of q-values derived from the search results of S. marcescens TDP data using the original and database updated with Proteoform-predictor; Table S1, total number of protein entries in UniProtKB released in July 2024 (.pdf).

Table S2, the summary of the possible amino acids for common PTM types; Table S3. identified proteins using the original database of S. marcescens; Table S4. Identified proteoforms using the original database of S. marcescens; Table S5. identified proteins using the database of S. marcescens updated by Proteoform-predictor; Table S6. identified proteoforms using the database of S. marcescens updated by Proteoform-predictor (.xlsx).