# Gene Annotation Errors in RNA-seq The most common cause of a suspicious RNA-seq result is not your library, your sequencer, or your statistical model. It is a mismatch between the genome FASTA you aligned against and the GTF or GFF3 file you counted reads with. A close second is a set of false splice junctions generated by the aligner itself, some of which have quietly made their way into widely used annotation databases [1]. Both problems produce the same symptom, plausible-looking numbers attached to the wrong genes, and both are cheap to check for. The first thing to do when results look wrong is boring but effective. Confirm that your genome FASTA, your annotation file, and your transcriptome reference, if you use one, come from the same assembly release and the same annotation pipeline. Then look at the twenty most highly expressed genes in your count matrix. If a small handful of models sitting in ribosomal DNA regions dominate your transcripts per million values, you are almost certainly looking at misannotated gene models rather than biology [2]. This guide walks through gene annotation errors in RNA-seq from the most frequent cause to the rarest, explains the mechanism behind each failure, gives you a diagnostic sequence you can run on real data, and ends with prevention practices that hold up across projects. ## What Gene Annotation Errors in RNA-seq Actually Are An annotation, in the practical sense most analysts use, is a GTF or GFF3 file. Each line describes a genomic feature, typically an exon, a coding sequence, or a transcript, with coordinates on a specific genome assembly, a feature type, a strand, and attributes such as gene_id and transcript_id. Your aligner, quantifier, and differential expression tool all inherit the assumptions packed into that file. When the file is wrong, every number downstream inherits the error. A gene model is the set of coordinates describing where a gene starts, where it ends, which pieces are exons, and how those exons are spliced together. Gene annotation errors are mistakes in those coordinates or in the identity assigned to the resulting sequence. The error types that matter for RNA-seq fall into a few recurring categories. - **Version and coordinate mismatches.** The annotation describes one assembly, the FASTA describes another. Chromosome names differ, coordinates are shifted, or genes exist in one file but not the other. - **False gene models in repetitive or ribosomal DNA regions.** Automated pipelines predict messenger RNA gene models inside rDNA arrays, where the actual genes encode rRNA, not protein [2]. - **Aligner-induced phantom splicing.** Splice-aware aligners such as STAR and HISAT2 can produce spliced alignments between repeated sequences that were never spliced in reality [1]. - **Artificial gene fusions.** Annotators merge two adjacent or overlapping genes into one model, creating a fusion gene that does not exist [3][4][5]. - **Homology transfer errors.** Functional labels are copied from a related species onto the wrong gene, a common strategy in poorly annotated organisms [6]. - **Assembly errors.** The underlying genome assembly is misjoined or misassembled, so no annotation can be fully correct on top of it [7][8][5]. Every one of these categories has a different signature in the data, which is why a structured diagnostic sequence beats ad hoc suspicion. ## Why Annotation Errors Corrupt Every Downstream Result Read counting is the first casualty. When a false gene model sits in an rDNA region, the enormous pool of rRNA-derived reads has somewhere legal to go. In plant assemblies examined with riboCleaner, a handful of genes appearing to contribute to more than 50 percent of the total transcripts per million values of entire RNA-seq datasets traced back to false mRNA gene models predicted within rDNA regions [2]. Any normalization scheme that divides by total counts then undercounts every genuine transcript in the sample. Differential expression is the second casualty. Most RNA-seq analysis tools only consider read abundance along annotated features, and annotations are typically incomplete. DiffSegR was built specifically because incomplete annotations lead to errors in differential expression analysis, using changepoint detection on per-base log2 fold change profiles to find differentially expressed regions without requiring prior annotation at all [9]. Genes that are missing, truncated, or fused in the annotation either disappear from the analysis or get their signal blended with a neighbor. Allele-specific and variant analyses are the third casualty. Variant calling from RNA-seq is already complicated by [alternative splicing](/blog/guides/alternative-splicing), allele-specific expression, RNA editing, and variable transcript abundance [10]. Layer an ambiguous or wrongly annotated reference on top, and mapping bias distorts allele counts. In case studies across four species, from Drosophila melanogaster to a highly heterozygous orangutan, inaccurate allele and [gene expression](/blog/guides/gene-expression) counts due to map bias and genome ambiguity produced high false positive and false negative rates in studies of allelic imbalance [7]. Functional interpretation is the final casualty. Even a perfectly counted, correctly tested gene list falls apart at the enrichment step if the gene labels are wrong. In non-model organisms, researchers often transfer functional annotation from related species using homology searches, and this strategy often introduces annotation errors while overlooking species-specific functions [6]. The result is a pathway story built on mislabeled genes. ## Causes of Gene Annotation Errors in RNA-seq, From Most to Least Common ### 1. Genome and annotation version mismatches This is the failure mode that hits nearly every analyst at some point. You download genome GRCh38 from one mirror and a GTF from another, or you rebuild an index two years later and pull a newer annotation release. Chromosome naming conventions drift between sources, so chr1 in the FASTA becomes 1 in the GTF, and every gene silently fails to match. Coordinate shifts between assembly versions are worse because alignment still succeeds, counting still succeeds, and the error is invisible until you notice that your gene coordinates make no sense. The fix begins with the boring discipline of recording the exact release for every file, but the diagnostic sequence below also gives you commands that surface mismatches in minutes. ### 2. Phantom introns from splice-aware aligners STAR and HISAT2 are excellent aligners, and both can introduce erroneous spliced alignments between repeated sequences. When read pairs span two copies of a repeat, the aligner can interpret the gap as an intron, producing a spliced alignment for a transcript that was never spliced. The authors of EASTR showed that these phantom introns can make their way into widely used genome annotation databases, meaning an error created during one project's transcript assembly can become reference truth for hundreds of later projects [1]. The mechanism is simple. Intron-flanking sequences in repeats look like splice signal-adjacent sequence similarity, and the aligner's junction discovery step has no way to distinguish a true intron from a repeat-mediated gap without evidence beyond the junction itself. EASTR addresses this by detecting sequence similarity between intron-flanking regions and removing the falsely spliced alignments or transcripts, improving accuracy across species as different as human, maize, and Arabidopsis thaliana [1]. ### 3. False gene models in rDNA and other repetitive regions Ribosomal DNA arrays contain hundreds of tandem copies of rRNA genes. Automated gene prediction pipelines, hunting for open reading frames, occasionally predict mRNA gene models inside these arrays. Because rRNA is transcribed at extraordinary levels, any sequencing read that maps into an rDNA-region gene model pulls a massive weight of reads with it. This is how a handful of models came to account for more than 50 percent of total TPM in entire plant RNA-seq datasets [2]. The same class of problem applies to any highly repetitive, transcribed locus, including transposon-rich regions where predicted models may be fragments rather than genes. ### 4. Artificial gene fusions Fusion artifacts appear when the annotation process joins two genuinely separate genes into one model. The helminth [genome analysis](/blog/guides/genome-analysis) is the clearest demonstration. Of 20,313 two-domain proteins annotated in current helminth genomes, 10,297 were apparently unique to helminths, and examination with RNA-seq data from 20 species suggested that the majority of helminth-specific fusion genes are gene prediction artifacts [3]. The evidence pattern is worth memorizing because it transfers to any organism. Artifactual fusions showed a lack of correlation between expression of the first and second fused domains, significantly longer interdomain regions, reduced continuity of RNA-seq coverage across the interdomain region consistent with coverage breakpoints, and little support in de novo transcriptome assemblies [3]. A plant parallel exists in the chalcone synthase gene family. Database searches found apparent triplication of the CHS domain-encoding region in gene models from Macadamia integrifolia, Musa balbisiana, Musa troglodytarum, and Nymphaea colorata, and manual inspection against large RNA-seq datasets indicated these models result from artificial fusions made during the annotation process [4]. The lesson generalizes. Any gene model encoding an unusual repeated domain architecture deserves a coverage check before you believe it. ### 5. Overlapping genes and polycistronic transcripts merged by mistake Manual curation of thousands of gene models in the nematode Pristionchus pacificus identified and corrected more than 7,500 gene models, roughly 24 percent of the annotation, after incorporating new Iso-seq and RNA-seq data. The recurring error sources were assembly errors, artificial transcript fusions resulting from overlapping genes and polycistronic RNAs, falsely called open reading frames, and error propagation based on homology data [5]. Overlapping genes are a special trap for automated pipelines. When two genes overlap on opposite strands or in different frames, a predictor can collapse them into one transcript, and every read from both real genes is then attributed to the wrong single model. ### 6. Homology-based annotation transfer in non-model organisms For organisms without curated functional annotation, BLAST-based transfer from a close relative is standard practice, and it is also a documented source of annotation errors [6]. The transferred label is only as good as the orthology assignment behind it, and orthology assignment across diverged genomes is genuinely hard. A comparative annotation effort across 301 species of Drosophilidae, generated with the Comparative Annotation Toolkit and BRAKER3 using RNA-seq and protein evidence, found that species-specific factors, including assembly error, contribute substantially to observed differences in gene number and coding sequence length [8]. In other words, some of what looks like biological variation between species is annotation noise. ### 7. Underlying assembly errors No annotation can be better than the assembly beneath it. Misassemblies create gene fragments, split genes, and chimeric scaffolds, and the annotator faithfully converts those physical errors into gene model errors. The Pristionchus curation effort listed assembly errors among its most frequent problems [5], and the Drosophilidae comparison reached the same conclusion at phylogenetic scale [8]. When apparent extreme allele bias shows up in an RNA-seq dataset, one recommended interpretation is exactly this, errors in the genome assembly or annotation rather than true biology [7]. ### 8. Long-read artifacts and the rare-transcript problem [Long-read sequencing](/knowledge/bioinformatics/long-read-sequencing-technologies-pacbio-and-oxford-nanopore) is the best tool for fixing annotation, but it introduces its own error class. Tens of thousands of novel transcripts have been reported even in well-annotated genomes using long reads, and the community benchmarking effort organized around the Long-read RNA-seq Genome Annotation Assessment Project validated many novel, lowly expressed, often sample-specific transcripts [11]. The open challenge is separating technology errors from rare but real molecules. Long-read accuracy is affected by sequencing errors, [mRNA degradation](/knowledge/molecular-biology/mrna-degradation), and incomplete cDNA synthesis [12], and downstream annotation pipelines must decide which of those novel structures deserve reference status. ## A Diagnostic Sequence for Bioinformatics Troubleshooting Work through these steps in order. Each one either clears a common cause or hands you evidence for the next step. **Step 1. Confirm the version triangle.** Check that your genome FASTA, annotation GTF, and any transcriptome index come from the same release. Read the release notes on the download page, record the assembly accession and annotation release in your project notes, and never mix sources for the same assembly. A single document in your project directory listing file names, download URLs, download dates, and release versions saves more debugging hours than any other habit in this field. **Step 2. Check chromosome name consistency.** ``` cut -f1 genome.fa.fai | sort -u > fasta_chrs.txt cut -f1 annotation.gtf | sort -u > gtf_chrs.txt comm -23 gtf_chrs.txt fasta_chrs.txt ``` If the second command prints anything, the annotation contains sequences absent from your FASTA. Do the reverse comparison as well. Mismatches here mean every count for those sequences is zero regardless of expression. **Step 3. Inspect the alignment summary.** Read the STAR Log.final.out or the HISAT2 summary. Look at the multimapping rate and the number of splices per junction. A multimapping rate that is high for your tissue and library type, or a small number of junctions absorbing a huge fraction of spliced reads, points at repetitive regions and phantom junctions [1]. **Step 4. Count reads and rank the most expressed genes.** ``` featureCounts -a annotation.gtf -t exon -g gene_id -o counts.txt aligned.bam ``` Then rank genes by counts or TPM and inspect the top twenty. Cross-check each one against the annotation. If several sit inside rDNA regions or carry rRNA-like names, you have found the riboCleaner failure mode, and a handful of models may be absorbing a large share of your library [2]. **Step 5. Screen for phantom introns before transcript assembly.** Run EASTR on your alignments and on your reference annotation before any assembly or counting step that trusts splice junctions. EASTR detects sequence similarity between intron-flanking regions and removes falsely spliced alignments or transcripts, and applying it before transcript assembly substantially reduces false positive introns, exons, and transcripts [1]. If you cannot run EASTR, a rough manual screen is to list junctions supported by few reads whose flanking intron sequences are highly similar to each other. **Step 6. Test candidate fusion models for coverage continuity.** For any gene model with an unusual domain architecture, extract the per-base coverage across the model from your BAM and look for a coverage breakpoint in the interdomain region. Genuine fusion genes show correlated expression between their fused domains and continuous coverage. Artifacts show uncorrelated domain expression, long interdomain regions, broken coverage, and weak support in de novo transcriptome assemblies [3]. A quick corollary screen is to assemble the transcriptome de novo and check whether any assembled transcript spans the putative fusion. **Step 7. Compare annotation versions directly.** When an annotation update is available, run gffcompare between your old and new GTF files to see how many models changed, merged, or vanished. A large change set is both a warning about your historical results and a reason to reprocess. If a gene central to your hypothesis changed between releases, read the curation notes before trusting either version. **Step 8. Evaluate extreme allele ratios as annotation signals.** If you are doing allele-specific work and see apparently extreme allele bias at a locus, treat it as a candidate annotation or assembly error first. The recommended workflow is to map to a personalized genome to increase allele assignments, track multimapping reads while tuning mapping parameters, and then evaluate apparent extreme allele bias specifically to identify errors in [genome assembly and annotation](/knowledge/molecular-biology/genome-assembly-and-annotation) [7]. ## Troubleshooting Table | Symptom | Likely cause | How to confirm | Fix | |--|--|--|--| | A handful of genes dominate total TPM | False mRNA gene models inside rDNA regions | Map top gene coordinates to rDNA loci, check for rRNA similarity | Remove or flag rDNA-region models, rerun normalization, quantify rRNA contamination with a tool such as riboCleaner [2] | | Many low-support novel junctions after assembly | Aligner phantom introns from repeats | EASTR screen for similarity between intron-flanking regions [1] | Run EASTR on alignments and annotation, then reassemble | | Gene coordinates make no sense or counts are zero | Version mismatch or chromosome naming mismatch | Chromosome set comparison with comm, release notes review | Align and count against a matched FASTA and GTF pair | | A "novel" gene looks like two genes stapled together | Artificial fusion in annotation | Uncorrelated domain expression, coverage breakpoint, poor de novo transcript support [3] | Split the model or exclude it, report the artifact to the database maintainers | | Many genes with unexpected domain repeats | Artifact from annotation fusion, as seen in CHS models | Manual inspection against large RNA-seq data [4] | Require RNA-seq support before trusting repeated-domain models | | Functional enrichment makes no biological sense | Homology transfer errors in a non-model organism | Check BLAST top hits, percent identity, and orthology evidence [6] | Build a custom annotation package from curated NCBI data compatible with clusterProfiler [6] | | Allele ratios near fixation at many loci | Map bias, assembly or annotation error | Personalized genome mapping, multimapping audit [7] | Remap to a personalized genome, tune parameters, re-evaluate | | DE signal at loci absent from the GTF | Incomplete annotation | Changepoint detection on per-base log2 fold change [9] | Use an annotation-free differential approach such as DiffSegR as a complementary screen [9] | | Novel transcript count explodes with long reads | Mix of real rare transcripts and technology errors | Benchmark against validated LR-GASP style datasets, check junction mismatch profiles [12][11] | Filter using splice junction mismatch profiles and stratified confidence criteria [12] | ## How to Fix RNA-seq Gene Annotation Errors ### Fix the reference pair first If your FASTA and GTF do not match, nothing else you do matters. Rebuild your STAR or HISAT2 index with the matched genome and annotation, regenerate the transcriptome index if you use a pseudoaligner, and rerun alignment and counting. This is the fix that resolves the majority of otherwise mysterious rna seq error reports, and it costs an afternoon. Then archive the exact file versions so the fix is permanent. ### Clean the alignment before you assemble Run EASTR on your alignments and your reference annotation before transcript assembly. Its application to reference annotation databases can detect and correct likely cases of mis-annotated transcripts, which means the same tool that protects your own analysis can also flag inherited errors in the reference itself [1]. The order matters. Emending alignments after assembly is far less useful, because the assembly has already committed the phantom junctions into transcript models. ### Handle rDNA and repetitive-region contamination explicitly Quantify rRNA-derived reads rather than guessing. riboCleaner is a containerized Snakemake workflow, with a prebuilt Docker image, designed to identify misannotated gene models in rDNA regions and quantify rRNA-derived reads in RNA-seq data, and it has been applied across several plant genome assemblies [2]. Even if you work outside plants, the principle transfers. Identify the rDNA coordinates in your organism, exclude or separate those models, and report rRNA fraction as a QC metric alongside alignment rate. For multimapping in repetitive regions generally, the allelic imbalance work offers a template that applies broadly. Track multimapping reads, tune mapping parameters, and make the treatment of ambiguous reads an explicit, documented decision rather than an aligner default [7]. ### Attack suspected fusions with four checks When a gene model looks like an artificial fusion, apply the helminth study's evidence framework. Check whether the two domains show correlated expression. Check the interdomain length. Check whether RNA-seq coverage is continuous across the junction or breaks there. Check whether de novo transcript assembly recovers a transcript spanning the model [3]. A model failing most of these checks should be split in your working annotation, and you should consider reporting it, since these artifacts have propagated into widely used sequence and protein databases including WormBase ParaSite and InterPro [3]. ### Improve the annotation itself for non-model organisms If you work on a poorly annotated organism, waiting for the reference database to improve is not a strategy. Build your own annotation resources. One adaptable approach creates custom annotation R packages from genomic data retrieved from NCBI, importable directly into R and compatible with clusterProfiler, demonstrated on Mycobacterium tuberculosis H37Rv and used to analyze differential expression from a rifampicin treatment dataset [6]. On the structural side, comparative annotation pipelines such as the Comparative Annotation Toolkit and BRAKER3, incorporating RNA-seq and protein evidence across related species, produce more consistent annotations than single-species pipelines [8]. If assembly is the bottleneck, consider avoiding it entirely for expression work. An assembly-free strategy aligns RNA-seq reads directly to a high-confidence [protein database](/blog/guides/protein-database), cutting runtime from the order of tens of hours to under half an hour and memory from tens of gigabytes to tens of megabytes in their benchmarks, while avoiding assembly errors that impact expression analysis [13]. This will not give you transcript structures, but for differential expression in a non-model organism it sidesteps an entire class of annotation and assembly failure. ### Use long reads to validate, then curate by hand where it counts Long-read RNA-seq is the strongest single investment for annotation quality. The Arabidopsis AtRTD3 effort built a reference transcript dataset of over 169,000 transcripts, twice the size of the best existing Arabidopsis transcriptome and including over 1,500 novel genes, with 78 percent of transcripts from Iso-seq carrying accurately defined splice junctions and transcription start and end sites. Their key trick was using mismatch profiles around splice junctions to distinguish correct junctions and remove false ones, a method that directly addresses the junction-error problem [12]. But long reads are not self-validating. The field's central unresolved problem is differentiating technology errors from rare but real molecules, and community benchmarking through the Long-read RNA-seq Genome Annotation Assessment Project is part of how standards will emerge [11]. In the meantime, manual curation of the models that matter for your project remains the gold standard. The Pristionchus community curation corrected more than 7,500 gene models, about 24 percent, by combining Iso-seq, RNA-seq, and genome-wide screening, and documented exactly which error classes to hunt for [5]. ### Rethink quantification where annotation is unreliable For differential expression in poorly annotated systems, an annotation-free option exists on the short-read side as well. DiffSegR identifies transcriptome-wide expression differences using multiple changepoint detection on per-base log2 fold change, requiring no prior annotation, and in a plant dataset correctly predicted the roles of chloroplast ribonuclease Mini-III in rRNA maturation and PNPase in precursor degradation [9]. Running an annotation-free analysis alongside your gene-level analysis is a cheap insurance policy. If both agree, your annotation is probably adequate. If they disagree at specific loci, you have found candidate annotation problems. For toxin and other specialized transcriptomes, modular pipelines that pair de novo assembly with curated protein homology and structural filtering show how to minimize misassemblies and annotation bias by design. ToxIR, validated in scorpion venom glands, combines rnaSPAdes assembly, curated UniProt homology searches, and structural analyses with SignalP, TMHMM, Pfam, and InterProScan to prioritize candidates, identifying 378 putative toxin candidates of which 192 were high-confidence [14]. The design lesson, matched evidence requirements at each step, is portable to any specialized transcriptome. ### For isoform work, decouple detection from annotation Annotation databases remain far from comprehensive, especially for condition-specific alternative splicing such as the cancer-specific splicing landscape. Freddie takes long-read sequencing plus its genomic alignment and computes isoforms annotation-independently, partitioning reads into independently processable sets and representing isoforms as subsets of canonical exon segments [15]. If your question is about isoforms rather than annotated genes, an annotation-independent detector plus a validation step against a curated annotation beats trusting the annotation alone. ## Common Mistakes and Limitations **Treating the annotation as ground truth.** The single most damaging habit in RNA-seq analysis is assuming the GTF is correct because it is official. Phantom introns have entered widely used databases [1], rDNA false models distort TPM totals [2], and a quarter of gene models in an actively maintained nematode genome needed correction [5]. Annotations are estimates, and version-to-version changes are evidence, not noise. **Filtering low-count genes before checking where they are.** Standard DE pipelines drop low-count genes early. If a real gene sits in a misannotated or merged model, its reads are misassigned, the model's count may survive, and the true gene vanishes silently. Check the annotation provenance of your DE hits, not just their statistics. **Trusting TPM rankings without checking composition.** When a few models absorb a huge share of reads, every normalization that divides by library total is distorted. Before interpreting any TPM-based comparison, verify that the top of your expression distribution looks biologically plausible rather than dominated by rDNA-region artifacts [2]. **Assuming multimapping is a solved problem.** Default aligner parameters make silent decisions about multimapping reads. In allele-aware and repeat-rich contexts, those decisions drive map bias [7]. Make the decision explicit, document it, and check its effect. **Confusing cell-type annotation errors with gene annotation errors.** In single-cell work, the phrase annotation error often refers to mislabeled cell types, a different problem with a different toolkit, addressed by methods that tolerate noisy reference labels [16] or identify novel cell types absent from training data [17]. If your collaborator says annotation error, confirm which annotation they mean before debugging. **Over-correcting after a fix.** When you switch annotation versions, results change, and the temptation is to cherry-pick the version that supports your hypothesis. Decide your annotation version before unblinding your differential results, and report the version in your methods. **Expecting long reads to be error-free.** Long-read data carries sequencing errors, degradation artifacts, and incomplete cDNA synthesis [12], and the community is still building methods to separate technology errors from rare real transcripts [11]. Use long reads as powerful evidence, not as oracles. One honest limitation of this entire workflow. Automated checks catch systematic problems, but borderline gene models in your organism of interest may need judgment that no pipeline provides, and if your project hinges on a specific gene model, verify it manually with RNA-seq coverage, long reads, and comparative evidence. ## Prevention Practices That Hold Up Across Projects **Pin your reference files.** Record assembly accession, annotation release, download URL, and download date for every reference file, in a file committed to your project repository. Re-verify checksums when you move between machines. **Never mix assembly and annotation sources.** If you must combine them, convert chromosome names, verify coordinate consistency on known genes, and document every transformation. **Emend alignments before assembling.** Make EASTR or an equivalent junction-filtering step a standard stage in your pipeline, applied to both your alignments and your reference annotation [1]. **Add rRNA fraction to your QC panel.** Quantify rRNA-derived reads and check the top-expressed gene list as routine QC, the same way you check mapping rate [2]. **Require RNA-seq evidence for surprising gene models.** Domain architectures that look like fusions or repeats need coverage continuity, correlated domain expression, and assembly support before you cite them [3][4]. **Reprocess together with annotation updates.** When a new annotation release lands, rerun gffcompare against your old version, quantify the change, and decide deliberately whether to reprocess historical datasets for comparability. **Validate with long reads where stakes are high.** For genes or isoforms central to a manuscript, long-read evidence with junction mismatch profiling [12] and, where feasible, manual curation [5] is the standard you should hold your key models to. **Report annotation versions in every output.** Methods sections, supplementary tables, and count matrices should all carry the annotation release string. Future you, and every reviewer, will want to know exactly which gene models produced the numbers. ## Frequently Asked Questions ### What are gene annotation errors in RNA-seq? They are mistakes in the reference gene models, coordinates, or functional labels used to interpret RNA-seq data, including wrong exon boundaries, false genes in repetitive regions, artificial fusions, and mislabeled functions. Because every downstream tool reads the annotation, these errors propagate into counts, differential expression, and interpretation. ### What are the most common causes of RNA-seq gene annotation errors? The most common causes are genome and annotation version mismatches, aligner-generated phantom introns in repetitive sequences, false gene models in rDNA regions, artificial gene fusions, homology-based label transfer in poorly annotated organisms, and errors in the underlying genome assembly. Community curation efforts found assembly errors, fusions from overlapping genes, false open reading frames, and homology error propagation to be the recurring classes. ### How do I know whether my RNA-seq results are affected? Check whether your FASTA and GTF match, inspect your alignment summary for high multimapping, rank your most highly expressed genes and look for rDNA or rRNA signatures, and test any surprising gene model for coverage continuity across putative fusion junctions. A handful of models absorbing a large share of total TPM is a classic warning sign. ### How do I fix RNA-seq gene annotation errors? Start by realigning against a matched genome and annotation pair, run EASTR on alignments and annotation before assembly to remove phantom junctions, quantify and exclude rRNA contamination, and split suspected fusion models that fail coverage and expression-correlation checks. For deeper problems, rebuild the annotation with long-read data or comparative pipelines, and curate key models by hand. ### Do I need to redo my entire analysis after an annotation update? Usually yes for any step that touches the annotation, including index building, alignment, counting, and differential expression. Run gffcompare between the old and new versions first so you understand the scale of the change, and report both versions if you need to bridge historical results. ### Are non-model organisms more prone to these problems? Yes. Curated annotations are scarce, so homology-based transfer from related species is common and introduces errors, assemblies are more likely to contain misassemblies, and fusion artifacts are more likely to survive into databases. Custom annotation resources built from curated NCBI data and comparative annotation pipelines reduce these problems substantially. ### Can the aligner itself create annotation errors? Yes. STAR and HISAT2 can produce erroneous spliced alignments between repeated sequences, creating phantom introns that have entered widely used annotation databases. Filtering with EASTR before transcript assembly substantially reduces these false junctions, exons, and transcripts. ### Is long-read sequencing worth adding just to improve annotation? If annotation quality limits your project, yes. Long reads capture full-length transcript structures, and Iso-seq-based approaches have used splice junction mismatch profiles to build much larger and more accurate reference transcriptomes. Keep in mind that long-read data has its own error sources, so validated, confidence-stratified integration is the goal rather than wholesale acceptance of every novel transcript. ## Sources 1. [EASTR: Identifying and eliminating systematic alignment errors in multi-exon genes.](https://pubmed.ncbi.nlm.nih.gov/37940654/) 2. [riboCleaner: a pipeline to identify and quantify rRNA read contamination from RNA-seq data in plants.](https://pubmed.ncbi.nlm.nih.gov/35731209/) 3. [Widespread gene fusion artifacts in helminth genome annotations.](https://pubmed.ncbi.nlm.nih.gov/41634541/) 4. [Identification of annotation artifacts concerning the chalcone synthase (CHS).](https://pubmed.ncbi.nlm.nih.gov/37340477/) 5. [Lessons learned from manual curation of thousands of gene models in the nematode Pristionchus pacificus.](https://pubmed.ncbi.nlm.nih.gov/41818166/) 6. [Enhanced RNA-Seq Expression Profiling and Functional Enrichment in Non-model Organisms Using Custom Annotations.](https://pubmed.ncbi.nlm.nih.gov/42367215/) 7. [The promise of long-read RNA-seq: reducing bias in analyses of allele imbalance.](https://pubmed.ncbi.nlm.nih.gov/42553517/) 8. [Comparative gene annotation and orthology assignments across 301 species of Drosophilidae.](https://pubmed.ncbi.nlm.nih.gov/41706752/) 9. [DiffSegR: an RNA-seq data driven method for differential expression analysis using changepoint detection.](https://pubmed.ncbi.nlm.nih.gov/37954572/) 10. [A comprehensive review of variant calling tools for RNA-seq: Challenges and advances.](https://pubmed.ncbi.nlm.nih.gov/42081940/) 11. [Notable challenges posed by long-read sequencing for the study of transcriptional diversity and genome annotation.](https://pubmed.ncbi.nlm.nih.gov/40032585/) 12. [A high-resolution single-molecule sequencing-based Arabidopsis transcriptome using novel methods of Iso-seq analysis.](https://pubmed.ncbi.nlm.nih.gov/35799267/) 13. [Assembly-free rapid differential gene expression analysis in non-model organisms using DNA-protein alignment.](https://pubmed.ncbi.nlm.nih.gov/35120462/) 14. [ToxIR: an accurate RNA-seq pipeline for high-precision toxin transcriptome profiling, validated in odontobuthus doriae venom glands.](https://pubmed.ncbi.nlm.nih.gov/41444383/) 15. [Freddie: annotation-independent detection and discovery of transcriptomic alternative splicing isoforms using long-read sequencing.](https://pubmed.ncbi.nlm.nih.gov/36478271/) 16. [scRCA: A Siamese network-based pipeline for annotating cell types using noisy single-cell RNA-seq reference data.](https://pubmed.ncbi.nlm.nih.gov/40158457/) 17. [A machine learning-based method for automatically identifying novel cells in annotating single-cell RNA-seq data.](https://pubmed.ncbi.nlm.nih.gov/36083008/)