The Role of Long-Read RNA-Seq in Annotating Complex Genomes: A Guide to Improving Gene Models

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

Genome annotation projects that rely exclusively on short-read RNA-seq or ab initio prediction often produce gene models that miss transcript isoforms, misassign splice junctions, and fail to capture allele-specific expression patterns. Long-read RNA-seq provides full-length transcript information that directly addresses these gaps, but integrating this data into an existing annotation pipeline requires deliberate workflow decisions. This article provides a practical framework for researchers who need to update gene models using long-read transcript data, covering library considerations, alignment and transcript assembly, quality filtering with tools such as SQANTI and GFFcompare, and the integration of evidence into gene prediction pipelines.

The intended reader is a biology student, researcher, or laboratory professional who has generated or plans to generate long-read RNA-seq data and needs a concrete path from raw reads to improved gene models. The guidance assumes familiarity with basic RNA-seq analysis but does not assume prior experience with long-read data. The focus is on eukaryotic genomes, where transcript complexity and genome size create the most pressing annotation challenges.

Key Takeaways

• Long-read RNA-seq improves genome annotation by sequencing full-length transcripts, directly revealing exon ordering, alternative splicing, and polyadenylation that short-read data cannot infer.
• Libraries with longer and more accurate sequences produce more accurate transcript models than libraries with increased read depth alone, so prioritize read length over throughput for annotation.
• Reference-based alignment tools outperform reference-free assembly methods in well-annotated genomes, while reference-free approaches suit fragmented assemblies or rare novel transcript discovery.
• The BRAKER3 pipeline integrates short-read RNA-seq and protein databases with long-read evidence, raising average transcript-level F1-scores by about 20 percentage points for complex genomes.
• For refining gene models, define the annotation objective first, then generate or obtain long-read data, preprocess reads, align to the genome, and assemble transcript models.

At a Glance

The table below summarizes the key workflow stages, the primary decisions at each stage, and the expected outcome for a genome annotation project.

The central finding from recent benchmarking is that libraries with longer and more accurate sequences produce more accurate transcripts than libraries with increased read depth alone, while greater read depth improves quantification accuracy. This distinction matters for annotation projects because the goal is transcript structure accuracy, not expression quantification. Researchers should prioritize read length and base accuracy over sheer throughput when the objective is gene model refinement.

Why Long-Read RNA-Seq Changes the Annotation Landscape

Short-read RNA-seq has been the standard evidence source for genome annotation for over a decade. Reads of 100 to 150 base pairs are aligned to the genome, and spliced alignments are used to infer exon-intron boundaries. This approach works well for highly expressed genes with canonical splice sites, but it struggles with lowly expressed isoforms, complex alternative splicing patterns, and genes with very large introns. The fundamental limitation is that short reads cannot span entire transcripts, so isoform structure must be inferred from read coverage patterns and splice junction evidence.

Long-read RNA-seq removes this limitation by sequencing full-length or near-full-length transcripts. Oxford Nanopore Technologies and Pacific Biosciences platforms produce reads that can span entire transcripts, providing direct evidence for exon ordering, alternative promoter usage, alternative splicing, and alternative polyadenylation. The value of this information for genome annotation is substantial. A large human long-read RNA-seq dataset generated from 88 Genotype-Tissue Expression samples identified over 70,000 novel transcripts for annotated genes, with protein expression validated for 10 percent of these novel transcripts. This scale of discovery is not possible with short-read data alone.

The practical implication for annotation projects is that long-read data can resolve transcript structures that short-read data cannot. When a gene model is built from short reads, the annotated isoforms represent a best guess based on coverage patterns. When the same gene model is built from long reads, each isoform is directly observed as a contiguous transcript sequence. The difference is between inference and observation.

Core Principles for Evidence-Driven Annotation

Transcript Accuracy Depends on Read Length and Base Quality

The Long-read RNA-Seq Genome Annotation Assessment Project Consortium generated over 427 million long-read sequences from complementary DNA and direct RNA datasets across human, mouse, and manatee species. The consortium's systematic assessment found that libraries with longer and more accurate sequences produce more accurate transcripts than libraries with increased read depth. This finding has a direct consequence for experimental design. If the goal is to improve gene models, the sequencing strategy should prioritize read length and base accuracy. Increasing read depth will improve quantification accuracy, but it will not compensate for short or error-prone reads when the objective is transcript structure determination.

For practical workflow design, this means that researchers should evaluate the expected read length distribution of their chosen platform and chemistry before committing to a sequencing run. Oxford Nanopore R9.4.1 chemistry and Pacific Biosciences Sequel II circular consensus sequencing represent two commonly used approaches, and the choice between them affects downstream analysis options. The evaluation of sequencing platforms and data processing methods for evidence-driven genome annotation found that incorporating PacBio transcripts into an annotation pipeline significantly outperformed traditional methods such as ab initio predictions and short-read-based annotations. This result does not mean that Nanopore data is unsuitable for annotation, but it does indicate that platform-specific error profiles and read length distributions influence the quality of the resulting transcript models.

Reference-Based Approaches Outperform Reference-Free Methods in Annotated Genomes

The consortium assessment also found that in well-annotated genomes, tools based on reference sequences demonstrated the best performance. Reference-based approaches align long reads to the genome and use the alignment structure to define transcript models. Reference-free approaches assemble transcripts from the reads themselves without a genome guide. For annotation projects in organisms with an existing genome assembly, reference-based methods are the appropriate starting point.

The exception is when the goal is to detect rare and novel transcripts or when the genome assembly is highly fragmented. In these cases, the consortium advises incorporating additional orthogonal data and replicate samples. Reference-free approaches can identify transcripts that are absent from the reference annotation, but they require more careful validation because the assembly process can introduce artifacts.

Orthogonal Data Improves Novel Transcript Detection

The consortium's recommendation to incorporate orthogonal data when aiming to detect rare and novel transcripts reflects a broader principle in genome annotation. No single data type provides complete evidence for all transcript structures. Protein evidence, short-read RNA-seq, and long-read RNA-seq each contribute different information. Protein evidence confirms that a transcript is translated. Short-read RNA-seq provides quantitative expression information and splice junction support. Long-read RNA-seq provides full-length transcript structure.

The BRAKER3 pipeline exemplifies this principle by integrating short-read RNA-seq and a large protein database with statistical models learned iteratively for the target genome. The pipeline outperformed earlier versions that used only RNA-seq or only protein data, with the average transcript-level F1-score increased by about 20 percentage points. The improvement was most pronounced for species with large and complex genomes. This result demonstrates that the combination of evidence types produces better gene models than any single evidence source.

Practical Workflow for Integrating Long-Read RNA-Seq

Step 1: Define the Annotation Objective and Data Requirements

Before generating or analyzing long-read RNA-seq data, define the specific annotation problem. The workflow differs depending on whether the goal is to refine existing gene models, discover novel transcripts, or annotate a newly assembled genome from scratch.

For refining existing gene models, the primary need is high-quality transcript models that can be compared against the current annotation. The sequencing strategy should prioritize read length and base accuracy. For discovering novel transcripts, additional considerations apply. The consortium found that incorporating orthogonal data and replicate samples is advised when aiming to detect rare and novel transcripts. This means that the experimental design should include biological replicates and possibly multiple tissue types or conditions.

For annotating a newly assembled genome, the workflow must integrate long-read transcript evidence with ab initio prediction and protein evidence. The BRAKER3 pipeline provides an automated approach that combines these evidence types, but the input data must be prepared appropriately.

Step 2: Generate or Obtain Long-Read RNA-Seq Data

The choice of sequencing platform and library preparation method determines the characteristics of the data. Complementary DNA protocols convert RNA to cDNA before sequencing, while direct RNA protocols sequence the RNA molecule itself. The consortium generated data from both complementary DNA and direct RNA datasets, and the choice affects read length, accuracy, and the ability to detect certain modifications.

For annotation purposes, the key parameters are read length and base accuracy. Longer reads are more likely to span entire transcripts, and more accurate reads produce more reliable splice junction calls. The consortium's finding that libraries with longer and more accurate sequences produce more accurate transcripts should guide the platform and chemistry selection.

If the data is being generated in-house, consider running a small test library first to assess read length distribution and throughput before committing to a full sequencing run. If the data is being obtained from a public repository, the NCBI provides search systems and sequence resources that can be used to identify suitable datasets. The NCBI Data Resources page describes the available databases and search systems for sequence data.

Step 3: Preprocess Reads and Assess Data Quality

Raw long-read data requires preprocessing before alignment. The specific steps depend on the platform. For Oxford Nanopore data, basecalling and adapter trimming are typically required. For PacBio data, circular consensus sequencing produces highly accurate reads that may require minimal preprocessing.

Quality assessment should include read length distribution, estimated base accuracy, and the number of reads that map to the genome. The EMBL-EBI Training program offers learning pathways for bioinformatics data analysis that cover quality assessment and preprocessing approaches. These training materials provide practical guidance for researchers who are new to long-read data analysis.

The preprocessing step should also include a check for contamination and adapter sequences. Contamination from other organisms can produce spurious transcript models, and adapter sequences can interfere with alignment.

Step 4: Align Reads to the Genome

Alignment of long reads to the genome is a critical step because the alignment structure defines the transcript model. Long-read aligners are designed to handle reads that span introns and contain sequencing errors. The choice of aligner affects the sensitivity and specificity of splice junction detection.

For reference-based approaches, the genome assembly must be of sufficient quality to support accurate alignment. Fragmented assemblies with many small contigs will produce fragmented transcript models. If the genome assembly is highly fragmented, consider whether additional genome assembly work is needed before proceeding with transcript annotation.

The alignment step produces BAM files that contain the read alignments. These files should be inspected for alignment statistics, including the percentage of reads that map, the distribution of mapping quality, and the distribution of alignment lengths. Poor alignment rates may indicate sample contamination, library preparation problems, or genome assembly issues.

Step 5: Assemble Transcript Models

Transcript assembly from long reads uses the aligned reads to construct full-length transcript models. The assembly process groups reads that share the same splice junctions and exon structure into transcript models. The output is a GTF or GFF file that describes the predicted transcripts.

The choice between reference-based and reference-free assembly depends on the annotation context. For well-annotated genomes, reference-based approaches demonstrate the best performance. For genomes with limited annotation or when the goal is to discover novel transcripts, reference-free approaches may be appropriate, but they require additional validation.

The assembly step produces a set of transcript models that must be filtered and classified before integration into the annotation. The raw assembly will contain artifacts from sequencing errors, alignment errors, and incomplete transcript coverage.

Step 6: Filter and Classify Transcripts with SQANTI

SQANTI is a tool designed to filter and classify long-read transcript models. It assigns each transcript to a structural category based on its relationship to the reference annotation. The categories include full splice match, incomplete splice match, novel in catalog, novel not in catalog, and others. This classification provides a basis for deciding which transcripts to retain for annotation.

The filtering step removes transcripts that are likely artifacts. SQANTI applies multiple filters based on the number of reads supporting a transcript, the presence of non-canonical splice sites, and other structural features. The specific filter thresholds should be set based on the data quality and the annotation objective.

For annotation projects, the transcripts classified as full splice match and novel in catalog are the most useful. Full splice match transcripts confirm existing annotation models. Novel in catalog transcripts use known splice sites but in new combinations, providing evidence for alternative isoforms. Novel not in catalog transcripts use unannotated splice sites and require additional validation before being added to the annotation.

Step 7: Compare Against Reference Annotation with GFFcompare

GFFcompare is a tool that compares transcript models against a reference annotation. It provides statistics on the number of transcripts that match known models, the number of novel transcripts, and the level of agreement between the two sets. This comparison is essential for quantifying the improvement that long-read data provides.

The comparison produces a set of statistics that can be used to assess the completeness of the long-read transcript set. The transcript-level agreement between long-read-based annotation and short-read-based annotation can be surprisingly low. In the Florida manatee study, the loci-level agreement between long-read-based and short-read-based annotation was 90 percent, but the transcript-level agreement was only 35 percent. This discrepancy highlights the value of long-read data for resolving isoform structure that short-read data cannot distinguish.

The GFFcompare output also identifies transcripts that are present in the long-read data but absent from the reference annotation. These novel transcripts require further investigation to determine whether they represent genuine biological isoforms or artifacts.

Step 8: Integrate Transcript Evidence into Gene Prediction

The filtered and classified transcript models serve as evidence for gene prediction. The integration approach depends on the annotation pipeline being used. BRAKER3 provides an automated pipeline that uses short-read RNA-seq and a large protein database, along with statistical models learned iteratively for the target genome. The pipeline outperformed other existing tools including MAKER2, Funannotate, and FINDER in benchmarking studies.

The integration of long-read transcript evidence into gene prediction can be approached in two ways. The first approach uses the long-read transcript models as a training set for ab initio gene predictors. The second approach uses the transcript models as direct evidence for gene structure, with ab initio prediction filling in regions where transcript evidence is absent.

The evaluation of evidence-driven genome annotation using long-read RNA-seq found that incorporating PacBio transcripts into the annotation pipeline significantly outperformed traditional methods. The study also identified 4906 novel loci represented by 5707 isoforms, with 64 percent of these isoforms matching known sequences in other mammalian species. This result demonstrates that long-read data can identify biologically meaningful novel loci that are absent from short-read-based annotations.

Step 9: Validate and Report the Updated Gene Models

Validation of updated gene models involves multiple checks. The transcript models should be examined for consistency with protein evidence. The presence of a complete open reading frame and homology to known proteins in other species provides support for the validity of a transcript model.

The reporting of annotation updates should include the number of transcripts added, the number of transcripts modified, and the number of transcripts that were not supported by the long-read data. The comparison statistics from GFFcompare provide a quantitative basis for this reporting.

The updated gene models should be made available in a format that supports downstream analysis. The NCBI provides databases and analysis services that can be used to submit and access annotation data.

Options and Tradeoffs in Long-Read Annotation

Platform Choice: Oxford Nanopore versus PacBio

The choice between Oxford Nanopore and PacBio platforms affects read length, base accuracy, throughput, and cost. The evaluation of sequencing platforms for evidence-driven genome annotation compared Oxford Nanopore R9.4.1 chemistry and PacBio Sequel II circular consensus sequencing. The study found that incorporating PacBio transcripts into the annotation pipeline significantly outperformed traditional methods, but this result does not establish that PacBio is always superior for annotation.

Oxford Nanopore offers the advantage of very long reads and real-time sequencing. The read length can exceed the length of most transcripts, providing full-length coverage. The base accuracy is lower than PacBio circular consensus sequencing, but the accuracy has improved with newer chemistry and basecalling algorithms.

PacBio circular consensus sequencing produces highly accurate reads by sequencing the same molecule multiple times and generating a consensus. The read length is typically shorter than Oxford Nanopore but still sufficient to span most transcripts. The higher base accuracy simplifies downstream analysis and reduces the need for error correction.

The choice between platforms should be based on the specific requirements of the annotation project. If the goal is to resolve complex isoform structures in genes with very long transcripts, Oxford Nanopore may be preferable. If the goal is to achieve the highest possible base accuracy for transcript models, PacBio circular consensus sequencing may be preferable.

Library Preparation: cDNA versus Direct RNA

Complementary DNA library preparation converts RNA to cDNA before sequencing. This process can introduce biases, but it also enables amplification and the use of standard sequencing protocols. Direct RNA sequencing avoids the conversion step and can detect RNA modifications, but it typically produces lower throughput and requires more input RNA.

For annotation purposes, cDNA libraries are the more common choice because they provide higher throughput and are compatible with standard analysis workflows. Direct RNA sequencing may be useful for specific applications, such as detecting RNA modifications that affect transcript structure.

Read Depth versus Read Length

The consortium's finding that longer and more accurate sequences produce more accurate transcripts than increased read depth has direct implications for experimental design. For annotation projects, the priority should be read length and accuracy. Increasing read depth will improve quantification accuracy, but it will not compensate for short reads that cannot span entire transcripts.

The practical implication is that researchers should allocate sequencing resources to maximize read length and accuracy instead of simply increasing the number of reads. This may mean using a library preparation method that preserves full-length transcripts, even if it produces lower throughput.

Reference-Based versus Reference-Free Assembly

Reference-based assembly aligns reads to the genome and uses the alignment structure to define transcript models. Reference-free assembly constructs transcripts from the reads themselves without a genome guide. The consortium found that reference-based approaches demonstrated the best performance in well-annotated genomes.

For annotation projects in organisms with an existing genome assembly, reference-based assembly is the appropriate choice. Reference-free assembly may be useful for detecting transcripts that are absent from the reference annotation, but it requires additional validation because the assembly process can introduce artifacts.

Observations and Measurements for Quality Control

Alignment Statistics

The alignment statistics provide the first indication of data quality. The percentage of reads that map to the genome should be high, typically above 80 percent for a well-prepared library. Low mapping rates may indicate contamination, library preparation problems, or genome assembly issues.

The distribution of mapping quality should be inspected. Reads with low mapping quality may align to multiple locations in the genome, which complicates transcript assembly. The distribution of alignment lengths should also be inspected. Reads that are much shorter than expected may indicate degradation during library preparation.

Transcript Model Statistics

The transcript models produced by the assembly step should be examined for structural features. The number of transcripts per gene, the number of exons per transcript, and the distribution of transcript lengths provide a basis for assessing the completeness of the transcript set.

The SQANTI classification provides a structured way to assess the quality of the transcript models. The proportion of transcripts classified as full splice match indicates how well the long-read data agrees with the reference annotation. The proportion of novel transcripts indicates the extent to which the long-read data reveals new biology.

Comparison Statistics

The GFFcompare comparison provides quantitative statistics on the agreement between the long-read transcript set and the reference annotation. The transcript-level agreement is typically lower than the loci-level agreement, as demonstrated by the Florida manatee study where loci-level agreement was 90 percent but transcript-level agreement was only 35 percent.

The comparison statistics should be interpreted in the context of the annotation objective. If the goal is to refine existing gene models, the focus should be on transcripts that are classified as full splice match or novel in catalog. If the goal is to discover novel transcripts, the focus should be on transcripts that are classified as novel not in catalog.

Records and Documentation for Reproducibility

Version Control for All Software and Reference Data

Reproducibility in genome annotation requires careful documentation of software versions, reference data versions, and analysis parameters. The nf-core documentation provides standards for community pipeline usage and configuration that can serve as a model for documenting analysis workflows.

The specific software versions used for alignment, transcript assembly, filtering, and gene prediction should be recorded. The reference genome version and the reference annotation version should also be recorded. Changes to any of these components can affect the final annotation.

Parameter Documentation

Analysis documentation must capture every methodological decision - tool versions, parameter choices, data preprocessing steps, and filtering thresholds - enabling independent reproduction of the reported findings.

The Galaxy Training Network (GTN) offers community-curated, browser-based pipelines that promote FAIR and reproducible bioinformatics analyses.

Output File Management

The output files from each analysis step should be organized and named consistently. The raw sequencing data, the aligned reads, the assembled transcripts, the filtered transcripts, and the final gene models should be stored in a structured directory layout.

Bioconductor's vignette-driven documentation model ensures that every R package ships with a reproducible analysis workflow demonstrating real data application.

Common Failure Patterns and How to Address Them

Failure Pattern 1: Low Mapping Rates

Low mapping rates indicate that a substantial fraction of the reads do not align to the genome. This can result from contamination, library preparation problems, or genome assembly issues. The first step is to check the read content for adapter sequences and contamination. The second step is to check the genome assembly for completeness and contiguity.

If contamination is the cause, the contaminated reads should be removed before alignment. If the genome assembly is the cause, additional genome assembly work may be needed before transcript annotation can proceed.

Failure Pattern 2: Fragmented Transcript Models

Fragmented transcript models result from reads that do not span the full transcript. This can occur when the library preparation method does not preserve full-length transcripts or when the sequencing platform produces short reads. The solution is to adjust the library preparation or sequencing strategy to produce longer reads.

The consortium's finding that longer and more accurate sequences produce more accurate transcripts supports the priority on read length. If the current data produces fragmented transcript models, consider whether the library preparation method is preserving full-length transcripts.

Failure Pattern 3: Excessive Novel Transcripts

An excessive number of novel transcripts may indicate that the filtering step is not stringent enough. SQANTI applies multiple filters to remove artifacts, but the filter thresholds may need to be adjusted based on the data quality.

The novel transcripts should be examined for structural features that distinguish genuine isoforms from artifacts. Transcripts with non-canonical splice sites, very low read support, or incomplete open reading frames are more likely to be artifacts.

Failure Pattern 4: Low Transcript-Level Agreement with Reference Annotation

Low transcript-level agreement with the reference annotation can be expected when the reference annotation was built from short-read data. The Florida manatee study found only 35 percent transcript-level agreement between long-read-based and short-read-based annotation. This does not necessarily indicate a problem with the long-read data.

The transcript-level disagreement reflects the ability of long-read data to resolve isoform structures that short-read data cannot distinguish. The loci-level agreement provides a more appropriate measure of overall concordance.

Failure Pattern 5: Poor Gene Prediction Performance

Poor gene prediction performance can result from insufficient training data, inappropriate parameters, or a mismatch between the evidence and the prediction algorithm. The BRAKER3 pipeline addresses this by using statistical models learned iteratively for the target genome.

If the gene prediction performance is poor, consider whether the transcript evidence has been filtered appropriately. Transcript models that are artifacts will confuse the gene prediction algorithm. Consider also whether the protein database used for evidence is appropriate for the target species.

Limitations of Long-Read RNA-Seq for Annotation

Detection of Lowly Expressed Transcripts

Long-read RNA-seq has lower throughput than short-read RNA-seq, which limits the detection of lowly expressed transcripts. The consortium found that incorporating orthogonal data and replicate samples is advised when aiming to detect rare and novel transcripts. This means that a single long-read RNA-seq library may not be sufficient for comprehensive annotation.

The practical implication is that long-read data should be combined with short-read data and protein evidence to achieve comprehensive annotation. The short-read data provides quantitative information that can guide the interpretation of the long-read transcript models.

Error Profiles and Their Impact on Transcript Models

Long-read sequencing errors can introduce artifacts into transcript models. The error profile differs between Oxford Nanopore and PacBio platforms. PacBio circular consensus sequencing produces highly accurate reads, while Oxford Nanopore reads have a higher error rate.

The impact of sequencing errors on transcript models depends on the analysis approach. Reference-based approaches that align reads to the genome can tolerate some sequencing errors because the alignment provides context. Reference-free approaches are more sensitive to sequencing errors because the assembly process must distinguish true variation from sequencing error.

Incomplete Annotation of Untranslated Regions

Long-read RNA-seq can identify transcript boundaries, but the annotation of untranslated regions may be incomplete. The 5-prime ends of transcripts are often difficult to capture because of the library preparation process. The 3-prime ends are typically better represented because of the poly-A tail.

The incomplete annotation of untranslated regions affects the interpretation of transcript models. A transcript that appears to lack a complete open reading frame may actually have an incomplete 5-prime untranslated region instead of a genuine non-coding transcript.

Safety and Regulatory Context for Annotation Data

Data Sharing and Database Submission

Genome annotation data is typically shared through public databases. The NCBI provides databases and search systems for sequence data, and the submission of annotation data follows established protocols. The EMBL-EBI Training program provides guidance on data submission and the use of bioinformatics data resources.

The decision to submit annotation data to a public database should be made early in the project. The submission format and the required metadata should be determined before the analysis is complete to avoid rework.

Ethical Considerations for Human Data

If the long-read RNA-seq data comes from human samples, ethical considerations apply. The GTEx dataset used in the human long-read RNA-seq study was generated from donated tissues with appropriate consent. Researchers working with human data must ensure that their data collection and analysis comply with ethical standards and data protection regulations.

The allele-specific analysis of long reads can reveal genetic variants that have implications for the research participants. The handling of this information requires careful consideration of privacy and confidentiality.

Professional Escalation Criteria

When to Seek Specialized Support

Genome annotation projects can encounter problems that require specialized support. The following situations warrant escalation to a bioinformatics core facility, a computational biology collaborator, or a specialized service provider.

If the genome assembly is highly fragmented and the fragmentation prevents accurate transcript annotation, seek support from a genome assembly specialist. The transcript annotation cannot be better than the genome assembly that supports it.

If the long-read RNA-seq data quality is poor and the cause is not apparent, seek support from a sequencing facility or a library preparation specialist. The problem may be in the library preparation protocol, the sequencing run, or the basecalling process.

If the gene prediction pipeline produces results that are inconsistent with the transcript evidence, seek support from a gene prediction specialist. The integration of multiple evidence types into gene prediction requires careful parameter tuning and validation.

When to Reconsider the Experimental Design

The decision to generate additional data should be based on the quality of the existing data and the annotation objectives. If the existing long-read data produces fragmented transcript models, generating more data with the same approach will not solve the problem. The experimental design should be revised to prioritize read length and accuracy.

If the goal is to detect rare and novel transcripts and the existing data does not provide sufficient depth, additional libraries and replicate samples may be needed. The consortium's recommendation to incorporate orthogonal data and replicate samples for rare transcript detection should guide this decision.

Frequently Asked Questions

What is the difference between using long-read RNA-seq for annotation versus quantification?

Long-read RNA-seq for annotation focuses on determining the structure of transcripts, including exon boundaries, splice junctions, and isoform composition. Long-read RNA-seq for quantification focuses on measuring the abundance of transcripts. The consortium found that libraries with longer and more accurate sequences produce more accurate transcripts, while greater read depth improves quantification accuracy. For annotation projects, the priority should be read length and accuracy. For quantification projects, the priority should be read depth.

How much long-read RNA-seq data is needed for genome annotation?

The amount of data needed depends on the genome complexity, the expression levels of the genes of interest, and the annotation objectives. The consortium generated over 427 million long-read sequences for their assessment, but this scale is not required for every project. The key consideration is whether the data provides sufficient coverage of the transcripts of interest. For rare transcripts, additional depth and replicate samples are needed. The consortium advises incorporating orthogonal data and replicate samples when aiming to detect rare and novel transcripts.

Should I use Oxford Nanopore or PacBio for genome annotation?

The choice between Oxford Nanopore and PacBio depends on the specific requirements of the annotation project. The evaluation of sequencing platforms found that incorporating PacBio transcripts into the annotation pipeline significantly outperformed traditional methods. However, this result does not establish that PacBio is always superior. Oxford Nanopore offers longer reads, while PacBio circular consensus sequencing offers higher base accuracy. The decision should be based on the read length and accuracy requirements of the project.

What is the role of SQANTI in the annotation workflow?

SQANTI filters and classifies long-read transcript models. It assigns each transcript to a structural category based on its relationship to the reference annotation. The classification provides a basis for deciding which transcripts to retain for annotation. Transcripts classified as full splice match confirm existing annotation models. Transcripts classified as novel in catalog provide evidence for alternative isoforms. Transcripts classified as novel not in catalog require additional validation.

How do I interpret the low transcript-level agreement between long-read and short-read annotation?

Low transcript-level agreement between long-read and short-read annotation is expected and reflects the ability of long-read data to resolve isoform structures that short-read data cannot distinguish. The Florida manatee study found 90 percent loci-level agreement but only 35 percent transcript-level agreement. The loci-level agreement provides a more appropriate measure of overall concordance. The transcript-level disagreement indicates that the long-read data provides additional isoform resolution.

Can I use long-read RNA-seq data to annotate a newly assembled genome?

Yes, long-read RNA-seq data can be used to annotate a newly assembled genome. The integration of long-read transcript evidence with ab initio prediction and protein evidence produces better gene models than any single evidence source. The BRAKER3 pipeline provides an automated approach that combines short-read RNA-seq and a large protein database with statistical models learned iteratively for the target genome. The pipeline outperformed other existing tools in benchmarking studies.

What should I do if my long-read data produces fragmented transcript models?

Fragmented transcript models result from reads that do not span the full transcript. The solution is to adjust the library preparation or sequencing strategy to produce longer reads. The consortium found that libraries with longer and more accurate sequences produce more accurate transcripts. If the current data produces fragmented transcript models, consider whether the library preparation method is preserving full-length transcripts.

How do I validate novel transcripts identified by long-read RNA-seq?

Novel transcripts should be validated using multiple lines of evidence. Protein evidence confirms that a transcript is translated. Homology to known transcripts in other species provides support for the biological validity of a transcript. The Florida manatee study found that 64 percent of novel isoforms matched known sequences in other mammalian species. Transcripts that lack protein evidence and homology should be treated with caution.

Related Bioinformatics Guides

• RNA-Seq Data Analysis Workflow: From Raw Reads to Insights
• Long-Read Sequencing for De Novo Assembly of Complex Genomes: Case Studies and Best Practices
• RNA-Seq Databases: Accessing and Using Public RNA-Seq Data
• RNA-Seq vs ChIP-Seq: Complementary Approaches for Gene Regulation
• De Novo Genome Assembly with Long Reads: A Practical Workflow

References and Further Reading

• NCBI Data Resources. National Center for Biotechnology Information.
• EMBL-EBI Training. European Bioinformatics Institute.
• Bioconductor. Bioconductor Project.
• Galaxy Training Network. Galaxy Project.
• nf-core Documentation. nf-core.
• The Carpentries Lessons. The Carpentries.
• BRAKER3: Fully automated genome annotation using RNA-seq and protein evidence with GeneMark-ETP, AUGUSTUS and TSEBRA.. bioRxiv : the preprint server for biology, 2024.
• Systematic assessment of long-read RNA-seq methods for transcript identification and quantification.. Nature methods, 2024.
• BRAKER3: Fully automated genome annotation using RNA-seq and protein evidence with GeneMark-ETP, AUGUSTUS, and TSEBRA.. Genome research, 2024.
• Evaluation of strategies for evidence-driven genome annotation using long-read RNA-seq.. Genome research, 2025.
• Transcriptome variation in human tissues revealed by long-read sequencing.. Nature, 2022.

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