DEXSeq

Unlike gene-level tools, DEXSeq partitions genes into non-overlapping exonic counting bins. Each bin corresponds to a unique exonic region, allowing the model to detect differential exon usage (DEU) independently of overall gene expression changes. This granularity captures isoform switching that gene-level analysis masks.

DEXSeq employs a negative binomial GLM with an interaction term between exon and condition. This design simultaneously models:

• Condition effect: Overall expression change
• Exon effect: Baseline exon inclusion levels
• Interaction term: The parameter of interest, testing whether the condition effect differs per exon This structure controls for gene-level expression changes, isolating true differential exon usage.

The package estimates exon-specific dispersions using Cox-Reid profile-adjusted likelihood, then applies empirical Bayes shrinkage to stabilize estimates. A mean-dispersion trend is fit across all counting bins, and individual estimates are shrunk toward this trend. This borrowing of information across exons prevents inflated false positives from highly variable, low-count bins.

DEXSeq provides built-in plotting functions to visualize results:

• Transcript annotation plots: Display known isoforms alongside the gene model
• Expression plots: Show normalized counts per exon across conditions
• Fitted expression plots: Overlay the GLM fitted values with confidence intervals
• MA plots: Visualize the relationship between mean count and fold change per exon These diagnostics help researchers validate computational findings against known transcript structures.

DEXSeq operates within the Bioconductor framework, accepting input from standard alignment and quantification pipelines:

• HTSeq-count output in exon-level mode
• featureCounts with exon-level annotation
• Rsubread alignments Results are stored as DEXSeqResults objects, compatible with downstream tools like GenomicRanges for annotation overlap and rtracklayer for exporting to genome browsers.

DEXSeq applies the Benjamini-Hochberg procedure to control the false discovery rate across all tested exonic regions. The package also implements independent filtering to remove exons with very low counts before p-value adjustment, increasing detection power. Per-gene adjusted p-values are reported to identify genes with at least one differentially used exon, while per-exon adjusted p-values pinpoint the specific regions.

The biological phenomenon that DEXSeq is designed to detect. DTU refers to condition-dependent shifts in the relative abundance of alternative mRNA isoforms from the same gene. Unlike differential gene expression, which measures total transcriptional output, DTU captures regulatory mechanisms like alternative splicing, alternative promoter usage, and alternative polyadenylation that change isoform proportions without altering overall gene-level counts.

DEXSeq partitions each gene into non-overlapping exonic counting bins rather than full transcripts. This design avoids the statistical challenges of isoform-level quantification uncertainty. Key steps:

• Flatten gene models into disjoint exonic regions
• Count reads overlapping each bin per sample
• Fit a generalized linear model (GLM) testing for condition-bin interaction terms
• A significant interaction indicates differential exon usage between conditions

The original DEXSeq uses a negative binomial GLM with per-exon dispersion estimation, similar to DESeq2. DEXSeq-VM is a faster variant that employs a variational Bayes approach to approximate posterior distributions, dramatically reducing computation time for large experiments. Both test the same null hypothesis: that the relative usage of an exonic region does not change between conditions.

DEXSeq belongs to a family of count-based DTU tools. Complementary approaches include:

• rMATS: Uses junction reads and a hierarchical framework to detect five splicing event types
• SUPPA2: Leverages transcript-level quantifications (e.g., Salmon, kallisto) to compute percent-spliced-in (PSI) values
• LeafCutter: Focuses exclusively on intron excision ratios from junction reads, enabling cluster-level splicing analysis
• MAJIQ: Builds local splice graphs to detect complex, unannotated splicing variations

The plotDEXSeq() function generates gene-level visualization showing:

• Normalized counts per exonic bin for each condition
• Fitted expression estimates from the GLM
• Flattened gene model with annotated transcripts below
• Significance annotations marking bins with significant differential usage This plot is essential for confirming that detected DTU events correspond to interpretable isoform switches rather than technical artifacts.

After identifying genes with significant differential exon usage, common follow-up analyses include:

• Domain annotation: Determine if differentially used exons encode functional protein domains (e.g., kinase domains, transmembrane regions)
• Nonsense-mediated decay (NMD) prediction: Assess whether exon skipping introduces premature stop codons
• Motif enrichment: Search for RNA-binding protein motifs near regulated exons to infer splicing regulators
• Integration with eQTL data: Overlap DTU events with genetic variants to identify splicing quantitative trait loci (sQTLs)