Proteogenomic Workflows for Refining Gene Models: From Mass Spectrometry Data to Updated Annotations

By Dr. Zubair Khalid, DVM, MS, PhD · 2026-08-21

Proteogenomic workflows integrate mass spectrometry-derived peptide evidence with genomic sequence data to validate, correct, and discover protein-coding gene models. For researchers working with organisms that have incomplete or imperfect genome annotations, this approach provides direct physical evidence that a genomic region is translated into protein. The workflow moves from constructing a searchable protein database, through peptide identification and genomic mapping, to evidence-based annotation revision. This article presents the complete workflow with practical decisions at each stage, using published examples from human, rat, wheat, and Atlantic salmon studies to illustrate what works and what fails.

Key Takeaways

• Proteogenomic workflows integrate mass spectrometry derived peptide evidence with genomic sequence data to validate, correct, and discover protein-coding gene models.
• Database construction is the first decision point, where expanding the search database with transcriptomic or six-frame translations increases sensitivity but raises false discovery risk.
• Stringent quality control, including target-decoy searching and conservative filtering, is essential to prevent false peptide identifications from propagating into incorrect annotation changes.
• Genomic mapping tools like tBLASTn connect peptides to loci, but they may miss true mappings for peptides with mid-sequence frame shifts.
• The wheat 2024 project validated 31.4 percent of High Confidence gene models with 83,015 unique peptides, while 2,934 orphan peptides were proposed as novel gene candidates.

The Problem Space: Why Gene Models Need Proteomic Evidence

Genome annotation remains an incomplete process for most organisms. Computational gene predictors, transcript mapping, and homology searches produce candidate gene models, but these predictions frequently contain errors. Exons may be missed, intron boundaries misassigned, start sites incorrectly placed, and entire genes overlooked. The consequences of annotation errors propagate through every downstream analysis, from differential expression studies to evolutionary comparisons and functional genomics.

Mass spectrometry provides a distinct evidence type that computational prediction cannot offer. When a peptide sequence is identified from tandem mass spectra, that peptide must have been translated from a messenger RNA, which in turn must have been transcribed from genomic DNA. This chain of physical evidence means that a confidently identified peptide confirms translation of a specific genomic locus. The GENCODE consortium, which maintains human gene annotation, has demonstrated that a high-stringency proteogenomics workflow can support the addition of novel protein-coding genes to the reference annotation, while also showing that many peptide identifications in pseudogenes cannot be annotated without orthogonal supporting evidence [<a href="#ref-1">1</a>].

The scale of the problem varies by organism. For the wheat genome, a 2024 community resource project retrieved publicly available proteomics datasets and mapped 861,759 unique peptides against the International Wheat Genome Sequencing Consortium RefSeq v2.1 reference genome. Of the resulting 92,719 hits, 83,015 unique peptides aligned along 33,612 High Confidence gene models, validating 31.4 percent of all wheat High Confidence gene models. A further 6,685 unique peptides mapped against 3,702 Low Confidence gene models, providing evidence that these models should be considered for promotion to High Confidence status. The remaining 2,934 orphan peptides could not be assigned to existing models and were proposed as candidates for novel gene discovery [<a href="#ref-2">2</a>].

For non-model species, the annotation gap is wider. A 2023 study using Atlantic salmon constructed proteogenomic databases from publicly available transcriptomic data and in-house generated RNA-Seq and LC-MS/MS data. The analysis identified approximately 80,000 peptides, providing direct evidence of translation for over 40,000 RefSeq structures. Critically, 183 co-located peptide groups supported a single transcript each, and in every case the evidence either corrected a previous annotation, supported an Ensembl annotation not present in RefSeq, or identified a novel previously unannotated gene [<a href="#ref-3">3</a>].

These examples establish the core value proposition. Proteogenomic evidence does also confirm what computational methods already predicted. It corrects errors, resolves discrepancies between annotation sources, and discovers genes that no computational method found.

Core Principles of Proteogenomic Analysis

The Evidence Chain from Spectra to Gene Models

The proteogenomic evidence chain has four links. First, mass spectrometry produces spectra from digested proteins. Second, search algorithms match those spectra to peptide sequences from a protein database. Third, the identified peptide sequences are mapped back to the genome. Fourth, the genomic mappings are interpreted in the context of existing gene models.

Each link introduces potential errors, and the workflow must control for errors at every stage. A false peptide identification propagates through the entire chain and can lead to incorrect annotation changes. The GENCODE proteogenomics workflow emphasized that a conservative approach using stringent filtering is required to generate valid identifications, based on reprocessing of three large-scale publicly available human datasets [<a href="#ref-1">1</a>].

Database Construction Is the First Decision Point

The protein database used for peptide searching determines what can be identified. A standard reference proteome database contains the protein sequences predicted from the current genome annotation. This database can only identify peptides that match those predictions. Novel exons, alternative splice variants, translated pseudogenes, and entirely novel genes will be invisible to a search against the reference proteome alone.

Proteogenomic workflows expand the database to include sequences that represent possible but unannotated translation products. The Atlantic salmon study constructed proteogenomic databases from publicly available transcriptomic data combined with in-house generated RNA-Seq data, using the six-frame translation of transcripts to capture potential protein sequences that might not appear in the reference annotation [<a href="#ref-3">3</a>]. The rat genome re-annotation study used an integrated transcriptomic-proteomic pipeline that incorporated RNA-Seq data to build search databases representing the actual transcriptional landscape of the tissue being analyzed [<a href="#ref-4">4</a>].

The database construction decision involves a tradeoff. A larger database increases sensitivity by including more possible peptide sequences, but it also increases the search space and raises the risk of false positive identifications. The multiple testing burden grows with database size, and stringent false discovery rate control becomes more critical.

Peptide Identification Requires Stringent Quality Control

Peptide identification from mass spectra relies on search algorithms that score the match between an observed spectrum and a theoretical spectrum derived from a candidate peptide sequence. The output of these searches requires statistical validation. Target-decoy searching, where the database is supplemented with reversed or shuffled sequences, provides a standard approach for estimating the false discovery rate.

The GENCODE workflow demonstrated that stringent filtering is essential. The authors reprocessed three large-scale publicly available human datasets and showed that a conservative approach was required to generate valid identifications. They also noted that many peptide identifications in pseudogenes could not be annotated due to the absence of orthogonal supporting evidence [<a href="#ref-1">1</a>]. This finding highlights a key principle. Peptide evidence alone may not be sufficient to support an annotation change. The genomic context must be examined, and additional evidence types may be required.

Genomic Mapping Connects Peptides to Loci

Once peptides are confidently identified, their sequences must be mapped to the genome. This step determines which genomic locus produced each peptide. The mapping must account for the possibility that a peptide spans an exon junction, in which case it will not align as a contiguous sequence to the genome.

The wheat proteogenomics project used the Basic Local Alignment Search Tool with the tBLASTn algorithm to map identified unique peptides against the wheat reference genome [<a href="#ref-2">2</a>]. The choice of mapping tool affects sensitivity and specificity. The wheat study demonstrated a specific limitation of tBLASTn. It could not map peptides exhibiting mid-sequence frame shifts, meaning that some true mappings were missed by this approach [<a href="#ref-2">2</a>].

Annotation Interpretation Requires Genomic Context

The final step interprets peptide mappings in the context of existing gene models. A peptide that maps within an annotated exon confirms that exon. A peptide that maps to an intronic region suggests the intron may actually be translated. A peptide that maps to an intergenic region suggests a novel gene. A peptide that spans an exon-exon junction confirms the splice site.

The rat genome study used four proteogenomic pipelines to analyze publicly available RNA-Seq and mass spectrometry data from rat brain microglia. The analysis discovered 363 novel peptides representing novel proteoforms for 249 gene loci. These novel peptides aided in the discovery of novel exons, translation of annotated untranslated regions, pseudogenes, and splice variants for various loci, many with known disease associations including schizophrenia and amyotrophic lateral sclerosis [<a href="#ref-4">4</a>].

At a Glance: Workflow Stages and Key Decisions

Building the Search Database

Reference Proteome as the Baseline

The reference proteome represents the current best prediction of all protein-coding sequences in the genome. For well-annotated organisms, this database contains the canonical protein sequences from the primary annotation source. The NCBI maintains reference sequence databases that provide this baseline for many organisms, along with search systems and analysis services for accessing and comparing these resources [<a href="#ref-5">5</a>].

Searching against the reference proteome alone answers a limited question. Which predicted proteins are actually translated in the sample being analyzed? This question has value for validation studies, but it cannot discover anything new. Every peptide identified will match an existing prediction.

Expanding with Transcriptomic Evidence

The integrated transcriptomic-proteomic approach used in the rat genome study demonstrates the value of incorporating RNA-Seq data into the search database. The EuGenoSuite pipeline, developed by the authors, used multiple algorithmic proteomic search tools within an integrated transcriptomic-proteomic pipeline to facilitate automated proteogenomic analysis [<a href="#ref-4">4</a>].

The logic is straightforward. RNA-Seq data reveals which genomic regions are transcribed in the tissue of interest. Some transcripts may represent novel exons, alternative splice forms, or previously unannotated genes. Translating these transcripts in all six reading frames produces candidate protein sequences that can be added to the search database. If mass spectrometry evidence matches peptides from these novel translations, the transcript is supported by protein-level evidence.

The Atlantic salmon study followed this approach by constructing proteogenomic databases from publicly available transcriptomic data combined with in-house generated RNA-Seq data [<a href="#ref-3">3</a>]. This strategy ensures that the search database reflects the actual transcriptional landscape of the samples being analyzed, instead of only the predicted gene models.

Six-Frame Translation for Complete Coverage

The most exhaustive database construction approach translates the entire genome in all six reading frames. This approach makes no assumptions about gene structure and can capture peptides from any translated region. The cost is a very large database that increases search time and multiple testing burden.

Six-frame translation databases are most useful for organisms with poor annotations or for detecting translation in regions that produce no annotated transcripts. The tradeoff is substantial. The false discovery rate must be carefully controlled when searching against a database that contains many sequences with no biological basis.

Practical Database Construction Decisions

The database size decision should be guided by the research question. For validating existing annotations, the reference proteome plus known isoforms may suffice. For discovering novel genes and correcting annotations, transcriptomic expansion is necessary. For complete discovery in poorly annotated organisms, six-frame translation may be required.

The Atlantic salmon study provides a practical example of database construction for a non-model species. The authors combined publicly available transcriptomic data with in-house generated RNA-Seq and LC-MS/MS data to construct their proteogenomic databases [<a href="#ref-3">3</a>]. This hybrid approach leverages existing community resources while adding tissue-specific transcriptomic evidence from the actual samples being analyzed.

Peptide Identification and Quality Control

Search Algorithms and Their Tradeoffs

Multiple search algorithms are available for peptide identification, each with different strengths. The rat genome study used EuGenoSuite, an open source multiple algorithmic proteomic search tool, within their integrated pipeline [<a href="#ref-4">4</a>]. The use of multiple algorithms provides complementary evidence, as different algorithms may identify different subsets of true peptides.

The choice of search algorithm affects sensitivity and specificity. Some algorithms are optimized for high-resolution mass spectrometry data, while others handle lower-resolution data. Some excel at identifying post-translational modifications, while others assume unmodified peptides. The search database size and complexity also affect algorithm performance.

False Discovery Rate Control

The false discovery rate (FDR) is the expected proportion of false positives among the identified peptides. Standard practice uses target-decoy searching, where the database contains both forward protein sequences and reversed or shuffled decoy sequences. Peptide matches to decoy sequences estimate the false positive rate.

The GENCODE proteogenomics workflow emphasized that stringent filtering is required to generate valid identifications [<a href="#ref-1">1</a>]. The authors reprocessed three large-scale publicly available human datasets and demonstrated that a conservative approach was necessary. This finding has direct implications for workflow design. The FDR threshold must be set appropriately for the intended use of the peptide identifications.

For annotation refinement, the consequences of a false positive are severe. A false peptide identification could lead to the addition of a nonexistent exon or the correction of a gene model that was actually correct. The GENCODE workflow therefore recommended a high-stringency approach for proteogenomic evidence intended to support annotation changes [<a href="#ref-1">1</a>].

Filtering Beyond the FDR

FDR control addresses the global false positive rate, but additional filtering is often needed. Peptide length, number of matched fragment ions, mass accuracy, and other quality metrics can distinguish high-confidence identifications from marginal ones.

The wheat proteogenomics project provides an example of scale. The study retrieved the largest recent wheat proteomics datasets publicly available and applied tBLASTn to map 861,759 identified unique peptides against the wheat reference genome [<a href="#ref-2">2</a>]. The sheer number of peptides required automated filtering and mapping pipelines to handle the data volume.

The Problem of One-Hit Wonders

A peptide identified from a single spectrum, sometimes called a one-hit wonder, carries less evidence than a peptide identified from multiple spectra. For annotation changes, the GENCODE workflow recommended caution with single-peptide evidence [<a href="#ref-1">1</a>]. The absence of orthogonal supporting evidence was cited as a reason why many peptide identifications in pseudogenes could not be annotated.

The practical implication is that novel annotation claims should ideally be supported by multiple peptide identifications from the same genomic region, or by peptide evidence combined with transcriptomic evidence. The rat genome study demonstrated this principle by using integrated transcriptomic-proteomic analysis, where RNA-Seq data provided supporting evidence for the novel proteoforms discovered through mass spectrometry [<a href="#ref-4">4</a>].

Mapping Peptides to the Genome

Alignment Tools and Their Limitations

The mapping step determines the genomic coordinates of each identified peptide. The wheat proteogenomics project used tBLASTn, a variant of the Basic Local Alignment Search Tool that translates the nucleotide database in all six reading frames and searches it with protein queries [<a href="#ref-2">2</a>].

The choice of mapping tool has practical consequences. The wheat study demonstrated that tBLASTn could not map peptides exhibiting mid-sequence frame shifts [<a href="#ref-2">2</a>]. This limitation means that some true peptide mappings were missed, potentially leading to underestimation of the number of genes supported by peptide evidence.

Alternative mapping approaches include direct nucleotide alignment of the peptide-coding sequence, splice-aware alignment tools that can handle exon junctions, and specialized proteogenomic mapping tools. Each approach has different sensitivity and specificity characteristics.

Handling Exon Junctions

Peptides that span exon junctions present a special mapping challenge. The peptide sequence is contiguous in protein space but discontinuous in genomic space. A standard alignment tool will fail to map such peptides because the genomic sequence contains an intron between the two exon segments.

Splice-aware alignment tools can handle this problem by allowing gaps in the genomic alignment that correspond to introns. The choice of whether to use splice-aware mapping depends on the expected frequency of junction-spanning peptides in the dataset. For organisms with many multi-exon genes, junction-spanning peptides provide valuable evidence for splice site confirmation.

Resolving Multi-Mapping Peptides

Some peptides map to multiple genomic locations. This situation arises when the peptide sequence is encoded by a repeated genomic region or by a gene family with conserved protein sequences. Multi-mapping peptides cannot provide unambiguous evidence for a specific gene model.

The standard approach assigns multi-mapping peptides to all possible locations or uses parsimony principles to assign them to the minimum number of loci. For annotation refinement, multi-mapping peptides are typically excluded from novel gene claims because they cannot localize the evidence to a specific locus.

The Wheat Example: Mapping at Scale

The wheat proteogenomics project provides a concrete example of mapping at scale. The study mapped 861,759 identified unique peptides against the wheat reference genome, producing 92,719 hits. Of these, 83,015 unique peptides aligned along 33,612 High Confidence genes, validating 31.4 percent of all wheat High Confidence gene models [<a href="#ref-2">2</a>].

The scale of this analysis required automated pipelines. The study supplied their Galaxy workflow and Python code, as well as Browser Extensible Data (BED) files, as a community resource [<a href="#ref-2">2</a>]. This sharing of analysis code and results enables other researchers to apply the same workflow to additional datasets.

Interpreting Peptide Evidence in Genomic Context

Categories of Evidence

Peptide mappings fall into distinct categories when interpreted against existing gene models. Each category has different implications for annotation refinement.

Peptides that map within annotated exons confirm the existing gene model. These peptides provide protein-level evidence that the predicted exon is actually translated. The wheat study validated 31.4 percent of High Confidence gene models through this type of evidence [<a href="#ref-2">2</a>].

Peptides that map to intronic regions suggest that the intron may actually be translated. This evidence can support the discovery of novel exons. The rat study discovered novel exons through this type of evidence [<a href="#ref-4">4</a>].

Peptides that map to intergenic regions suggest novel genes. The wheat study identified 2,934 orphan peptides that could not be mapped to existing gene models and proposed them as candidates for novel gene discovery [<a href="#ref-2">2</a>].

Peptides that map to pseudogenes provide evidence of pseudogene translation. The GENCODE workflow found that many peptide identifications in pseudogenes could not be annotated due to the absence of orthogonal supporting evidence [<a href="#ref-1">1</a>]. This finding suggests that pseudogene translation evidence requires additional support before annotation changes are made.

Correcting Existing Annotations

The Atlantic salmon study provided a clear example of annotation correction. The analysis identified 183 co-located peptide groups that supported a single transcript each. In every case, the evidence either corrected a previous annotation, supported an Ensembl annotation not present in RefSeq, or identified a novel previously unannotated gene [<a href="#ref-3">3</a>].

This finding demonstrates that proteogenomic evidence can resolve discrepancies between annotation sources. When RefSeq and Ensembl disagree about a gene model, peptide evidence can determine which version is correct. When both sources miss a gene, peptide evidence can identify the novel locus.

Supporting Low Confidence Models

The wheat study demonstrated how peptide evidence can support the promotion of Low Confidence gene models to High Confidence status. The study found 6,685 unique peptides mapped against 3,702 Low Confidence gene models and argued that these gene models should be considered for High Confidence status [<a href="#ref-2">2</a>].

This application of proteogenomic evidence has practical value for genome projects. Low Confidence models are typically considered unreliable for downstream analysis. Peptide evidence provides a basis for promoting specific models, improving the overall quality of the annotation.

Discovering Novel Genes

The discovery of novel genes requires peptide evidence that maps to genomic regions with no existing gene model. The wheat study exemplified this approach on chromosome 4D, using orphan peptides for novel gene discovery [<a href="#ref-2">2</a>].

Novel gene discovery through proteogenomics is most productive in organisms with incomplete annotations. The rat study discovered novel proteoforms for 249 gene loci, including novel exons, translation of annotated untranslated regions, pseudogenes, and splice variants [<a href="#ref-4">4</a>]. Many of these novel proteoforms were for genes with known disease associations, including neurological disorders, cardiovascular diseases, and breast cancer [<a href="#ref-4">4</a>].

Practical Workflow Implementation

Step 1: Define the Annotation Question

The workflow should begin with a clear statement of the annotation question. Are you validating existing gene models? Correcting specific errors? Discovering novel genes? The answer determines database construction strategy, filtering stringency, and interpretation criteria.

For validation studies, the reference proteome database may be sufficient. For correction and discovery, transcriptomic expansion is needed. The Atlantic salmon study combined publicly available transcriptomic data with in-house generated RNA-Seq data to construct databases that could support both validation and discovery [<a href="#ref-3">3</a>].

Step 2: Assemble Input Data

The required inputs are mass spectrometry data, a reference genome, and optionally transcriptomic data. The mass spectrometry data should come from samples relevant to the annotation question. The reference genome should be the version being annotated. The transcriptomic data should represent the same tissues or conditions as the mass spectrometry data.

Public data resources can supplement in-house data. The NCBI provides access to sequence databases, search systems, and analysis services that can support data assembly [<a href="#ref-5">5</a>]. The wheat study retrieved the largest recent wheat proteomics datasets publicly available, demonstrating the value of community data resources [<a href="#ref-2">2</a>].

Step 3: Construct the Search Database

Database construction follows the decisions made in Step 1. The reference proteome provides the baseline. Transcriptomic data can be assembled and translated to add novel sequences. Six-frame translation of the genome provides complete coverage for poorly annotated organisms.

The database construction should be documented carefully. The exact sequences included, their sources, and the translation approach all affect downstream results. Reproducibility requires that this documentation be complete.

Step 4: Search and Filter

Peptide identification uses the chosen search algorithm against the constructed database. The search output requires FDR control and additional quality filtering. The GENCODE workflow demonstrated that stringent filtering is required for valid identifications [<a href="#ref-1">1</a>].

The filtering stringency should match the intended use of the results. For annotation changes, high stringency is essential. For exploratory analysis, slightly relaxed filtering may be acceptable, with the understanding that novel claims will require additional validation.

Step 5: Map Peptides to the Genome

The identified peptides are mapped to the reference genome using the chosen alignment tool. The wheat study used tBLASTn and documented its limitation with mid-sequence frame shifts [<a href="#ref-2">2</a>]. Alternative tools may be needed for specific mapping challenges.

The mapping output should include genomic coordinates, strand information, and alignment quality metrics. Multi-mapping peptides should be flagged for careful interpretation.

Step 6: Interpret in Genomic Context

The final step interprets peptide mappings against existing gene models. This interpretation requires a genome browser or annotation tool that can display peptide mappings alongside gene models. The wheat study provided their results via Apollo Jbrowse, enabling community access to the proteogenomic evidence [<a href="#ref-2">2</a>].

The interpretation should categorize each peptide mapping as confirming an existing model, correcting a model, supporting a Low Confidence model, or identifying a novel locus. Each category has different implications for annotation updates.

Step 7: Document and Share

The workflow should be documented for reproducibility. The wheat study supplied their Galaxy workflow and Python code, as well as BED files, as a community resource [<a href="#ref-2">2</a>]. This sharing enables other researchers to apply the same workflow to additional datasets.

The Galaxy Training Network provides accessible workflow training and analysis tutorials that can support the implementation of reproducible proteogenomic pipelines [<a href="#ref-6">6</a>]. The nf-core documentation provides community pipeline standards for reproducible workflow configuration [<a href="#ref-7">7</a>].

Records and Measurements

What to Record

The proteogenomic workflow generates multiple types of records that should be preserved. The search database construction should be documented with version information for all input sequences. The search parameters should be recorded, including the algorithm version, precursor mass tolerance, fragment mass tolerance, and enzyme specificity. The FDR threshold and filtering criteria should be specified.

The peptide identification results should be stored in a standard format that includes peptide sequences, scores, and quality metrics. The genomic mappings should include coordinates, strand, and alignment scores. The interpretation results should document which category each peptide mapping falls into and what annotation action, if any, is recommended.

Quality Metrics to Track

Several quality metrics provide ongoing assessment of workflow performance. The number of identified peptides at a given FDR threshold indicates the sensitivity of the analysis. The proportion of peptides mapping to annotated exons indicates the completeness of the annotation. The proportion of novel peptides indicates the discovery potential of the dataset.

The Atlantic salmon study reported approximately 80,000 identified peptides providing evidence for over 40,000 RefSeq structures [<a href="#ref-3">3</a>]. The wheat study reported 861,759 identified unique peptides with 83,015 mapping to High Confidence genes [<a href="#ref-2">2</a>]. These numbers provide benchmarks for comparing workflow performance across datasets.

Reproducibility Records

Reproducibility requires complete records of the computational environment. Software versions, database versions, and parameter settings all affect results. Container-based approaches can capture the computational environment for future replication.

From command-line data manipulation to collaborative version control, the Carpentries' open curriculum equips researchers with essential reproducible science skills.

Common Failure Patterns

Overly Permissive Filtering

The most common failure in proteogenomic workflows is insufficiently stringent filtering. The GENCODE workflow demonstrated that a conservative approach using stringent filtering is required to generate valid identifications [<a href="#ref-1">1</a>]. Overly permissive filtering produces false positive peptide identifications that can lead to incorrect annotation changes.

The failure mode is insidious because false positives are not randomly distributed. They tend to cluster in specific sequence contexts, potentially creating false evidence for novel exons or genes. The consequences for annotation quality are severe.

Database Contamination

The search database can contain sequences that do not represent real biological translation products. Six-frame translations of non-coding regions, assembly errors, and contaminating sequences all contribute to database contamination. These sequences increase the search space and can produce false identifications.

The Atlantic salmon study addressed this challenge by constructing databases from transcriptomic data, which filters out many non-transcribed genomic regions [<a href="#ref-3">3</a>]. Transcriptomic evidence provides a biological filter that reduces database contamination.

Mapping Tool Limitations

The mapping step can fail to identify true peptide mappings. The wheat study demonstrated that tBLASTn could not map peptides exhibiting mid-sequence frame shifts [<a href="#ref-2">2</a>]. This limitation means that some true mappings are missed, leading to underestimation of annotation support.

The practical response is to use multiple mapping approaches or to validate mapping failures manually. The choice of mapping tool should be informed by the expected characteristics of the peptides in the dataset.

Ignoring Multi-Mapping Peptides

Multi-mapping peptides present an interpretation challenge. Assigning them to a single locus can create false evidence for that locus. Excluding them entirely loses potentially useful evidence. The appropriate approach depends on the annotation question.

For novel gene discovery, multi-mapping peptides should be excluded because they cannot localize evidence to a specific locus. For validation of existing models, multi-mapping peptides can provide supporting evidence if the locus is already annotated.

Confirmation Bias in Interpretation

The interpretation step is vulnerable to confirmation bias. Researchers may be more likely to accept peptide evidence that supports existing hypotheses and reject evidence that contradicts them. The Atlantic salmon study found that peptide evidence corrected previous annotations, supported Ensembl annotations not present in RefSeq, and identified novel genes [<a href="#ref-3">3</a>]. Each type of outcome requires equal consideration.

The rat study discovered novel proteoforms for genes with known disease associations [<a href="#ref-4">4</a>]. These discoveries required accepting evidence that contradicted existing annotations. The interpretation process should apply consistent criteria regardless of whether the evidence confirms or contradicts existing models.

Limitations and Interpretation Boundaries

Peptide Evidence Is Necessary but Not Sufficient

Peptide evidence confirms that a genomic region is translated. It does not determine the full structure of the translated product. A peptide may map to a novel exon without revealing the complete transcript structure. The GENCODE workflow noted that many peptide identifications in pseudogenes could not be annotated due to the absence of orthogonal supporting evidence [<a href="#ref-1">1</a>].

The practical implication is that peptide evidence should be combined with transcriptomic evidence for annotation changes. The rat study used integrated transcriptomic-proteomic analysis, where RNA-Seq data provided supporting evidence for novel proteoforms [<a href="#ref-4">4</a>].

Sample Coverage Limits Discovery

The peptides identified in a proteogenomic experiment depend on the proteins expressed in the analyzed samples. Proteins that are not expressed, or expressed below detection limits, will not produce peptide evidence. The wheat study noted that their workflow could be applied to additional datasets from biotically and abiotically stressed samples to expand the resource with proteins and peptides from different conditions [<a href="#ref-2">2</a>].

The practical implication is that a single proteogenomic experiment provides partial evidence. Multiple experiments across tissues, conditions, and developmental stages are needed for comprehensive annotation support.

Technical Limitations of Mass Spectrometry

Mass spectrometry has inherent limitations that affect proteogenomic analysis. Some peptides do not ionize well and are never detected. Some protein sequences produce no detectable peptides under standard digestion conditions. These limitations create blind spots in peptide evidence.

The Atlantic salmon study analyzed multiple tissue types and detected distinct expression profiles for each [<a href="#ref-3">3</a>]. This multi-tissue approach expands the coverage of the proteome and reduces the impact of tissue-specific expression on annotation support.

Annotation Source Discrepancies

Different annotation sources often disagree about gene models. The Atlantic salmon study found that peptide evidence sometimes supported Ensembl annotations not present in RefSeq [<a href="#ref-3">3</a>]. These discrepancies require careful interpretation. The peptide evidence may support one annotation source over another, but the absence of peptide evidence for a model does not prove the model is wrong.

The interpretation should acknowledge the limitations of the evidence. Peptide evidence can confirm translation, but it cannot prove that a gene model is complete or that alternative isoforms do not exist.

Safety and Regulatory Context

Data Integrity and Reproducibility

Proteogenomic analyses that support annotation changes have downstream consequences for the research community. Incorrect annotations propagate through databases and affect future studies. The GENCODE workflow emphasized the need for stringent filtering to generate valid identifications [<a href="#ref-1">1</a>].

The practical implication is that annotation changes should be supported by reproducible workflows with complete documentation. The Galaxy Training Network provides accessible workflow training and analysis tutorials that support reproducibility [<a href="#ref-6">6</a>]. The nf-core documentation provides community pipeline standards for reproducible workflow configuration [<a href="#ref-7">7</a>].

Community Standards for Annotation

The annotation community has established standards for evidence quality. The GENCODE consortium strives to provide complete annotation of the human genome, augmenting computational and experimental evidence with manual annotation [<a href="#ref-1">1</a>]. Proteogenomic evidence is one input to this process, but it is not the only input.

The practical implication is that proteogenomic results should be submitted to the appropriate annotation community for integration. The wheat study provided their results as a community resource via Apollo Jbrowse and GitHub repositories [<a href="#ref-2">2</a>]. This sharing enables community evaluation and integration.

Ethical Use of Public Data

Proteogenomic workflows often use publicly available data. The wheat study retrieved the largest recent wheat proteomics datasets publicly available [<a href="#ref-2">2</a>]. The use of public data requires proper attribution and compliance with data usage policies.

NCBI's documentation portal covers the technical architecture, record format standards, and programmatic API access patterns for each of its major biological databases. Researchers should consult these resources for data usage guidance.

Professional Escalation Criteria

When to Seek Expert Assistance

Proteogenomic workflows require expertise in mass spectrometry, bioinformatics, and genomics. Researchers who lack experience in any of these areas should seek collaboration or training before attempting annotation changes.

EMBL-EBI's freely accessible training courses cover database literacy, sequence analysis, structural biology, and expression data interpretation across the institute's core biological resources. Software Carpentry and Data Carpentry workshops build core computational literacy - Unix shell, scripted data analysis, and Git version control - that underpins reproducible scientific computing. These resources can build the skills needed for proteogenomic analysis.

When to Question Results

Certain results should trigger additional scrutiny. Novel gene claims supported by a single peptide require validation. Annotation corrections that contradict multiple existing evidence sources require careful review. Peptide identifications in pseudogenes require orthogonal supporting evidence before annotation changes [<a href="#ref-1">1</a>].

The GENCODE workflow provided a model for this scrutiny. The authors used a conservative approach with stringent filtering and required orthogonal evidence for annotation changes [<a href="#ref-1">1</a>].

When to Engage the Annotation Community

Annotation changes should be communicated to the relevant annotation community. The wheat study provided their proteogenomics results as a community resource [<a href="#ref-2">2</a>]. This sharing enables community evaluation and integration into official annotations.

The practical approach is to submit evidence to the appropriate annotation database or consortium. The NCBI provides search systems and analysis services that support this process [<a href="#ref-5">5</a>].

Frequently Asked Questions

What is the difference between proteomics and proteogenomics?

Proteomics identifies and quantifies proteins in a sample using mass spectrometry. Proteogenomics uses proteomics data to inform genome annotation by mapping identified peptides back to the genome. The Atlantic salmon study demonstrated this distinction by using proteogenomic databases constructed from transcriptomic data to identify peptides and then mapping those peptides to the genome to validate and revise gene models [<a href="#ref-3">3</a>].

How many peptides are needed to support a novel gene model?

There is no fixed number that applies across all situations. The GENCODE workflow recommended a conservative approach with stringent filtering and noted that many peptide identifications in pseudogenes could not be annotated due to the absence of orthogonal supporting evidence [<a href="#ref-1">1</a>]. In practice, multiple peptides from the same genomic region provide stronger evidence than a single peptide, and orthogonal transcriptomic evidence strengthens the case for annotation changes.

Can proteogenomics discover genes that transcriptomics misses?

Proteogenomics can discover genes that are not represented in the reference annotation, but the discovery depends on the proteins being expressed in the analyzed samples. The Atlantic salmon study identified novel previously unannotated genes through peptide evidence [<a href="#ref-3">3</a>]. The wheat study identified 2,934 orphan peptides that could not be mapped to existing gene models and proposed them as candidates for novel gene discovery [<a href="#ref-2">2</a>].

What causes false positive peptide identifications in proteogenomics?

False positives arise from multiple sources. Insufficiently stringent filtering is a primary cause, as demonstrated by the GENCODE workflow [<a href="#ref-1">1</a>]. Database contamination from six-frame translations of non-coding regions can also contribute. The multiple testing burden increases with database size, requiring careful FDR control.

How should multi-mapping peptides be handled?

Multi-mapping peptides map to multiple genomic locations and cannot provide unambiguous evidence for a specific locus. For novel gene discovery, they should be excluded because they cannot localize evidence. For validation of existing models, they can provide supporting evidence if the locus is already annotated. The interpretation should document how multi-mapping peptides were handled.

What is the role of RNA-Seq data in proteogenomic workflows?

RNA-Seq data serves two roles in proteogenomic workflows. First, it can be used to construct search databases that include novel transcripts not present in the reference annotation. The Atlantic salmon study constructed proteogenomic databases from publicly available transcriptomic data combined with in-house generated RNA-Seq data [<a href="#ref-3">3</a>]. Second, it provides orthogonal evidence that supports annotation changes suggested by peptide evidence. The rat study used integrated transcriptomic-proteomic analysis for this purpose [<a href="#ref-4">4</a>].

How does annotation quality affect downstream analysis?

Annotation errors propagate through every downstream analysis. Differential expression studies may miss genes that are not annotated. Evolutionary comparisons may be distorted by incorrect gene models. Functional genomics may attribute functions to incorrect loci. The wheat study demonstrated that peptide evidence can validate 31.4 percent of High Confidence gene models and support the promotion of Low Confidence models [<a href="#ref-2">2</a>].

What resources are available for learning proteogenomic analysis?

The Galaxy Training Network hosts community-maintained analytical workflow guides designed to minimize computational barriers for new practitioners.

Related Bioinformatics Guides

• Proteomics Mass Spectrometry: From Sample Preparation to Data Analysis
• Mass Spectrometry-Based Proteomics: Data Analysis Pipelines and Tools
• Spatial Proteomics Mass Spectrometry: Techniques and Applications
• Spatial Transcriptomics Workflow: From Sample Preparation to Data Analysis
• Mass Spectrometry Protein Identification: From Raw Spectra to Confident Hits

References and Further Reading

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.