Why Science? | limma-voom, Precision Weights and RNA-seq Linear Models

Section 1 of 36

1. Begin with unequal precision

RNA-seq log-counts do not have constant variance across abundance. Voom begins by estimating how measurement precision changes with the fitted expression level.

Voom converts counts to log-counts per million and estimates a mean–variance trend, producing observation-level precision weights for linear modelling.

Archive the sample manifest, count provenance, feature identifiers, formula, contrast, filters, package version, random seeds and exclusions so another analyst can reconstruct the decision path.

Contents · Next section

Section 2 of 36

2. Define counts per million

Counts per million adjust counts for library size on a relative scale. The logarithm improves visual and modelling behaviour, while offsets and small-count handling still matter.

The weights quantify estimated reliability on the transformed scale. They are not probabilities that a gene is true or important.

Keep the biological sample as the unit of replication. More reads or cells can improve measurement, but they do not create more independent people, animals or cultures.

Contents · Previous section · Next section

Section 3 of 36

3. Understand observation weights

Each gene–sample observation receives an estimated precision weight. Larger weights mean greater model influence, not greater biological importance.

Limma fits gene-wise linear models and moderates variance estimates across genes with empirical Bayes methods, improving stability in small replicated studies.

Pair adjusted evidence with effect direction, uncertainty, sample-level plots and sensitivity checks; a short ranked table is not a complete scientific result.

Contents · Previous section · Next section

Section 4 of 36

4. Meet the linear model

The design matrix describes expected log expression as a combination of coefficients. Linear modelling makes complex comparisons possible when the design is full rank.

Contrasts are algebraic questions asked of fitted coefficients. A correct contrast must match the design coding and the biological comparison.

Use negative controls, simulated nulls or label permutations only when their assumptions match the design, and name the particular false signal each check could reveal.

Contents · Previous section · Next section

Section 5 of 36

5. Understand empirical Bayes moderation

Limma shares information across genes to stabilise residual variance estimates. Moderation helps small studies but cannot substitute for missing experimental units.

Quality weights and observation weights solve different problems: one can downweight a sample, while the other follows the mean-dependent precision pattern.

Create an evidence card naming the question, measured material, statistical unit, model, comparison, result, validation status, alternatives and narrowest defensible claim.

Contents · Previous section · Next section

Section 6 of 36

6. Choose the right voom family

Standard voom, limma-trend, sample-quality weights and specialised voomLmFit routes address different mean–variance or sample-quality structures. Follow current documentation for the design at hand.

Weighted residuals, library-level plots and design-rank checks reveal failures that a smooth mean–variance curve alone can miss.

Write association or differential expression when that is what was tested. Reserve cause, mechanism, diagnosis and benefit for designs with stronger supporting evidence.

Contents · Previous section · Next section

Section 7 of 36

7. Read the 2014 voom evidence in scope

Law and colleagues showed how precision weights unlock limma’s linear-model tools for RNA-seq counts. The paper evaluates statistical performance, not biological truth for every detected gene.

Voom converts counts to log-counts per million and estimates a mean–variance trend, producing observation-level precision weights for linear modelling.

Archive the sample manifest, count provenance, feature identifiers, formula, contrast, filters, package version, random seeds and exclusions so another analyst can reconstruct the decision path.

Contents · Previous section · Next section

Section 8 of 36

8. Start from documented counts

Record quantification, annotation, transcript-to-gene summarisation and matrix orientation. Do not feed already normalised or transformed expression into a workflow expecting counts.

The weights quantify estimated reliability on the transformed scale. They are not probabilities that a gene is true or important.

Keep the biological sample as the unit of replication. More reads or cells can improve measurement, but they do not create more independent people, animals or cultures.

Contents · Previous section · Next section

Section 9 of 36

9. Filter weakly expressed features

Retain features with enough expression across relevant libraries to support a stable variance estimate. Base the rule on the design rather than the outcome labels alone.

Limma fits gene-wise linear models and moderates variance estimates across genes with empirical Bayes methods, improving stability in small replicated studies.

Pair adjusted evidence with effect direction, uncertainty, sample-level plots and sensitivity checks; a short ranked table is not a complete scientific result.

Contents · Previous section · Next section

Section 10 of 36

10. Normalise effective library sizes

A common workflow uses edgeR objects and composition factors before voom. Record these factors because voom uses effective, not merely raw, library sizes.

Contrasts are algebraic questions asked of fitted coefficients. A correct contrast must match the design coding and the biological comparison.

Use negative controls, simulated nulls or label permutations only when their assumptions match the design, and name the particular false signal each check could reveal.

Contents · Previous section · Next section

Section 11 of 36

11. Construct the design matrix

Represent conditions, batches, donors, time points and interactions only when supported by the samples. Inspect column names and rank before fitting.

Quality weights and observation weights solve different problems: one can downweight a sample, while the other follows the mean-dependent precision pattern.

Create an evidence card naming the question, measured material, statistical unit, model, comparison, result, validation status, alternatives and narrowest defensible claim.

Contents · Previous section · Next section

Section 12 of 36

12. Run voom with diagnostic plotting

Estimate the mean–variance trend and inspect the curve, points and residual standard-deviation pattern. A smooth line is a model component, not automatic validation.

Weighted residuals, library-level plots and design-rank checks reveal failures that a smooth mean–variance curve alone can miss.

Write association or differential expression when that is what was tested. Reserve cause, mechanism, diagnosis and benefit for designs with stronger supporting evidence.

Contents · Previous section · Next section

Section 13 of 36

13. Inspect the weight distribution

Compare weights across abundance, samples and groups. Systematically low weights in one library may signal quality problems requiring investigation.

Voom converts counts to log-counts per million and estimates a mean–variance trend, producing observation-level precision weights for linear modelling.

Archive the sample manifest, count provenance, feature identifiers, formula, contrast, filters, package version, random seeds and exclusions so another analyst can reconstruct the decision path.

Contents · Previous section · Next section

Section 14 of 36

14. Consider sample-quality weights

When whole libraries differ in reliability, quality weights may complement observation weights. Use them because diagnostics justify the model, not because they improve a preferred ranking.

The weights quantify estimated reliability on the transformed scale. They are not probabilities that a gene is true or important.

Keep the biological sample as the unit of replication. More reads or cells can improve measurement, but they do not create more independent people, animals or cultures.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with a fictional voom table

This classroom table is invented and is not a result from the voom paper.

Limma fits gene-wise linear models and moderates variance estimates across genes with empirical Bayes methods, improving stability in small replicated studies.

Pair adjusted evidence with effect direction, uncertainty, sample-level plots and sensitivity checks; a short ranked table is not a complete scientific result.

Contents · Previous section · Next section

Section 16 of 36

16. Fit weighted gene-wise models

lmFit uses the design and weights to estimate coefficients for every gene. Preserve the fitted object and the exact commands so contrasts can be audited.

Contrasts are algebraic questions asked of fitted coefficients. A correct contrast must match the design coding and the biological comparison.

Use negative controls, simulated nulls or label permutations only when their assumptions match the design, and name the particular false signal each check could reveal.

Contents · Previous section · Next section

Section 17 of 36

17. Specify contrasts visibly

Use makeContrasts or direct coefficient tests only after translating the comparison into plain language. Confirm reference levels and sign direction.

Quality weights and observation weights solve different problems: one can downweight a sample, while the other follows the mean-dependent precision pattern.

Create an evidence card naming the question, measured material, statistical unit, model, comparison, result, validation status, alternatives and narrowest defensible claim.

Contents · Previous section · Next section

Section 18 of 36

18. Apply empirical Bayes moderation

eBayes or treat moderates gene-wise variance information. treat asks whether effects exceed a chosen fold-change threshold rather than merely differing from zero.

Weighted residuals, library-level plots and design-rank checks reveal failures that a smooth mean–variance curve alone can miss.

Write association or differential expression when that is what was tested. Reserve cause, mechanism, diagnosis and benefit for designs with stronger supporting evidence.

Contents · Previous section · Next section

Section 19 of 36

19. Control false discoveries

Adjust across the declared gene set and comparisons. The adjusted value answers a testing-family question, not the probability that a gene is biologically useful.

Voom converts counts to log-counts per million and estimates a mean–variance trend, producing observation-level precision weights for linear modelling.

Archive the sample manifest, count provenance, feature identifiers, formula, contrast, filters, package version, random seeds and exclusions so another analyst can reconstruct the decision path.

Contents · Previous section · Next section

Section 20 of 36

20. Rank with effect and precision

Examine log fold change, average expression, moderated statistics, uncertainty and sample plots together. A strong statistic can accompany a modest effect when precision is high.

The weights quantify estimated reliability on the transformed scale. They are not probabilities that a gene is true or important.

Keep the biological sample as the unit of replication. More reads or cells can improve measurement, but they do not create more independent people, animals or cultures.

Contents · Previous section · Next section

Section 21 of 36

21. Inspect residuals by sample

Plot residual summaries and sample relationships after fitting. Batch structure or nonlinearity left in residuals can invalidate a simple contrast.

Limma fits gene-wise linear models and moderates variance estimates across genes with empirical Bayes methods, improving stability in small replicated studies.

Pair adjusted evidence with effect direction, uncertainty, sample-level plots and sensitivity checks; a short ranked table is not a complete scientific result.

Contents · Previous section · Next section

Section 22 of 36

22. Check the mean–variance trend again

Review whether weighted residual variation is approximately stabilised. Persistent abundance-dependent patterns may call for a different voom or trend setting.

Contrasts are algebraic questions asked of fitted coefficients. A correct contrast must match the design coding and the biological comparison.

Use negative controls, simulated nulls or label permutations only when their assumptions match the design, and name the particular false signal each check could reveal.

Contents · Previous section · Next section

Section 23 of 36

23. Audit low-count sensitivity

Repeat headline analyses after reasonable filtering changes. Very low-count genes should not dominate scientific interpretation.

Quality weights and observation weights solve different problems: one can downweight a sample, while the other follows the mean-dependent precision pattern.

Create an evidence card naming the question, measured material, statistical unit, model, comparison, result, validation status, alternatives and narrowest defensible claim.

Contents · Previous section · Next section

Section 24 of 36

24. Audit sample influence

Use leave-one-library checks or sample weights to see whether one observation creates the effect. Report influential samples instead of silently removing them.

Weighted residuals, library-level plots and design-rank checks reveal failures that a smooth mean–variance curve alone can miss.

Write association or differential expression when that is what was tested. Reserve cause, mechanism, diagnosis and benefit for designs with stronger supporting evidence.

Contents · Previous section · Next section

Section 25 of 36

25. Model repeated measures appropriately

For repeated observations, use supported correlation or mixed-model extensions and enough subjects. Repeated measurements are not independent replication.

Voom converts counts to log-counts per million and estimates a mean–variance trend, producing observation-level precision weights for linear modelling.

Archive the sample manifest, count provenance, feature identifiers, formula, contrast, filters, package version, random seeds and exclusions so another analyst can reconstruct the decision path.

Contents · Previous section · Next section

Section 26 of 36

26. Separate batch from condition

A design cannot recover a condition effect when batch and condition are perfectly aligned. Narrow the claim or redesign the experiment.

The weights quantify estimated reliability on the transformed scale. They are not probabilities that a gene is true or important.

Keep the biological sample as the unit of replication. More reads or cells can improve measurement, but they do not create more independent people, animals or cultures.

Contents · Previous section · Next section

Section 27 of 36

27. Compare with DESeq2

DESeq2 models counts directly with negative-binomial GLMs; voom uses precision-weighted log-count models. Compare them only with aligned inputs, design and contrasts.

Limma fits gene-wise linear models and moderates variance estimates across genes with empirical Bayes methods, improving stability in small replicated studies.

Pair adjusted evidence with effect direction, uncertainty, sample-level plots and sensitivity checks; a short ranked table is not a complete scientific result.

Contents · Previous section · Next section

Section 28 of 36

28. Build a voom evidence card

Record counts, filters, scaling factors, design, voom variant, weight diagnostics, contrasts, moderation, testing family, sensitivity and biological validation.

Contrasts are algebraic questions asked of fitted coefficients. A correct contrast must match the design coding and the biological comparison.

Use negative controls, simulated nulls or label permutations only when their assumptions match the design, and name the particular false signal each check could reveal.

Contents · Previous section · Next section

Section 29 of 36

29. Connect voom to school science

Repeated measurements do not all deserve equal confidence. Precision weighting formalises the familiar idea that reliable measurements should influence a conclusion more.

Quality weights and observation weights solve different problems: one can downweight a sample, while the other follows the mean-dependent precision pattern.

Create an evidence card naming the question, measured material, statistical unit, model, comparison, result, validation status, alternatives and narrowest defensible claim.

Contents · Previous section · Next section

Section 30 of 36

30. Try a weighted-measurement activity

Combine thermometer readings with different stated uncertainties. Compare an ordinary average with a reliability-weighted summary and discuss when weighting can mislead.

Weighted residuals, library-level plots and design-rank checks reveal failures that a smooth mean–variance curve alone can miss.

Write association or differential expression when that is what was tested. Reserve cause, mechanism, diagnosis and benefit for designs with stronger supporting evidence.

Contents · Previous section · Next section

Section 31 of 36

31. Connect to mathematics

Logarithms, regression, residuals, variance trends, weights and moderated t-statistics show how mathematical models adapt to changing precision.

Voom converts counts to log-counts per million and estimates a mean–variance trend, producing observation-level precision weights for linear modelling.

Archive the sample manifest, count provenance, feature identifiers, formula, contrast, filters, package version, random seeds and exclusions so another analyst can reconstruct the decision path.

Contents · Previous section · Next section

Section 32 of 36

32. Connect to computing and AI literacy

A smooth algorithmic result rests on encoded assumptions. Inspect weights and design columns the way one would inspect features and labels in an AI system.

The weights quantify estimated reliability on the transformed scale. They are not probabilities that a gene is true or important.

Keep the biological sample as the unit of replication. More reads or cells can improve measurement, but they do not create more independent people, animals or cultures.

Contents · Previous section · Next section

Section 33 of 36

33. Connect to school and career pathways

Precision modelling links biology, mathematics, statistics, computing, medicine and data science through multiple education choices.

Limma fits gene-wise linear models and moderates variance estimates across genes with empirical Bayes methods, improving stability in small replicated studies.

Pair adjusted evidence with effect direction, uncertainty, sample-level plots and sensitivity checks; a short ranked table is not a complete scientific result.

Contents · Previous section · Next section

Section 34 of 36

34. Create a family evidence habit

When a gene-expression plot looks decisive, ask how precision was estimated, how many samples contributed and whether one library was downweighted.

Contrasts are algebraic questions asked of fitted coefficients. A correct contrast must match the design coding and the biological comparison.

Use negative controls, simulated nulls or label permutations only when their assumptions match the design, and name the particular false signal each check could reveal.

Contents · Previous section · Next section

Section 35 of 36

35. Use precise limma-voom language

Say the weighted linear model supports a differential-expression estimate under the stated contrast. Avoid claiming a direct regulatory mechanism without further experiments.

Quality weights and observation weights solve different problems: one can downweight a sample, while the other follows the mean-dependent precision pattern.

Create an evidence card naming the question, measured material, statistical unit, model, comparison, result, validation status, alternatives and narrowest defensible claim.

Contents · Previous section · Next section

Section 36 of 36

36. Finish with a weights-to-claim bundle

Deliver counts, metadata, scaling, code, design, voom plots, weights, fits, contrasts, complete tables, sample diagnostics, sensitivity and validation.

Weighted residuals, library-level plots and design-rank checks reveal failures that a smooth mean–variance curve alone can miss.

Write association or differential expression when that is what was tested. Reserve cause, mechanism, diagnosis and benefit for designs with stronger supporting evidence.

## Source date and scope note Primary studies and current official software documentation were checked for this article on 11 October 2026. Software interfaces and recommendations can change, so readers should consult the linked current documentation before reproducing an analysis.

## Final reader checklist Before accepting a claim, confirm the biological sample count, condition definition, measured material, preprocessing, statistical unit, design matrix, exact contrast, effect size, uncertainty, software version, negative controls, sensitivity checks, independent validation and the boundary between association and causation.

Contents · Previous section · Continue to the Science Learning Hub