RNA-seq: quantification, design, and differential expression

RNA-seq is not just about finding a list of significant genes. The quality of the biological design, the consistency of the library preparation, and the clarity of the model formula usually matter more than the choice of one aligner or one plotting style.

Two common analysis strategies

Genome alignment

STAR · HISAT2

Best when you need splice junctions, coverage QC, or novel feature discovery.

Transcript quantification

Salmon · Kallisto

Fast when a good transcriptome exists and you mainly want abundance for DE.

Lightweight alignment

JapalitySplice

For teaching or small genomes when a 30 GB STAR index is not available.

Minimal conceptual workflow

QC and trim if needed
Quantify genes or transcripts
Build a design matrix that reflects the biology
Normalize counts and inspect sample structure
Fit differential expression models
Interpret with pathways, markers, and experiment context

Statistics you cannot skip

Replicates: without biological replicates, DE claims are often weak.
Batch effects: prep date, lane, donor, and operator can dominate structure.
Multiple testing: raw p-values across thousands of genes are not enough.
Effect size: significant is not the same as biologically meaningful.
Normalization: library size helps; composition bias still matters.

Example command (Salmon)

Salmon writes transcript-level estimates. For gene-level DE with DESeq2 or edgeR you still need a gene mapping step such as tximport. Do not feed TPM into those count models.

salmon index -t transcripts.fa -i tx_index
salmon quant -i tx_index -l A \
  -1 trimmed_R1.fastq.gz -2 trimmed_R2.fastq.gz \
  -p 8 -o sample_salmon

Interactive RNA-seq interpretation guide

Click a stage to see the main question it answers and the most common way learners get misled.

Design decides what you are allowed to conclude

Before thinking about software, define your comparison: treatment vs control, paired vs unpaired, donor effects, time points, and replicates. A weak design cannot be repaired statistically afterward.

Include replicates whenever you want differential expression claims.
Write the experimental factors explicitly before you run the pipeline.

Quantification converts reads into comparable summaries

Gene-level counts are commonly used for DE, while transcript-level abundance can be useful for isoform questions. Choose a summary that matches the question, not just the easiest tool.

Counts are typically preferred for DE models like DESeq2 or edgeR.
TPM is useful for within-sample expression profiles, not as a drop-in replacement for DE counts.

Normalization reduces unfair sample-to-sample differences

Different libraries have different depths and composition. Normalization tries to make expression more comparable without erasing true biology.

Large composition shifts can distort naive interpretations of raw counts.
Always inspect sample-level plots after normalization.

Batch structure can overwhelm biology

If samples cluster by run date or operator rather than condition, the model and interpretation must account for that. Ignoring batch effects can create elegant but false results.

PCA is often the fastest way to see hidden sample structure.
Do not blindly remove variation you do not understand.

Differential expression is the start of interpretation, not the end

A volcano plot is only an overview. The real work is understanding whether the changing genes fit the biology, known pathways, cell composition shifts, or technical confounding.

Read effect size and FDR together.
Validate key genes with annotation and biological knowledge.

Volcano plot (example)

The y-axis here is −log10(FDR), not raw p. Genes with large effect size and strong adjusted support appear in the upper corners; biology still decides which of them matter.

Marker overview (example)

A small marker set can be visualized quickly, but this is a communication aid, not a substitute for proper modeling.

PCA-style sample structure (example)

The safest RNA-seq projects show samples separating primarily by biology, not by library date or lane.

MA plot (example)

MA plots help you see whether expression changes are balanced, intensity-dependent, or dominated by low-count noise.

Design checklist before you trust a DE result

Replicates and metadata

Every sample should have condition labels, batch information, and any paired/blocked structure clearly recorded.

QC consistency

Outlier samples with bad QC can dominate DE results if left in without explanation.

Interpretation guardrails

Significant genes should be checked for annotation quality, known pathways, and plausibility within the experiment.