Differential Isoform Expression Analysis with Long-Read RNA-Seq: How to Use DEXSeq and DRIMSeq

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

Long-read RNA sequencing produces full-length transcript reads that can resolve isoform structure directly, but the statistical frameworks most bioinformaticians know from short-read work need careful adaptation before they can be applied to this data type. This article explains how to prepare long-read data for differential isoform expression analysis using DEXSeq and DRIMSeq, what each tool actually tests, where their assumptions break down with long reads, and how to interpret results without overstating biological conclusions.

The practical problem is straightforward. Standard differential expression tools were built around short-read counting models that assume reads map to small exonic fragments. Long-read data changes the unit of observation from the fragment to the full transcript, which alters count distributions, library size calculations, and the meaning of isoform-level comparisons. Researchers who run DEXSeq or DRIMSeq on long-read data without adjusting their workflow often produce results that reflect technical artifacts instead of biology.

This article gives you a workflow that accounts for those differences. You will learn what inputs each tool requires, how to construct the annotation files that make isoform-level testing possible, which quality checks protect against false discoveries, and how to report results in a way that other researchers can reproduce.

Key Takeaways

• Differential isoform expression analysis with long-read RNA-Seq requires adapting DEXSeq and DRIMSeq workflows because long-read data changes count distributions, library size calculations, and isoform-level comparisons compared to short-read data.
• DEXSeq tests for differential exon usage by counting reads overlapping each exon within a gene, while DRIMSeq tests for differential transcript usage at the level of full isoforms using a Dirichlet-multinomial model.
• Long-read data violates short-read assumptions because a single sequencing molecule can produce multiple reads through rolling circle amplification, creating non-independent counts that inflate confidence in differential calls.
• For long-read data, DEXSeq requires exon-level counts derived from full-length transcript alignments, while DRIMSeq requires transcript-level counts where each read is assigned to a transcript based on its full alignment.
• A significant DEXSeq result indicates differential exon usage but does not identify which isoform is responsible, so examining isoform-level changes with DRIMSeq or direct isoform quantification is needed for biological interpretation.

What DEXSeq and DRIMSeq Actually Test

DEXSeq and DRIMSeq answer different questions about isoform usage, and confusing the two is a common source of analytical error.

DEXSeq tests for differential exon usage. It counts reads overlapping each exon or exon fragment within a gene and asks whether the proportion of total gene counts assigned to a particular exon changes between conditions. The tool was originally designed for short-read data where exon-level counts are the natural unit of observation. When applied to long-read data, DEXSeq can still be used, but the exon counts must be derived from full-length transcript alignments instead of from fragmented read overlaps.

DRIMSeq tests for differential transcript usage. It works at the level of full isoforms and asks whether the relative proportion of each transcript within a gene changes between conditions. This is conceptually closer to what long-read data can directly measure, because each read represents a complete or near-complete transcript molecule. DRIMSeq uses a Dirichlet-multinomial model that accounts for the compositional nature of isoform counts within a gene.

The distinction matters for interpretation. A gene can show no change in total expression but still have significant differential transcript usage if the proportions of its isoforms shift. The reverse is also possible. A gene can change in overall expression while maintaining the same isoform proportions. DEXSeq detects the first pattern at exon resolution. DRIMSeq detects it at transcript resolution. Neither tool tests for differential gene expression in the conventional sense, and neither should be used for that purpose.

Long-read data provides an advantage for both tools because isoform structure is observed directly instead of inferred. Short-read approaches must reconstruct isoforms from exon coverage patterns, which introduces uncertainty about which transcript a given read came from. Long-read sequencing eliminates much of that ambiguity, as described in a 2024 review of isoform expression analysis that notes the ability of long-read methods to characterize isoform expression in single-cell and spatial contexts [<a href="#ref-1">1</a>].

Why Short-Read Assumptions Fail With Long-Read Data

The statistical models in DEXSeq and DRIMSeq assume that read counts follow certain distributions and that technical variation behaves in predictable ways. Long-read data violates several of these assumptions in ways that matter for downstream inference.

Count Distribution Mismatch

Short-read count models assume that each read represents an independent sampling event from a pool of transcripts. With long-read data, a single sequencing molecule can produce multiple reads through rolling circle amplification or concatemer formation, depending on the library preparation method. This creates non-independence in the counts that standard models do not account for.

The practical consequence is inflated confidence in differential calls. When counts are not truly independent, the effective sample size is smaller than the raw read count suggests, and p-values become overly optimistic.

Library Size Calculation Differences

Library size normalization in short-read pipelines typically uses total mapped reads or a robust summary statistic like median-of-ratios. Long-read data has much lower throughput per sample, often by orders of magnitude. A sample with 10 million short reads and a sample with 500,000 long reads are not directly comparable in terms of count depth, and normalization methods that work well for one may behave poorly for the other.

The uncertainty in isoform expression estimates is also larger for long-read data because of the lower counts. A 2023 paper introducing DELongSeq makes this point directly, arguing that estimation efficiency, meaning the degree of uncertainty in expression estimates, is an important factor for downstream analysis and that conventional read count approaches do not capture this uncertainty [<a href="#ref-2">2</a>].

Isoform Assignment Ambiguity

Even with full-length reads, some reads will not cover the complete transcript. Partial reads can be ambiguous about which isoform they represent, particularly for genes with many alternative exons or alternative transcription start and polyadenylation sites. This ambiguity is different from short-read ambiguity, where the problem is reconstructing isoforms from fragments. With long reads, the problem is deciding whether a partial read belongs to one isoform or another when the missing portion would distinguish them.

Data Preparation for Long-Read Differential Isoform Analysis

The success of DEXSeq and DRIMSeq analysis depends almost entirely on the quality of the input data and annotations. This section covers the preparation steps that determine whether downstream statistical testing is meaningful.

Base Calling and Demultiplexing

Start with high-quality base calling. For nanopore data, this means using the current base calling models available in the sequencing platform software. For PacBio data, the circular consensus sequencing mode provides higher per-read accuracy than continuous long-read mode and is generally preferred for isoform analysis.

Demultiplexing must be done carefully because barcode misassignment creates cross-sample contamination that inflates apparent biological variation. Check the barcode assignment rates in the sequencing summary and investigate any sample with an unusually low or high proportion of assigned reads.

Alignment to the Reference Genome

Align long reads to the reference genome using a splice-aware aligner that can handle the read lengths and error profiles of long-read data. The choice of aligner affects downstream quantification, so document the version and parameters used.

For nanopore data, aligners designed for error-prone long reads are appropriate. For PacBio data, aligners that handle the higher accuracy of circular consensus reads may perform differently. Test a small subset of reads with candidate aligners and compare mapping rates, splice junction accuracy, and alignment quality metrics before committing to one for the full dataset.

The reference genome version matters. Isoform annotation is genome-version-specific, and mixing annotations from one genome build with alignments from another creates systematic errors. Record the exact genome build and annotation version used for every analysis.

Transcript Annotation and Quantification

Long-read data can be analyzed with reference annotations, de novo transcript discovery, or a combination of both. The choice affects what DEXSeq and DRIMSeq can test.

With reference annotations, you quantify reads against known isoforms. This is the simplest approach and works well when the organism and tissue have well-curated annotations. The limitation is that novel isoforms are invisible to the analysis.

With de novo transcript discovery, you assemble transcripts from the long-read data itself. This can reveal novel isoforms but introduces additional complexity because the transcript models must be validated and filtered for artifacts.

A hybrid approach uses the reference annotation as a scaffold and adds novel transcripts that pass quality filters. This is often the most informative but requires careful curation.

The 2024 study of tau pathology in the entorhinal cortex used long-read sequencing with a novel bioinformatics tool called FICLE to profile transcript diversity and identified hundreds of novel isoforms, including differentially expressed transcripts of genes such as Apoe, App, Cd33, Clu, Fyn, and Trem2 [<a href="#ref-3">3</a>]. This demonstrates that reference-only analysis would have missed a substantial portion of the biologically relevant transcript diversity.

Generating Count Matrices for DEXSeq and DRIMSeq

DEXSeq requires exon-level counts. For long-read data, you need to generate these counts from transcript alignments instead of from short-read fragment overlaps. The DEXSeq package documentation on Bioconductor provides the standard workflow for building exon count matrices from aligned reads [<a href="#ref-4">4</a>].

DRIMSeq requires transcript-level counts. Each read is assigned to a transcript based on its alignment, and the counts are aggregated per transcript per sample. The DRIMSeq package documentation on Bioconductor describes the expected input format and the steps for constructing the count matrix from quantification results [<a href="#ref-4">4</a>].

The key difference from short-read workflows is that transcript assignment for long reads uses the full alignment instead of a probabilistic assignment based on fragment compatibility. This reduces the ambiguity in the count matrix but does not eliminate it entirely.

Running DEXSeq on Long-Read Data

DEXSeq was designed for short-read data, but it can be adapted for long-read data with appropriate preparation. The workflow below describes the steps and the decisions that matter.

Building the Exon Count Matrix

The DEXSeq workflow requires a count matrix where rows are exonic regions and columns are samples. For long-read data, you construct this matrix from transcript alignments by counting the number of reads that overlap each exonic region.

The critical decision is how to define exonic regions. DEXSeq can use the union of all exons from all isoforms of a gene, or it can use the disjoint exonic regions that result from merging overlapping exons across isoforms. The latter approach, which DEXSeq implements in its own annotation preparation step, creates bins that distinguish alternative splicing events.

For long-read data, you should verify that the exonic bins match the isoform structures present in your data. If your long-read analysis has identified novel exons or novel splice junctions, the reference-based bins will not capture them. You may need to add novel exons to the annotation before building the count matrix.

Statistical Model and Dispersion Estimation

DEXSeq fits a generalized linear model with a negative binomial distribution for each exonic region. The model includes terms for the condition effect and the interaction between condition and exon, which tests whether the relative usage of each exon changes between conditions.

Dispersion estimation is a critical step. DEXSeq borrows information across exons to estimate dispersion, which works well when there are many exons with similar count levels. Long-read data often has lower counts per exon than short-read data, which makes dispersion estimation noisier.

Check the dispersion plot after fitting the model. If the dispersion estimates are highly variable or show poor fit to the mean-dispersion relationship, consider whether the count depth is sufficient for reliable inference.

Interpreting DEXSeq Results

DEXSeq produces a p-value for each exonic region, testing whether the relative usage of that region differs between conditions. Significant results indicate differential exon usage, which can arise from alternative splicing, alternative promoter usage, or alternative polyadenylation.

The interpretation requires care. A significant exon usage change does not tell you which isoform is responsible. It only tells you that the proportion of reads covering that exon differs between conditions. To understand the biological mechanism, you need to examine the isoform-level changes, which is where DRIMSeq or direct isoform quantification becomes useful.

The 2025 study of the adult human heart used long-read single-nucleus RNA sequencing to identify differential isoform usage across cell types and conditions, finding that approximately 30 percent of cell type-specific genes were polyform, meaning they used multiple isoforms tailored to cell type-specific programs [<a href="#ref-5">5</a>]. This illustrates the biological importance of exon and isoform usage changes that gene-level analysis would miss.

Running DRIMSeq on Long-Read Data

DRIMSeq is conceptually better suited to long-read data because it works at the transcript level. The workflow below describes the steps and the decisions that matter.

Building the Transcript Count Matrix

DRIMSeq requires a count matrix where rows are transcripts and columns are samples. Each cell contains the number of reads assigned to that transcript in that sample.

For long-read data, transcript assignment is based on the alignment of each read to the reference. Reads that align unambiguously to a single transcript are assigned directly. Reads that are compatible with multiple transcripts require a decision. Some workflows assign the read to the best-matching transcript based on alignment score. Others use a probabilistic assignment that distributes the read across compatible transcripts.

The choice matters for downstream results. Probabilistic assignment introduces uncertainty that DRIMSeq does not model. Direct assignment to the best match is simpler but can bias counts toward the most abundant isoform when reads are partial.

Filtering Low-Count Transcripts

DRIMSeq requires filtering of transcripts with very low counts before fitting the model. The standard approach is to remove transcripts that do not reach a minimum count in a minimum number of samples. The exact thresholds depend on your data and should be chosen based on the count distribution.

The filtering step is important because transcripts with near-zero counts in most samples create numerical problems in the Dirichlet-multinomial model and inflate the number of tests. Document the filtering criteria and report how many transcripts were retained.

Statistical Model and Hypothesis Testing

DRIMSeq fits a Dirichlet-multinomial model for each gene, where the response is the vector of transcript counts and the model tests whether the transcript proportions differ between conditions. The model accounts for overdispersion relative to a simple multinomial distribution.

The output is a p-value per gene for differential transcript usage. This tests whether the relative proportions of isoforms within the gene change between conditions, regardless of whether total gene expression changes.

Interpreting DRIMSeq Results

A significant DRIMSeq result indicates that the isoform composition of a gene differs between conditions. This is a gene-level test, not a transcript-level test. To identify which specific transcripts drive the difference, you need to examine the fitted proportions and their confidence intervals.

The biological interpretation should consider the functional consequences of the isoform changes. Isoforms can differ in protein-coding sequences, functional domains, and molecular networks, as noted in the heart isoform study [<a href="#ref-5">5</a>]. A shift in isoform usage that changes the protein product is more likely to have functional consequences than a shift between isoforms that differ only in untranslated regions.

Combining DEXSeq and DRIMSeq in One Workflow

DEXSeq and DRIMSeq provide complementary information, and a complete analysis should use both. The workflow below integrates them into a coherent pipeline.

Step 1: Quantify Transcripts From Long-Read Alignments

Start with aligned long reads and generate transcript-level quantification. Use a quantification tool that can handle long-read data and produces both transcript-level counts and exon-level counts.

Step 2: Run DEXSeq for Exon-Level Analysis

Build the exon count matrix from the alignments and run DEXSeq to identify exonic regions with differential usage between conditions. This analysis is sensitive to alternative splicing events that change exon inclusion.

Step 3: Run DRIMSeq for Transcript-Level Analysis

Build the transcript count matrix and run DRIMSeq to identify genes with differential transcript usage. This analysis captures the combined effect of all isoform changes within a gene.

Step 4: Cross-Reference Results

Compare the DEXSeq and DRIMSeq results to identify genes that show significant changes in both analyses. These are the most robust findings because they are supported by two independent statistical approaches.

Genes that show significant DRIMSeq results but not significant DEXSeq results may have isoform changes that do not alter exon usage patterns, such as changes in alternative transcription start sites that do not affect the exonic bins tested by DEXSeq.

Genes that show significant DEXSeq results but not significant DRIMSeq results may have exon usage changes that are distributed across multiple isoforms in a way that does not produce a significant transcript proportion shift.

Step 5: Validate With Visualization

Visualize the results for genes of interest. The ScisorWiz tool provides visualization of differential isoform expression in single-cell long-read data, with clustering methods and highlighting of alternative exons and single-nucleotide variants [<a href="#ref-6">6</a>]. While designed for single-cell data, the visualization principles apply to bulk long-read data as well.

Plot the isoform proportions for each sample, colored by condition, to confirm that the statistical results reflect visible biological patterns instead of artifacts of the model.

At a Glance: Tool Selection and Workflow Decisions

Practical Implementation Steps

The following steps provide a concrete workflow for implementing differential isoform expression analysis with DEXSeq and DRIMSeq on long-read data.

Step 1: Verify Data Quality Before Analysis

Check the sequencing summary for each sample. Record the number of reads, the read length distribution, and the estimated read accuracy. Samples with very low read counts or poor accuracy should be flagged before analysis.

Confirm that the number of biological replicates is sufficient for the statistical models. DEXSeq and DRIMSeq both require replication to estimate dispersion. Three replicates per condition is the practical minimum, and more is better.

Step 2: Align Reads and Assess Mapping Quality

Align reads to the reference genome and record the mapping rate for each sample. Investigate samples with mapping rates that are substantially lower than the rest of the cohort.

Check the distribution of alignment lengths and the number of reads that align to multiple locations. Multi-mapping reads should be handled consistently across all samples.

Step 3: Quantify Transcripts and Build Count Matrices

Use a long-read-aware quantification approach to assign reads to transcripts. Record the number of reads assigned to each transcript for each sample.

Build the exon count matrix for DEXSeq and the transcript count matrix for DRIMSeq. Verify that the matrices have the expected dimensions and that the counts match the alignment files.

Step 4: Run Quality Control on Count Matrices

Check for sample-level outliers using principal component analysis or clustering. Samples that cluster separately from their condition group may have technical issues.

Check for genes with extreme count distributions that could indicate alignment or annotation problems.

Step 5: Run DEXSeq and DRIMSeq Analyses

Run both tools with documented parameters. Record the version numbers of all software packages used.

Save the full results tables, beyond the significant hits, so that the analysis can be revisited with different thresholds.

Step 6: Interpret and Report Results

Report the number of significant exonic regions from DEXSeq and the number of significant genes from DRIMSeq. Include the filtering criteria and the multiple testing correction method.

For genes of interest, report the direction of the isoform usage change and the specific transcripts involved.

Records and Measurements for Reproducible Analysis

Reproducibility requires detailed records of every step in the analysis. The following measurements should be recorded for each analysis run.

Sequencing and Alignment Records

Record the sequencing platform, the base calling model, the aligner version, and the alignment parameters. Record the reference genome build and the annotation version.

Record the mapping rate, the median read length, and the number of reads per sample. These measurements allow other researchers to assess whether the data quality is adequate for the analysis.

Quantification Records

Record the quantification tool and version, the transcript assignment method, and any filtering thresholds applied. Record the number of transcripts detected per sample and the proportion of reads assigned to known versus novel transcripts.

Statistical Analysis Records

Record the DEXSeq and DRIMSeq package versions, the model formulas, the dispersion estimation method, and the multiple testing correction method. Record the filtering criteria for low-count exons and transcripts.

Record the number of tests performed and the significance thresholds used. This information is essential for interpreting the results and for comparing across studies.

Common Failure Patterns and How to Avoid Them

Several recurring problems appear when researchers apply DEXSeq and DRIMSeq to long-read data. Recognizing these patterns helps avoid wasted effort and false conclusions.

Failure Pattern 1: Insufficient Sequencing Depth

Long-read sequencing produces fewer reads per sample than short-read sequencing. When the read depth is too low, the count matrices are sparse, dispersion estimates are unstable, and the statistical tests lack power.

The solution is to sequence more deeply or to increase the number of biological replicates. There is no fixed threshold that works for all experiments, but examining the count distribution and the dispersion plot will reveal whether the depth is adequate.

Failure Pattern 2: Ignoring Novel Isoforms

Reference annotations are incomplete for most organisms and tissues. When the analysis only considers annotated isoforms, novel isoforms that are biologically relevant are missed entirely.

The tau pathology study identified hundreds of novel isoforms using long-read sequencing [<a href="#ref-3">3</a>]. An analysis restricted to reference annotations would have missed these transcripts and the differential expression patterns associated with them.

The solution is to perform transcript discovery on the long-read data and add validated novel transcripts to the annotation before quantification.

Failure Pattern 3: Treating All Reads as Full-Length

Not all long reads cover the complete transcript. Partial reads are common, particularly for very long transcripts. When partial reads are assigned to isoforms based on incomplete information, the counts can be biased.

The solution is to examine the read length distribution relative to the transcript length distribution and to consider whether partial reads should be filtered or handled differently in the assignment step.

Failure Pattern 4: Overinterpreting Exon-Level Results

A significant DEXSeq result for an exonic region does not identify the specific isoform change responsible. Researchers who conclude that a particular isoform is differentially expressed based only on DEXSeq results are overinterpreting the data.

The solution is to cross-reference DEXSeq results with DRIMSeq results and to examine the isoform-level counts directly for genes of interest.

Failure Pattern 5: Ignoring Compositional Effects

DRIMSeq tests the relative proportions of transcripts within a gene. A significant result can arise from changes in one transcript or from coordinated changes in multiple transcripts. The test does not distinguish these scenarios.

The solution is to examine the fitted proportions and their uncertainty for each transcript in significant genes.

Limitations of DEXSeq and DRIMSeq for Long-Read Data

Both tools have limitations that should be acknowledged in any report of results.

DEXSeq Limitations

DEXSeq was designed for short-read data and its exon bin construction assumes that exonic regions can be defined from the reference annotation. Long-read data often reveals novel exons and novel splice junctions that are not in the reference, and these are invisible to the standard DEXSeq workflow.

The negative binomial model in DEXSeq assumes that counts are independent, which may not hold for long-read data where a single molecule can generate multiple reads.

DRIMSeq Limitations

DRIMSeq requires transcript-level counts, which depend on the accuracy of transcript assignment. Partial reads and reads that are compatible with multiple transcripts introduce uncertainty that the model does not capture.

The Dirichlet-multinomial model assumes that the transcript proportions are constant within a condition, which may not hold if there is substantial biological variation between replicates.

General Limitations

Both tools require biological replication and have limited power with small sample sizes. Both tools test for relative usage changes, not absolute expression changes. A gene can show significant differential transcript usage while having identical total expression between conditions.

The interpretation of results depends on the quality of the annotation and the accuracy of the quantification. Errors in either propagate through the statistical analysis.

Quality Controls and Validation Approaches

Quality control should be applied at multiple stages of the analysis to protect against false discoveries.

Alignment Quality Control

Check the mapping rate and the distribution of alignment scores. Reads with low alignment scores may be chimeric or may contain sequencing errors that affect transcript assignment.

Check the splice junction accuracy by comparing the junctions detected in the long-read data with known junctions in the annotation. A high proportion of novel junctions may indicate alignment errors or may indicate genuine novel splicing.

Quantification Quality Control

Check the correlation between replicate samples. High correlation within conditions and lower correlation between conditions is expected. Samples that do not follow this pattern should be investigated.

Check the proportion of reads assigned to each transcript. Transcripts with very low counts in most samples may be artifacts of the quantification process.

Statistical Quality Control

Examine the dispersion estimates from DEXSeq and DRIMSeq. Extreme dispersion values can indicate model misspecification or data quality problems.

Check the p-value distributions. A well-behaved analysis should produce a uniform distribution of p-values under the null hypothesis, with a peak near zero for true positives. An excess of very small p-values may indicate that the model is not accounting for all sources of variation.

Validation Approaches

Validate a subset of results using an independent method. The heart isoform study validated isoform compositions of representative cardiac genes using reverse transcription quantitative polymerase chain reaction and targeted amplicon sequencing [<a href="#ref-5">5</a>]. This type of validation provides confidence that the computational results reflect biological reality.

For novel isoforms, validation is particularly important because the computational evidence for their existence may be incomplete.

Safety and Regulatory Context for Research Applications

Differential isoform expression analysis is a research tool, not a clinical diagnostic. The results should not be used for clinical decision-making without appropriate validation and regulatory approval.

Research Use Only

The tools and workflows described in this article are for research purposes. They are not approved for clinical diagnostic use. Researchers who plan to use isoform expression data in clinical contexts should consult with their institutional review board and regulatory authorities.

Data Management and Privacy

Long-read sequencing data can contain sensitive information about research participants. Data should be stored securely and shared only in accordance with applicable regulations and institutional policies. The NCBI provides data resources and search systems for sequence data, and researchers should follow the data submission and access guidelines for these resources [<a href="#ref-7">7</a>].

Reproducibility Requirements

Practitioners seeking reproducible bioinformatics education will find GTN's open-access tutorial library a valuable starting point for pipeline design.

Researchers should deposit their raw sequencing data in an appropriate repository and make their analysis code available under a suitable license.

Professional Escalation Criteria

Some situations require escalation to a more experienced bioinformatician or a specialist in long-read analysis. The following criteria indicate when additional expertise is needed.

Escalate When Data Quality Is Inadequate

If the mapping rate is below 50 percent, if the read accuracy is poor, or if the read length distribution is substantially shorter than expected, escalate the issue before proceeding with analysis. Poor data quality cannot be fixed by statistical analysis.

Escalate When Results Are Biologically Implausible

If the analysis produces hundreds of significant genes with no plausible biological connection to the experimental question, escalate the issue. This pattern may indicate a technical artifact, a batch effect, or a problem with the statistical model.

Escalate When Replicates Do Not Cluster

If biological replicates from the same condition do not cluster together in principal component analysis, escalate the issue. This pattern indicates that technical variation is dominating the signal, and the results may not be reliable.

Escalate When Validation Fails

If independent validation of a subset of results fails to confirm the isoform changes, escalate the issue. This may indicate a systematic problem with the quantification or the statistical analysis.

Frequently Asked Questions

What is the difference between differential gene expression and differential isoform expression?

Differential gene expression tests whether the total expression level of a gene changes between conditions. Differential isoform expression tests whether the relative proportions of isoforms within a gene change between conditions. A gene can show no change in total expression while having significant isoform usage shifts, and the reverse is also possible. The heart isoform study found that cell state-specific programs tend to operate on monoform genes, while cell type-specific programs often involve polyform genes with multiple isoforms [<a href="#ref-5">5</a>]. This distinction is biologically important because isoform changes can alter protein products without changing total gene expression.

Can DEXSeq and DRIMSeq be used interchangeably for long-read data?

No. DEXSeq tests for differential exon usage at the level of exonic regions, while DRIMSeq tests for differential transcript usage at the level of full isoforms. They answer different questions and provide complementary information. DEXSeq is more sensitive to specific alternative splicing events, while DRIMSeq captures the combined effect of all isoform changes within a gene. A complete analysis should use both tools and cross-reference the results.

What is the minimum number of biological replicates needed for DEXSeq and DRIMSeq analysis?

Both tools require biological replication to estimate dispersion. Three replicates per condition is the practical minimum, but more replicates provide more stable dispersion estimates and greater statistical power. The exact number needed depends on the biological variability of the system and the sequencing depth per sample. Examine the dispersion plot and the power of the analysis to determine whether the replication is adequate.

How should partial long reads be handled in isoform quantification?

Partial reads that do not cover the complete transcript create ambiguity in isoform assignment. The best approach depends on the data. Some workflows assign partial reads to the best-matching transcript based on alignment score. Others filter partial reads and analyze only full-length reads. The choice affects the count matrix and the downstream results. Document the approach and consider whether the proportion of partial reads differs between samples or conditions.

What should be done when the reference annotation is incomplete?

Reference annotations are incomplete for most organisms and tissues. Long-read data can be used for de novo transcript discovery to identify novel isoforms. The tau pathology study identified hundreds of novel isoforms using long-read sequencing [<a href="#ref-3">3</a>]. Novel transcripts should be validated and added to the annotation before quantification. The hybrid approach uses the reference annotation as a scaffold and adds novel transcripts that pass quality filters.

How should results from DEXSeq and DRIMSeq be validated?

Validation can be computational or experimental. Computational validation includes examining the fitted proportions, checking the p-value distributions, and cross-referencing results between tools. Experimental validation uses an independent method such as reverse transcription quantitative polymerase chain reaction or targeted amplicon sequencing, as was done in the heart isoform study [<a href="#ref-5">5</a>]. Experimental validation is particularly important for novel isoforms and for results that will be used in downstream applications.

What are the main limitations of using short-read tools on long-read data?

The main limitations are count distribution mismatches, library size calculation differences, and isoform assignment ambiguity. Short-read count models assume independent sampling events, which may not hold for long-read data where a single molecule can generate multiple reads. Library size normalization methods designed for short reads may behave poorly with the lower throughput of long-read data. Isoform assignment for partial reads introduces uncertainty that the statistical models do not capture.

How should differential isoform expression results be reported for publication?

Report the sequencing platform, the base calling model, the aligner version, the reference genome build, and the annotation version. Report the quantification tool and the filtering criteria. Report the DEXSeq and DRIMSeq package versions, the model formulas, and the multiple testing correction method. Report the number of tests performed and the significance thresholds. Deposit the raw sequencing data in an appropriate repository and make the analysis code available. The EMBL-EBI Training provides learning pathways for bioinformatics data resources and practical analysis education that can help researchers meet reporting standards [<a href="#ref-8">8</a>].

Related Bioinformatics Guides

• RNA Sequencing Data Analysis: From Raw Reads to Differential Expression
• RNA-Seq Data Analysis Workflow: From Raw Reads to Insights
• Evaluating Metagenomic Assembly Tools: A Benchmarking Framework for Short-Read and Long-Read Data
• Proteomics Data Analysis in R: A Practical Workflow for Differential Expression and Visualization
• RNA-Seq Data Analysis in Galaxy: A User-Friendly Platform

References and Further Reading

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