A gene produces more sequencing fragments in one sample than another. That can reflect more RNA molecules. It can also reflect a change toward longer transcript isoforms, which offer more opportunities to generate fragments.
This guide separates those explanations using a small, invented RNA-seq model. It shows how sample-specific effective length enters a gene-level comparison, why a single fixed gene length can miss the problem, and when a length correction would itself be inappropriate. The arithmetic explains a mechanism; it is not a complete differential-expression analysis.
eduKateSG · The Gold Standard of Bioinformatics
Choose your reading route
Separate molecules from fragments · Chapters 1–2. Build the quantities before using a formula.
Calculate the effect of an isoform switch · Chapters 3–5. Follow original examples, assumptions and changed conditions.
Check the protocol and inference · Chapters 6–8. Test transfer, answer common questions and follow connected reading.
Full chapter index · Practice and evidence checks · Return to Bioinformatics
Chapter index
Separate molecules from fragments · Chapters 1–2
Calculate the effect of an isoform switch · Chapters 3–5
Check the protocol and inference · Chapters 6–8
CHAPTER 1 OF 8 · Separate molecules from fragments
1. Distinguish a gene, its transcripts and the fragments counted
A gene can produce different RNA transcripts. These isoforms may share much of their sequence while differing in included exons or transcript ends. Consequently, “the length of this gene” is not automatically one relevant RNA length.
In a conventional fragmented RNA-seq library, a longer transcript can yield more sampled fragments per molecule than a shorter transcript. A gene-level count pools evidence from its transcripts. If the mixture changes between samples, the count can change even when the total number of molecules assigned to that gene does not.
Three questions therefore need separate answers. How many gene molecules are represented? Which transcripts make up that total? How does the experimental protocol turn those molecules into counted observations?
This does not mean that gene-level analysis is useless. It means the measurement model must match the biological question. A change in transcript usage may be biologically meaningful even when total gene abundance stays constant. Calling every count change an abundance change would lose that distinction.
State the comparison unit
For the invented examples, N means a relative amount of gene molecules in the represented RNA population. It does not automatically mean molecules per cell. The spike-in reference-unit guide explains what additional reference information is needed for a per-cell interpretation.
CHAPTER 2 OF 8 · Separate molecules from fragments
2. Build the sampling relationship before using a correction
Let n₁ and n₂ be the represented amounts of two isoforms, and let ℓ₁ and ℓ₂ be their effective lengths. In a deliberately simple model, the expected gene count is:
E(C) = q(n₁ℓ₁ + n₂ℓ₂) = qNℓ̄.
Here N = n₁ + n₂, and ℓ̄ = (n₁ℓ₁ + n₂ℓ₂)/N. The factor q summarises sample-specific sampling depth and the composition-dependent chance of observing the represented RNA. It is assumed known or appropriately accounted for in each exercise.
Effective length is a sampling quantity, not simply the annotated number of nucleotides. The Salmon output documentation distinguishes Length from EffectiveLength and describes the latter’s dependence on the fragment distribution and modelled biases. Its NumReads values are estimates, potentially involving ambiguous mappings, rather than an independent census of molecules.
For a simple uniform-fragment picture, a transcript of length L and a fixed fragment length f has L − f + 1 possible starting positions when L ≥ f. That counting argument helps explain why fragment length matters. Real effective-length estimation is more involved and should not be replaced by this classroom shortcut.
The key relationship is multiplicative: count combines abundance, average sampling opportunity and sample scale. A single observed increase cannot reveal which factor changed without additional information.
CHAPTER 3 OF 8 · Calculate the effect of an isoform switch
3. Work an isoform switch with unchanged gene abundance
Imagine two isoforms with effective lengths 1,000 and 2,000 sampling units. Sample A contains 100 represented molecules of the short isoform and none of the long one. Sample B contains 100 of the long isoform and none of the short one. Set q = 0.01 in both samples.
| Quantity | Sample A | Sample B |
|---|---|---|
| Short-isoform amount | 100 | 0 |
| Long-isoform amount | 0 | 100 |
| Total gene amount N | 100 | 100 |
| Average effective length ℓ̄ | 1,000 | 2,000 |
| Expected count qNℓ̄ | 1,000 | 2,000 |
The expected count ratio is 2. The gene-amount ratio is 1. The extra fragments arise from doubled sampling opportunity per represented gene molecule.
Correcting by each sample’s effective length exposes the difference:
A: 1,000/1,000 = 1. B: 2,000/2,000 = 1.
These corrected quantities still contain the common q factor. Their equality supports unchanged N within the stated model; their numerical value is not an absolute molecule count.
Why one fixed gene length fails here
Suppose both counts are divided by a single annotated length of 1,500. The resulting values are about 0.667 and 1.333. Their ratio remains 2. Dividing both sides by the same constant cannot remove a sample-dependent difference.
Likewise, doubling every sample-B count with a single library-depth explanation would be inappropriate if only this gene changed its isoform mixture. A library-scale factor and a gene-specific effective-length factor play different roles.
The corrected result also does not say “nothing happened.” An entire transcript switch happened by construction. A separate transcript-usage question is needed to describe it.
CHAPTER 4 OF 8 · Calculate the effect of an isoform switch
4. Change both abundance and transcript mixture
Real contrasts need not be all-short versus all-long. Keep the same two effective lengths and q = 0.01. Sample A contains 80 short and 20 long represented molecules. Sample B contains 40 short and 160 long molecules.
For A, N = 100 and ℓ̄ = (80 × 1,000 + 20 × 2,000)/100 = 1,200. The expected count is 1,200.
For B, N = 200 and ℓ̄ = (40 × 1,000 + 160 × 2,000)/200 = 1,800. The expected count is 3,600.
The raw expected ratio is 3. After accounting for average length, the amount ratio is:
(3,600/1,800)/(1,200/1,200) = 2.
The threefold count increase comprises a twofold abundance increase multiplied by a 1.5-fold change in average effective length. Neither explanation alone describes the constructed data.
Now change sampling depth
If q for B were half that for A, B’s expected count would be 1,800 rather than 3,600. The observed count ratio would then be 1.5, despite the same twofold amount increase. In general:
N_B/N_A = (C_B/C_A) × (q_A/q_B) × (ℓ̄_A/ℓ̄_B),
using expected counts or sufficiently supported estimates within the model. Here 1.5 × 2 × (1,200/1,800) = 2.
This algebra is a diagnostic explanation, not a replacement for fitting a count model across biological replicates. Sampling variation, uncertainty in transcript assignment and between-sample biological variation remain.
CHAPTER 5 OF 8 · Calculate the effect of an isoform switch
5. Match the analysis route to the library protocol
The tximport vignette describes gene-level summarisation that retains sample-specific effective-length information. Its documented routes include original estimated counts with appropriate offsets, or suitable abundance-derived counts without applying the length offset again. Keep the import choice and downstream method consistent.
An offset accounts for a known or estimated exposure term in the model. In this context, it can tell the model that one sample offered more fragment-generating opportunity for a gene. It is not an instruction to subtract a fixed number of reads.
Avoid correcting the same effect twice
In the first example, length adjustment already reduces the apparent ratio from 2 to 1. If a second, unintended adjustment divides the B-to-A ratio by another factor of 2, the reported ratio becomes 0.5. A false decrease has been manufactured from unchanged gene amount.
Do not impose a full-transcript model on every assay
For 3′-tagged counting protocols, counts need not have the full-transcript length dependence assumed here. The tximport authors explicitly distinguish that case. Applying the ordinary full-length correction can introduce bias. Unique molecular identifiers help address duplicate counting, but the acronym UMI alone does not define every capture and sampling property.
Record how the library was made before choosing the correction. A method cannot infer the right physical model from a spreadsheet column named “counts.”
CHAPTER 6 OF 8 · Check the protocol and inference
6. Check the evidence and practise transfer
| Check | Why it changes the interpretation |
|---|---|
| Transcript reference and annotation version | Defines the isoforms and gene membership being estimated |
| Library protocol and fragment distribution | Determines whether length affects observation probability |
| Transcript-to-gene mapping | Prevents aggregating the wrong identifiers |
| Sample-specific effective lengths and count-import mode | Reveals missing or duplicated adjustments |
| Replicates, design and mapping ambiguity | Supports uncertainty and biological inference |
Transfer question one: an unchanged mixture
Two samples both contain 50 short and 50 long molecules under equal q. Their average effective lengths are both 1,500. If the second sample has twice as many molecules of each isoform, what happens?
Its total amount and expected count both double, while average length stays 1,500. A sample-specific length adjustment does not remove the real abundance change. Explain why before calculating.
Transfer question two: ambiguous assignment
A gene’s isoforms share almost all their sequenced regions. A program returns a precise-looking mixture. Does that prove the average length is known accurately?
No. A model can allocate ambiguous fragments while leaving substantial uncertainty about the allocation. Inspect the quantifier’s uncertainty information and relevant read evidence. More decimal places do not create more distinguishing sequence.
Transfer question three: a protocol change
The second experiment counts one captured end per molecule rather than fragments sampled across transcript bodies. Can you copy the first experiment’s length correction?
No. Rebuild the observation model. The difference in protocol changes the relationship that justified the correction. That is the transferable skill: preserve the reasoning, then test whether its conditions still hold.
CHAPTER 7 OF 8 · Check the protocol and inference
7. Questions about length and gene expression
Does the DNA gene become longer when isoform usage changes?
Not in this model. The mixture of RNA products changes. The abundance-weighted effective length can therefore change while the genomic locus remains the same.
Is the average weighted by read counts?
The model weights effective lengths by represented transcript amounts. Raw fragment counts already reflect length, so using them as molecule weights without adjustment can count the length effect twice.
Does unchanged gene abundance imply unchanged protein function?
No. Transcript composition can matter, and protein abundance and function require further evidence. A gene-level summary does not answer every downstream biological question.
Should TPM values be entered into any count-analysis program?
No. Relative abundance measures, original estimated counts and specifically constructed abundance-derived count matrices have different roles. Follow the chosen method’s documented inputs and statistical assumptions.
Does length correction establish statistical significance?
No. It addresses a measurement effect. Replication, dispersion, experimental design, hypothesis tests and multiple-testing treatment still matter.
CHAPTER 8 OF 8 · Check the protocol and inference
8. Continue through the connected RNA evidence
Understand the quantities
- The Gold Standard of Bioinformatics connects data, references, computation and biological claims.
- RNA-seq count normalisation and differential expression develops library scaling, count models and experimental design.
- GTF gene and transcript annotations explains the reference records behind transcript-to-gene membership.
Ask what changed within the gene
- DRIMSeq and differential transcript usage follows changes in transcript proportions.
- rMATS and replicate-aware splicing examines event-level evidence.
- IsoformSwitchAnalyzeR and functional isoform consequences connects a switch to further hypotheses.
- sleuth and transcript-quantification uncertainty develops uncertainty beyond a point estimate.
Complete the interpretation
Independent filtering and missing adjusted p-values addresses a separate statistical reporting question. Return to The Gold Standard of Education for transferable reasoning, or explore How X Works.
The official Salmon and Bioconductor documentation linked above was checked on 22 October 2026. It supports the quantity definitions and documented analysis routes. All molecule amounts, sampling factors and effective lengths in the worked examples are invented, with uncertainty deliberately omitted from the arithmetic so the mechanism is visible.