# Building Custom Protein Databases from RNA-Seq Data for Proteogenomics: A Step-by-Step Guide

Proteogenomics integrates RNA sequencing (RNA-seq) data with mass spectrometry proteomics to improve peptide identification by constructing sample-specific protein databases. This approach captures variants, novel isoforms, and previously unannotated open reading frames that standard reference proteomes miss. This article provides a practical workflow for researchers who need to build custom protein databases from their own RNA-seq data, covering transcript assembly, translation, redundancy reduction, database formatting, quality checks, and common pitfalls.

## Scope and Reader Context

This workflow serves biology students, researchers, laboratory professionals, and life-science practitioners who generate or access RNA-seq data and want to use it to improve peptide identification in their proteomics experiments. The protocol assumes familiarity with basic command-line operations and access to a computing environment capable of running bioinformatics tools. The methods described here apply to any organism with a reference genome or transcriptome, though the specific tool choices may vary based on data type and research question.

The core problem this workflow solves is the limitation of reference protein databases. Standard databases such as UniProt or RefSeq contain annotated proteins, but they miss sample-specific mutations, alternative splicing events, RNA editing sites, and novel genes. RNA-seq data captures the actual transcripts expressed in your sample, providing a direct route to the proteins that could be present. By building a custom database from your RNA-seq data, you expand the search space for peptide identification and reduce the risk of missing biologically relevant proteins.

## Understanding Proteogenomics and Database Design Principles

Proteogenomics sits at the intersection of genomics, transcriptomics, and proteomics. The central idea is to use nucleotide-level information to inform protein identification. RNA-seq data provides the expressed transcript sequences in a given sample or tissue, and these sequences can be translated in silico to create a protein database tailored to that sample.

The design of a custom protein database requires balancing sensitivity and specificity. A larger database captures more potential protein sequences but increases the search space, which can lead to higher false discovery rates. A smaller database reduces the search space but risks missing relevant proteins. The workflow described here aims to produce a database that is comprehensive enough to capture sample-specific biology while remaining manageable for peptide-spectrum matching.

The National Center for Biotechnology Information (NCBI) maintains reference sequence databases that serve as the foundation for many proteogenomic studies. NCBI provides access to genome assemblies, transcript sequences, and protein annotations that can be used to validate or supplement custom databases. Understanding what is available in these public resources helps you decide whether to build a database from scratch or augment an existing reference.

The European Bioinformatics Institute (EMBL-EBI) offers training materials and data resources that support bioinformatics analysis, including proteomics and transcriptomics. These resources can help you understand the structure of public databases and the standards used for sequence data deposition.

## Core Principles of Custom Database Construction

### Transcript Assembly Approaches

The first major decision in building a custom protein database is how to generate transcript sequences from your RNA-seq data. Two primary approaches exist: reference-based assembly and de novo assembly.

Reference-based assembly aligns RNA-seq reads to a reference genome and then assembles transcripts from the aligned reads. This approach works well when a high-quality reference genome is available for your organism. Tools such as HISAT2 or STAR align reads to the genome, and then transcript assemblers such as StringTie or Cufflinks reconstruct isoform structures from the alignments. The resulting transcript models can include novel isoforms and splice variants that are not present in existing annotations.

De novo assembly builds transcripts without a reference genome by overlapping reads directly. This approach is necessary for organisms without a reference genome or when you want to capture sequences that are absent from the reference. Tools such as Trinity or rnaSPAdes generate contigs that represent expressed transcripts. De novo assembly is computationally intensive and may produce fragmented transcripts, but it can capture novel sequences that reference-based approaches miss.

The choice between these approaches depends on your research question and data availability. If you study a well-annotated model organism, reference-based assembly provides more accurate transcript structures. If you work with a non-model organism or suspect that the reference is incomplete, de novo assembly may be necessary.

### Translation of Transcripts to Protein Sequences

Once you have transcript sequences, you need to translate them into protein sequences. This step requires identifying open reading frames (ORFs) within each transcript. The longest ORF is often used as the default, but this approach can miss small proteins and alternative ORFs that have biological relevance.

Several tools can perform ORF prediction and translation. TransDecoder identifies long ORFs in transcript sequences and can be trained to recognize coding sequences based on sequence features. OrfPredictor and getorf from the EMBOSS suite provide similar functionality. For proteogenomics applications, you may want to retain all ORFs above a minimum length threshold instead of only the longest ORF, because short proteins and alternative reading frames can produce detectable peptides.

The translation process must account for the genetic code of your organism. Most tools default to the standard genetic code, but mitochondrial genomes and some organisms use alternative codes. Verify that your translation tool uses the correct genetic code for your species.

### Redundancy Reduction and Database Formatting

After translation, you will have a large set of protein sequences with substantial redundancy. Multiple transcripts from the same gene may translate to identical proteins, and different genes may produce similar sequences. Redundancy reduction improves search efficiency and reduces the risk of inflated false discovery rates.

CD-HIT is a widely used tool for clustering protein sequences and removing redundancy. It groups sequences by sequence identity thresholds, typically 90% or 100% identity for proteogenomics applications. The representative sequence from each cluster is retained in the final database. This step can dramatically reduce database size without losing biologically distinct sequences.

The final database must be formatted for use with peptide identification tools. Most search engines such as MaxQuant, Proteome Discoverer, or MS-GF+ accept FASTA-formatted protein databases. Some tools require additional indexing or formatting steps. Ensure that the protein identifiers in your FASTA file are unique and contain useful information such as the source transcript or gene identifier.

## At a Glance: Workflow Overview

| Workflow Step | Primary Tools | Input Data | Output | Key Quality Check |
| --- | --- | --- | --- | --- |
| Read quality control and preprocessing | FastQC, Trimmomatic, cutadapt | Raw RNA-seq reads | Cleaned reads | Per-base quality scores, adapter contamination levels |
| Transcript assembly | StringTie, Trinity, rnaSPAdes | Cleaned reads, reference genome (optional) | Assembled transcripts or contigs | Assembly completeness, transcript length distribution |
| ORF prediction and translation | TransDecoder, getorf | Assembled transcripts | Protein sequences | ORF length distribution, stop codon frequency |
| Redundancy reduction | CD-HIT | Translated proteins | Non-redundant protein set | Cluster statistics, sequence identity distribution |
| Database formatting and validation | Custom scripts, FASTA tools | Non-redundant proteins | Search-ready FASTA database | Protein count, sequence length distribution, identifier uniqueness |

## Practical Workflow: Step-by-Step Protocol

### Step 1: Data Preparation and Quality Control

Before assembling transcripts, you must ensure that your RNA-seq reads are of sufficient quality. Raw sequencing data often contains adapter sequences, low-quality bases, and contaminating sequences that can interfere with assembly.

Start by running FastQC on your raw reads to assess quality metrics including per-base quality scores, GC content, adapter contamination, and duplication levels. This initial assessment tells you whether your data requires trimming and what parameters to use.

Trim low-quality bases and adapter sequences using tools such as Trimmomatic or cutadapt. The trimming parameters depend on your sequencing platform and library preparation method. For Illumina data, common settings include removing bases below a Phred quality score of 20 and clipping adapter sequences when they appear at read ends.

After trimming, run FastQC again to confirm that quality issues have been resolved. Document the number of reads retained after trimming, as this information is important for interpreting downstream results and for reproducibility.

The Galaxy Training Network provides accessible tutorials on quality control and preprocessing that can help you understand the metrics and parameters involved. These tutorials are designed for researchers at various skill levels and include practical exercises.

### Step 2: Transcript Assembly

The assembly step converts cleaned reads into transcript sequences. The specific commands depend on whether you use reference-based or de novo assembly.

For reference-based assembly, first align your cleaned reads to the reference genome. STAR is a fast and accurate splice-aware aligner that works well for this purpose. After alignment, use StringTie to assemble transcripts from the aligned reads. StringTie can use existing gene annotations as a guide, which improves assembly accuracy for known genes while still allowing discovery of novel isoforms.

For de novo assembly, use Trinity or rnaSPAdes directly on the cleaned reads. These tools perform a series of steps including read normalization, contig assembly, and transcript reconstruction. De novo assembly requires substantial memory and computing time, especially for large genomes or deep sequencing data.

After assembly, assess the quality of your transcripts. Check metrics such as the number of assembled transcripts, the N50 length, and the proportion of reads that map back to the assembled transcripts. The nf-core documentation describes community standards for bioinformatics pipelines that include assembly quality assessment, and these standards can guide your evaluation.

### Step 3: ORF Prediction and Translation

With assembled transcripts in hand, you now predict open reading frames and translate them to protein sequences.

TransDecoder is a common choice for this step. It first identifies long ORFs in each transcript, then optionally uses homology searches and sequence composition features to distinguish true coding regions from spurious ORFs. For proteogenomics applications, you may want to retain all ORFs above a minimum length, typically 30 to 50 amino acids, instead of only the longest ORF per transcript.

The translation step produces protein sequences in FASTA format. Each protein entry should have a unique identifier that links back to the source transcript. This linkage is essential for interpreting which transcripts produce which proteins and for validating novel identifications.

The Bioconductor project provides R packages for genomic analysis that can assist with ORF prediction and sequence manipulation. These packages follow reproducible analysis standards and can be integrated into larger workflows.

### Step 4: Redundancy Reduction

The translated protein set will contain many identical or near-identical sequences. Redundancy reduction using CD-HIT clusters sequences at a specified identity threshold.

For proteogenomics databases, a threshold of 100% identity removes exact duplicates while retaining all distinct sequences. A threshold of 90% identity removes more redundancy but may collapse closely related isoforms that produce distinct peptides. The choice depends on your research question and the diversity of your sample.

After clustering, extract the representative sequences to create your non-redundant protein set. Record the number of clusters and the distribution of cluster sizes, as these metrics describe the complexity of your sample's proteome.

### Step 5: Database Formatting and Validation

The final step is formatting your protein sequences into a search-ready database and validating its quality.

Ensure that your FASTA file has unique identifiers and that the sequence lines are properly formatted. Some search engines have limits on identifier length or character types, so check the documentation for your specific tool.

Validate the database by checking the number of protein sequences, the distribution of sequence lengths, and the proportion of sequences that contain internal stop codons. Internal stop codons indicate translation errors or pseudogenes and should be investigated.

You can also compare your custom database to the reference proteome for your organism. The NCBI provides reference protein databases that can serve as a baseline for comparison. The overlap between your custom database and the reference indicates how much novel sequence information you have captured.

## Tool Selection and Tradeoffs

### Reference-Based vs. De Novo Assembly

Reference-based assembly produces more accurate transcript structures when a high-quality reference genome is available. It leverages the reference to resolve splice junctions and gene boundaries, resulting in transcripts that are easier to interpret biologically. The main limitation is that sequences absent from the reference genome cannot be assembled, even if they are expressed in your sample.

De novo assembly captures sequences that are missing from the reference, including novel genes and highly divergent alleles. The tradeoff is lower accuracy in transcript structure and higher computational requirements. De novo assemblies often contain chimeric transcripts and fragmented contigs that require additional filtering.

For most proteogenomics applications, a hybrid approach works well. Use reference-based assembly to capture known and novel isoforms, then add de novo assembly to capture sequences missing from the reference. Merge the two sets and remove redundancy to create a comprehensive database.

### ORF Prediction Strategies

The choice of ORF prediction strategy significantly affects database composition. Retaining only the longest ORF per transcript produces a smaller database with fewer false positives but may miss biologically relevant small proteins and alternative ORFs.

Retaining all ORFs above a minimum length produces a larger database that captures more potential proteins but increases the search space and false discovery risk. For proteogenomics discovery studies, retaining all ORFs above 30 amino acids is a reasonable starting point. You can then filter based on peptide identification results.

Some ORF prediction tools incorporate evidence from sequence conservation and codon usage to improve accuracy. TransDecoder can use BLAST searches against known proteins and Pfam domain searches to identify likely coding regions. These additional evidence sources reduce false positives but require access to external databases and increase runtime.

### Redundancy Thresholds

The redundancy reduction threshold affects both database size and the ability to distinguish closely related proteins. A 100% identity threshold removes only exact duplicates, preserving all sequence variants. This is appropriate when you want to detect single amino acid variants that could be biologically significant.

A 90% identity threshold collapses sequences that differ by a small number of amino acids. This reduces database size and improves search speed but may merge isoforms that produce distinct peptides. For quantitative proteomics, where you want to assign peptides to specific proteins, a higher identity threshold is preferable.

The EMBL-EBI training resources provide guidance on sequence similarity and clustering concepts that inform this decision.

## Observations and Measurements

### Key Metrics to Track

Throughout the workflow, you should record metrics that describe your data and the decisions you make. These records support reproducibility and help you troubleshoot problems.

For read preprocessing, record the number of raw reads, the number of reads retained after trimming, and the proportion of reads removed due to quality or adapter issues. These numbers indicate the quality of your starting material and the stringency of your trimming parameters.

For transcript assembly, record the number of assembled transcripts, the N50 length, and the proportion of reads that map back to the assembly. These metrics indicate the completeness and accuracy of your assembly.

For ORF prediction and translation, record the number of predicted ORFs, the length distribution of predicted proteins, and the proportion of transcripts that yield at least one ORF. These metrics indicate how much coding potential exists in your transcriptome.

For redundancy reduction, record the number of input sequences, the number of output clusters, and the reduction ratio. These metrics describe the redundancy level in your sample.

For the final database, record the total number of protein sequences, the sequence length distribution, and the number of sequences with internal stop codons. These metrics describe the database you will use for peptide identification.

### Quality Control Checkpoints

At each workflow step, you should perform quality checks before proceeding. The Galaxy Training Network provides practical guidance on quality assessment at various stages of analysis.

After read trimming, verify that quality scores have improved and that adapter contamination has been removed. After assembly, check that transcript lengths are reasonable and that known housekeeping genes are present. After translation, verify that protein lengths follow an expected distribution and that stop codons appear only at sequence ends. After redundancy reduction, confirm that the representative sequences are biologically meaningful.

These checkpoints catch errors early, before they propagate through the workflow and compromise your final database.

## Common Failure Patterns and Troubleshooting

### Low Assembly Completeness

If your transcript assembly produces few transcripts or short transcripts, several issues may be responsible. Insufficient sequencing depth is a common cause, particularly for lowly expressed genes. Check your sequencing depth and consider whether you need additional sequencing.

Poor read quality after trimming can also reduce assembly completeness. Verify that your trimming parameters are appropriate and that you have not removed too much sequence. Overly aggressive trimming can remove biologically relevant sequence along with low-quality bases.

Contamination from other organisms can interfere with assembly. Check for contamination by aligning a sample of reads to your reference genome or by running taxonomic classification tools.

### Excessive Redundancy in the Final Database

If your final database contains many near-identical sequences, your redundancy reduction may be insufficient. Check the identity threshold used in CD-HIT and consider whether a lower threshold is appropriate.

Alternatively, excessive redundancy may indicate that your transcript assembly produced many similar isoforms. This can happen when assembly parameters are too permissive, allowing minor variations to be assembled as separate transcripts. Review your assembly parameters and consider filtering transcripts by expression level.

### High False Discovery Rates in Peptide Identification

A custom database that is too large can lead to inflated false discovery rates in downstream peptide identification. If you observe high false discovery rates, consider whether your database contains too many spurious sequences.

Review your ORF prediction parameters and consider increasing the minimum ORF length or requiring additional evidence for coding potential. You can also filter your database based on peptide identification results, removing proteins that are never identified.

The nf-core documentation describes quality control standards for bioinformatics pipelines that can help you establish appropriate filtering criteria.

## Limitations and Interpretation Constraints

### Database Completeness

A custom protein database built from RNA-seq data captures the transcripts expressed in your sample at the time of sequencing. It does not capture proteins that are present due to post-transcriptional regulation, protein stability differences, or translation from stored mRNAs. The database reflects the transcriptome, not the proteome, and this distinction is important for interpreting results.

### Variant Detection Limitations

RNA-seq data provides information about transcribed sequences, but variant calling from RNA-seq has limitations. Mapping errors, sequencing errors, and RNA editing can produce apparent variants that are not present in the genome. Validation of novel variants requires additional evidence, such as genomic sequencing or targeted mass spectrometry.

### Sample Specificity

A custom database built from one sample may not be appropriate for other samples from the same organism. Gene expression varies across tissues, developmental stages, and conditions. If you are studying multiple samples, consider building a database that represents the union of transcripts across all samples, or build separate databases for each sample.

The GEPIA2 web server provides tools for comparing gene expression across tumor and normal samples from public databases, which can help you understand the expected expression patterns in your system.

## Safety and Reproducibility Context

### Computational Resource Requirements

Building custom protein databases requires substantial computational resources, particularly for de novo assembly of large transcriptomes. Ensure that your computing environment has sufficient memory, disk space, and processing power before starting the workflow. The Carpentries lessons provide foundational training in command-line computing that can help you manage computational tasks efficiently.

### Reproducibility Standards

Reproducibility is essential for proteogenomics research. Document all parameters used at each workflow step, including software versions, reference genome versions, and threshold settings. Consider using workflow management systems such as nf-core pipelines, which enforce reproducibility standards and provide structured documentation.

The Bioconductor project emphasizes reproducible research practices and provides tools for version control and workflow documentation. Adopting these practices ensures that your results can be reproduced by others and that your methods are transparent.

### Data Management

RNA-seq data and derived databases are large files that require careful management. Store raw data in public repositories such as the NCBI Sequence Read Archive to ensure long-term availability. Document file naming conventions and directory structures to facilitate data sharing and collaboration.

## Professional Escalation Criteria

### When to Seek Additional Expertise

If you encounter persistent problems with assembly quality, ORF prediction accuracy, or database performance, consider consulting with bioinformatics specialists or core facilities. These experts can review your workflow, identify issues, and recommend alternative approaches.

The EMBL-EBI training resources and the Galaxy Training Network provide pathways for developing additional skills. If your project requires advanced analysis beyond your current expertise, these resources can help you build the necessary knowledge.

### When to Validate with Orthogonal Methods

Novel protein identifications from a custom database should be validated with orthogonal methods before drawing biological conclusions. Targeted mass spectrometry, western blotting, or antibody-based validation can confirm the presence of predicted proteins. If you identify proteins that are not in reference databases, validation is particularly important.

The NCBI provides tools and resources for sequence analysis that can help you characterize novel sequences and assess their biological plausibility.

## Decision Framework for Database Scope and Search Strategy

The workflow steps described above produce a functional custom protein database, but the choices you make about database scope and search strategy determine whether that database delivers meaningful proteogenomic discoveries or simply increases your false discovery rate. This section provides a practical decision framework for matching database construction choices to your specific research question, sample type, and validation capacity. The framework is organized around three decision points that occur before you begin assembly and two decision points that occur after you build your first database version.

### Decision Point 1: Define the Biological Question Before Choosing Assembly Strategy

The first decision is not technical but biological. What type of protein diversity do you expect in your sample, and what type of novel identifications would be biologically meaningful for your study? This decision determines whether you need reference-based assembly, de novo assembly, or a hybrid approach.

For studies focused on known genes and isoforms, reference-based assembly with StringTie using existing annotations as a guide is sufficient. This approach captures alternative splicing events and sample-specific isoform usage while maintaining transcript structures that align with the reference genome. The resulting database will be smaller and more specific, which reduces false discovery risk in downstream peptide identification.

For studies investigating novel genes, unannotated open reading frames, or highly divergent sequences, de novo assembly with Trinity or rnaSPAdes is necessary. This approach captures sequences absent from the reference genome, including novel genes and highly divergent alleles. The tradeoff is lower accuracy in transcript structure and higher computational requirements.

For discovery-oriented proteogenomics where you want maximum coverage, use a hybrid approach. Assemble transcripts using both reference-based and de novo methods, then merge the results and remove redundancy. This produces the most comprehensive database but requires careful filtering to control database size.

The decision framework for assembly strategy can be summarized as follows. If you study a well-annotated model organism and your question concerns known biology, use reference-based assembly alone. If you study a non-model organism or suspect the reference is incomplete, use de novo assembly. If you want to maximize novel identifications and have the computational resources and validation capacity, use the hybrid approach.

### Decision Point 2: Match Database Size to Search Engine Capabilities

The size of your custom database directly affects peptide-spectrum matching performance and false discovery rates. Before building your database, consider the capabilities of your peptide identification tool and the computational resources available for the search step.

Most modern search engines such as MaxQuant, Proteome Discoverer, and MS-GF+ can handle databases with hundreds of thousands of protein sequences, but search time increases with database size. A database that is too large can make the search computationally prohibitive, especially for large-scale proteomics experiments with many samples.

The practical size limit depends on your search engine and hardware. As a general framework, databases with fewer than 100,000 protein sequences are manageable with standard desktop workstations. Databases with 100,000 to 500,000 sequences require a server or high-performance computing cluster. Databases exceeding 500,000 sequences may require specialized search strategies such as two-pass searching or peptide-centric approaches.

If your initial database exceeds your search capacity, you have several options. Increase the minimum ORF length threshold to remove short spurious sequences. Apply expression filtering to retain only transcripts above a minimum expression level. Use a higher redundancy reduction threshold in CD-HIT to collapse closely related sequences. Each of these options reduces database size but may remove biologically relevant sequences, so document the tradeoffs.

### Decision Point 3: Plan for Decoy-Based False Discovery Rate Control

Custom protein databases require careful false discovery rate control because they contain many sequences that are not present in reference databases. The standard approach is to append reversed or shuffled decoy sequences to your target database before searching. The proportion of decoy hits among your identified peptides estimates the false discovery rate.

The decision here is whether to use a single combined target-decoy database or separate target and decoy databases. Most search engines support the combined approach, where decoy sequences are appended to the target database and the search engine distinguishes them by a prefix in the protein identifier. This approach is simpler and is the default for most tools.

The key decision is the decoy strategy itself. Reversed decoys are the most common choice because they preserve amino acid composition and length distribution. Shuffled decoys preserve composition but not sequence order. For custom databases that contain many novel sequences, reversed decoys are preferred because they provide a more accurate estimate of false discoveries.

The nf-core documentation describes community standards for false discovery rate control in proteomics pipelines. Review these standards before designing your search strategy to ensure your approach aligns with current best practices.

### Decision Point 4: Evaluate Database Performance with a Pilot Search

Before committing to a full-scale search across all your samples, run a pilot search using a subset of your mass spectrometry data. This pilot search serves two purposes: it validates that your database is properly formatted and searchable, and it provides preliminary metrics on identification rates and false discovery rates.

For the pilot search, use a single mass spectrometry run or a small pool of runs that represent the diversity of your samples. Search against your custom database with standard parameters and record the following metrics: the number of peptide-spectrum matches, the number of unique peptides identified, the number of unique proteins identified, and the false discovery rate at the peptide and protein levels.

Compare these metrics to a search against the reference proteome for your organism. The custom database should identify at least as many peptides as the reference database, and ideally more. If the custom database identifies fewer peptides, your database construction may have introduced errors or removed sequences that are present in the reference.

The pilot search also reveals whether your database contains problematic sequences. A high proportion of peptides that map to decoy sequences indicates that your database contains too many spurious sequences. A high proportion of peptides that map to multiple proteins indicates that your redundancy reduction was insufficient.

The GEPIA2 web server provides tools for comparing gene expression across tumor and normal samples from public databases. If you are working with cancer samples, you can use this resource to check whether the genes represented in your custom database show expected expression patterns in your sample type.

### Decision Point 5: Iterate Based on Identification Results

Database construction is not a single-pass process. After your pilot search, you will likely need to iterate on your database to improve identification rates and reduce false discoveries. The iteration framework below describes common scenarios and the appropriate response for each.

If your false discovery rate is too high, your database likely contains too many spurious sequences. Increase the minimum ORF length threshold, apply expression filtering, or require additional evidence for coding potential such as homology to known proteins. You can also filter your database based on pilot search results, removing proteins that are never identified.

If your identification rate is lower than expected, your database may be missing relevant sequences. Check whether your transcript assembly captured known housekeeping genes and highly expressed transcripts. If these are missing, your assembly parameters may be too stringent or your sequencing depth may be insufficient.

If you identify many peptides that map to multiple proteins, your redundancy reduction was insufficient. Rerun CD-HIT with a lower identity threshold or apply additional filtering to collapse closely related isoforms.

If you identify peptides that map to novel sequences not present in the reference proteome, validate these identifications carefully. The NCBI provides tools for sequence analysis that can help you characterize novel sequences and assess their biological plausibility. Novel identifications should be validated with orthogonal methods before drawing biological conclusions.

### Record System for Database Construction Decisions

Reproducibility in proteogenomics requires documenting also the commands you run but also the decisions you make and the rationale behind them. The record system below provides a structured approach to documenting your database construction workflow.

Create a decision log that records each decision point, the options considered, the choice made, and the rationale. For each decision, record the date, the person making the decision, and any data that informed the choice. This log serves as the primary record of your database construction process.

Create a parameter file that records all software versions and parameters used at each workflow step. Include the version of each tool, the reference genome version, the assembly parameters, the ORF prediction parameters, and the redundancy reduction threshold. This file enables exact reproduction of your database.

Create a quality metrics file that records the metrics described in the observations section above. Include the number of reads before and after trimming, the number of assembled transcripts, the N50 length, the number of predicted ORFs, the number of proteins after redundancy reduction, and the final database size. These metrics provide a quantitative description of your database.

Create a search results file that records the pilot search metrics and any iterations performed. Include the false discovery rate at each iteration, the number of identified peptides and proteins, and the changes made to the database between iterations. This file documents the evolution of your database.

The Carpentries lessons provide foundational training in data management and reproducible research practices. These lessons cover file organization, version control, and documentation standards that support the record system described here.

### Troubleshooting Method for Database Construction Failures

When database construction fails, the cause is often identifiable from the quality metrics recorded at each step. The troubleshooting method below describes a systematic approach to diagnosing and resolving common failures.

If your transcript assembly produces few transcripts or short transcripts, check your sequencing depth and read quality. Insufficient sequencing depth is a common cause, particularly for lowly expressed genes. Poor read quality after trimming can also reduce assembly completeness. Verify that your trimming parameters are appropriate and that you have not removed too much sequence.

If your ORF prediction produces few proteins or many proteins with internal stop codons, check your transcript sequences for frameshift errors. These errors can result from sequencing errors or assembly mistakes. Consider using a tool that incorporates homology evidence to identify likely coding regions.

If your redundancy reduction removes too many sequences, check your identity threshold. A threshold that is too low can collapse biologically distinct isoforms. Consider using a higher identity threshold or reviewing the cluster representatives to ensure they are biologically meaningful.

If your final database produces poor search results, check the database formatting. Some search engines have specific requirements for FASTA headers and sequence line lengths. Verify that your identifiers are unique and that your sequences contain only standard amino acid characters.

The Galaxy Training Network provides practical tutorials on quality assessment and troubleshooting at various stages of bioinformatics analysis. These tutorials can help you diagnose specific problems in your workflow.

### Comparison of Database Construction Approaches

The table below compares the three main database construction approaches across key dimensions. This comparison supports the decision framework described above.

| Dimension | Reference-Based Only | De Novo Only | Hybrid Approach |
| --- | --- | --- | --- |
| Reference genome required | Yes | No | Yes |
| Captures novel genes | No | Yes | Yes |
| Transcript structure accuracy | High | Moderate | High for known, moderate for novel |
| Computational requirements | Moderate | High | Highest |
| Database size | Smaller | Larger | Largest |
| False discovery risk | Lower | Higher | Highest |
| Best use case | Well-annotated organisms | Non-model organisms | Discovery-oriented studies |

The choice between these approaches depends on your research question, your organism, and your computational resources. The hybrid approach provides the most comprehensive database but requires the most careful filtering and validation.

### Professional Escalation Criteria for Database Construction

Some database construction problems require expertise beyond what this workflow provides. The escalation criteria below describe situations where you should consult with bioinformatics specialists or core facilities.

If your transcript assembly consistently produces poor results across multiple parameter sets, consult with a bioinformatics specialist. The problem may be in your input data, your reference genome, or your assembly strategy. A specialist can review your data and recommend alternative approaches.

If your peptide identification results show unexplained patterns such as systematic false discoveries or poor identification rates, consult with a proteomics specialist. The problem may be in your search parameters, your mass spectrometry data, or your database construction. A specialist can help you diagnose the issue.

If you identify novel proteins that are biologically significant but cannot validate them with available methods, consult with experts in your biological system. They can help you design validation experiments and assess the biological plausibility of your findings.

The EMBL-EBI training resources and the Galaxy Training Network provide pathways for developing additional skills. If your project requires advanced analysis beyond your current expertise, these resources can help you build the necessary knowledge.

## Frequently Asked Questions

### What is the minimum sequencing depth required for building a custom protein database?

The required sequencing depth depends on your organism, sample complexity, and research question. Deeper sequencing captures more lowly expressed transcripts and improves assembly completeness. For most applications, 20 to 50 million paired-end reads per sample provides a reasonable starting point, but you should assess assembly completeness and add sequencing if needed.

### Can I use publicly available RNA-seq data instead of generating my own?

Public RNA-seq data from repositories such as the NCBI Sequence Read Archive or the Gene Expression Omnibus can be used to build custom databases. This approach is useful when you want to study samples that you cannot collect yourself. Ensure that the data quality is sufficient and that the sample conditions match your research question.

### How do I choose between reference-based and de novo assembly?

Reference-based assembly is preferred when a high-quality reference genome is available for your organism. De novo assembly is necessary when no reference exists or when you want to capture sequences absent from the reference. A hybrid approach that combines both methods provides the most comprehensive database.

### What minimum ORF length should I use for translation?

A minimum ORF length of 30 amino acids is a common starting point for proteogenomics applications. Shorter ORFs may be biologically relevant but are more difficult to distinguish from spurious sequences. You can adjust the threshold based on your peptide identification results and the expected protein sizes in your organism.

### How does database size affect peptide identification?

Larger databases increase the search space for peptide-spectrum matching, which can lead to higher false discovery rates. Smaller databases reduce the search space but may miss relevant proteins. The goal is to build a database that captures the proteins present in your sample without excessive spurious sequences.

### Should I include known reference proteins in my custom database?

Including reference proteins from databases such as NCBI RefSeq or UniProt can improve coverage and provide a baseline for comparison. Many proteogenomics workflows combine custom and reference sequences, then remove redundancy. This approach ensures that well-annotated proteins are included while adding sample-specific sequences.

### How do I validate novel proteins identified from a custom database?

Novel proteins should be validated using orthogonal methods such as targeted mass spectrometry, western blotting, or antibody-based detection. You can also assess the biological plausibility of novel sequences by checking for conserved domains, expression patterns, and evolutionary conservation.

### What are the common mistakes in building custom protein databases?

Common mistakes include using poor quality RNA-seq data without adequate preprocessing, choosing inappropriate assembly parameters, retaining only the longest ORF and missing small proteins, failing to remove redundancy, and not validating the final database. Each of these mistakes can compromise the quality of your peptide identification results.

## Related Bioinformatics Guides

- [RNA-Seq Databases: Accessing and Using Public RNA-Seq Data](/knowledge/bioinformatics/rna-seq-databases-accessing-and-using-public-rna-seq-data)
- [RNA-Seq Data Analysis Workflow: From Raw Reads to Insights](/knowledge/bioinformatics/rna-seq-data-analysis-workflow-from-raw-reads-to-insights)
- [Metabolomics Data Analysis in R: A Practical Workflow](/knowledge/bioinformatics/metabolomics-data-analysis-in-r-a-practical-workflow)
- [Spatial Transcriptomics Workflow: From Sample Preparation to Data Analysis](/knowledge/bioinformatics/spatial-transcriptomics-workflow-from-sample-preparation-to-data-analysis)
- [Genomic Data Infrastructure: Building and Managing Large-Scale Genomic Databases](/knowledge/bioinformatics/genomic-data-infrastructure-building-and-managing-large-scale-genomic-databases)

## References and Further Reading

- [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/). National Center for Biotechnology Information.
- [EMBL-EBI Training](https://www.ebi.ac.uk/training). European Bioinformatics Institute.
- [Bioconductor](https://bioconductor.org/). Bioconductor Project.
- [Galaxy Training Network](https://training.galaxyproject.org/). Galaxy Project.
- [nf-core Documentation](https://nf-co.re/docs). nf-core.
- [The Carpentries Lessons](https://carpentries.org/lessons). The Carpentries.
- [GEPIA2: an enhanced web server for large-scale expression profiling and interactive analysis.](https://pubmed.ncbi.nlm.nih.gov/31114875). Nucleic acids research, 2019.
- [Nicotinamide N-methyltransferase negatively regulates metastasis-promoting property of cancer-associated fibroblasts in lung adenocarcinoma.](https://pubmed.ncbi.nlm.nih.gov/39623600). Cancer communications (London, England), 2025.
- [Single-nucleus RNA sequencing and network pharmacology reveal the mediation of fisetin on neuroinflammation in Alzheimer's disease.](https://pubmed.ncbi.nlm.nih.gov/40215814). Phytomedicine : international journal of phytotherapy and phytopharmacology, 2025.
- [Targeted delivery of CCL3 reprograms macrophage antigen presentation and enhances the efficacy of immune checkpoint blockade therapy in hepatocellular carcinoma.](https://pubmed.ncbi.nlm.nih.gov/39988347). Journal for immunotherapy of cancer, 2025.
- [Single-cell sequencing combined with spatial transcriptomics reveals that the IRF7 gene in M1 macrophages inhibits the occurrence of pancreatic cancer by regulating lipid metabolism-related mechanisms.](https://pubmed.ncbi.nlm.nih.gov/39118300). Clinical and translational medicine, 2024.

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