Why Science? | ALDEx2, Compositional Counts and Log-Ratio Evidence

Section 1 of 36

1. Begin with the scientific unit

Start by identifying the independently sampled unit and the feature table. Compositional reasoning cannot rescue pseudoreplication or ambiguous provenance.

Sequencing counts are compositional: each feature is observed relative to a library total, so a rise in one component can make unchanged components appear to fall.

Archive raw-data accessions, sample metadata, annotations, software versions, commands, filters and random seeds so another analyst can reconstruct every decision.

Contents · Next section

Section 2 of 36

2. Separate measurement from biology

A library reports parts of a constrained whole. Without an external scale, absolute increases and relative reallocations can generate the same observed proportions.

ALDEx2 samples Dirichlet Monte Carlo instances and transforms each instance into log-ratios. The output summarises uncertainty across many plausible compositions.

Count independently collected organisms, people, cultures or specimens as the biological replicates; computational resamples and repeated reads do not create new experimental units.

Contents · Previous section · Next section

Section 3 of 36

3. Write the comparison first

Specify groups, covariates, denominator strategy and whether the aim is differential relative abundance, not a vague search for important microbes or genes.

The denominator defines the reference for log-ratios. Using all features, robust subsets or external references expresses different assumptions and must be documented.

Pair adjusted evidence with effect direction, uncertainty, sample-level plots, diagnostics and at least one sensitivity analysis capable of contradicting the headline.

Contents · Previous section · Next section

Section 4 of 36

4. Meet the uncertainty

Finite counts create sampling uncertainty, and biological samples vary within groups. ALDEx2 carries both ideas through Monte Carlo log-ratio summaries.

The current vignette cautions that centred log-ratios are scale invariant but can mislead when scale differs; gamma-based scale modelling makes that uncertainty explicit.

Declare the testing family, comparisons and exclusions before selecting examples. Multiple testing belongs to the complete decision set, not merely the displayed rows.

Contents · Previous section · Next section

Section 5 of 36

5. Identify the central assumption

Every log-ratio needs a reference. The denominator and assumed scale determine what unchanged means, even when software supplies a convenient default.

The ALDEx2 effect statistic combines a between-group difference with within-group dispersion across Monte Carlo instances; it complements, rather than duplicates, an adjusted test result.

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

Contents · Previous section · Next section

Section 6 of 36

6. Keep the claim bounded

Differential relative abundance does not by itself show absolute growth, ecological interaction, disease causation or treatment benefit.

Zero counts need probabilistic handling because logarithms of zero are undefined. A Dirichlet model represents uncertainty; it does not prove a feature is biologically present or absent.

Use association, differential abundance or differential expression when that is what was estimated; reserve cause, mechanism, diagnosis and benefit for stronger evidence.

Contents · Previous section · Next section

Section 7 of 36

7. Read the primary evidence in scope

Fernandes and colleagues framed count analysis compositionally and evaluated ALDEx2. The paper supports the approach under studied settings, not universal superiority.

Sequencing counts are compositional: each feature is observed relative to a library total, so a rise in one component can make unchanged components appear to fall.

Archive raw-data accessions, sample metadata, annotations, software versions, commands, filters and random seeds so another analyst can reconstruct every decision.

Contents · Previous section · Next section

Section 8 of 36

8. Build a trustworthy input matrix

Provide non-negative counts with consistent feature identifiers and documented preprocessing. Do not mix percentages, normalised values and raw counts in one matrix.

ALDEx2 samples Dirichlet Monte Carlo instances and transforms each instance into log-ratios. The output summarises uncertainty across many plausible compositions.

Count independently collected organisms, people, cultures or specimens as the biological replicates; computational resamples and repeated reads do not create new experimental units.

Contents · Previous section · Next section

Section 9 of 36

9. Create a sample sheet

Record group, batch, subject, pairing, extraction run and other covariates. Verify that columns and sample rows match one-to-one.

The denominator defines the reference for log-ratios. Using all features, robust subsets or external references expresses different assumptions and must be documented.

Pair adjusted evidence with effect direction, uncertainty, sample-level plots, diagnostics and at least one sensitivity analysis capable of contradicting the headline.

Contents · Previous section · Next section

Section 10 of 36

10. Inspect library composition

Inspect library sizes, zeros, dominant features and prevalence. Extreme compositions can shift many log-ratios even when most absolute abundances are stable.

The current vignette cautions that centred log-ratios are scale invariant but can mislead when scale differs; gamma-based scale modelling makes that uncertainty explicit.

Declare the testing family, comparisons and exclusions before selecting examples. Multiple testing belongs to the complete decision set, not merely the displayed rows.

Contents · Previous section · Next section

Section 11 of 36

11. Check sample relationships

Ordination and distance plots can reveal group and batch patterns, but the distance and transformation used must be named.

The ALDEx2 effect statistic combines a between-group difference with within-group dispersion across Monte Carlo instances; it complements, rather than duplicates, an adjusted test result.

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

Contents · Previous section · Next section

Section 12 of 36

12. Filter by information

Use a prespecified prevalence or abundance rule suited to the question. Filtering changes the reference set and therefore can change log-ratios.

Zero counts need probabilistic handling because logarithms of zero are undefined. A Dirichlet model represents uncertainty; it does not prove a feature is biologically present or absent.

Use association, differential abundance or differential expression when that is what was estimated; reserve cause, mechanism, diagnosis and benefit for stronger evidence.

Contents · Previous section · Next section

Section 13 of 36

13. Confirm the design is estimable

Check that each group and covariate combination has replication and that pairing identifiers are correct. Confounding remains confounding after transformation.

Sequencing counts are compositional: each feature is observed relative to a library total, so a rise in one component can make unchanged components appear to fall.

Archive raw-data accessions, sample metadata, annotations, software versions, commands, filters and random seeds so another analyst can reconstruct every decision.

Contents · Previous section · Next section

Section 14 of 36

14. Choose normalisation deliberately

In ALDEx2, normalisation is expressed through log-ratio denominators and scale assumptions. Choose these scientifically, not by discovery count.

ALDEx2 samples Dirichlet Monte Carlo instances and transforms each instance into log-ratios. The output summarises uncertainty across many plausible compositions.

Count independently collected organisms, people, cultures or specimens as the biological replicates; computational resamples and repeated reads do not create new experimental units.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with a fictional evidence table

This classroom table is invented and is not a result from the ALDEx2 study.

The denominator defines the reference for log-ratios. Using all features, robust subsets or external references expresses different assumptions and must be documented.

Pair adjusted evidence with effect direction, uncertainty, sample-level plots, diagnostics and at least one sensitivity analysis capable of contradicting the headline.

Contents · Previous section · Next section

Section 16 of 36

16. Run the method with recorded settings

Set a justified Monte Carlo sample count, denominator and test path. For stable final estimates, the official vignette recommends more instances than quick exploration.

The current vignette cautions that centred log-ratios are scale invariant but can mislead when scale differs; gamma-based scale modelling makes that uncertainty explicit.

Declare the testing family, comparisons and exclusions before selecting examples. Multiple testing belongs to the complete decision set, not merely the displayed rows.

Contents · Previous section · Next section

Section 17 of 36

17. Interpret the central statistic

Read the median between-group difference, within-group dispersion and effect statistic together. Each summarises a different part of the evidence.

The ALDEx2 effect statistic combines a between-group difference with within-group dispersion across Monte Carlo instances; it complements, rather than duplicates, an adjusted test result.

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

Contents · Previous section · Next section

Section 18 of 36

18. Correct for many features

Report the expected adjusted values across all tested features and planned contrasts, not just the lowest raw p-values.

Zero counts need probabilistic handling because logarithms of zero are undefined. A Dirichlet model represents uncertainty; it does not prove a feature is biologically present or absent.

Use association, differential abundance or differential expression when that is what was estimated; reserve cause, mechanism, diagnosis and benefit for stronger evidence.

Contents · Previous section · Next section

Section 19 of 36

19. Read effects beside probabilities

An effect describes separation relative to dispersion across Monte Carlo instances; it should be shown beside uncertainty and sample-level log-ratios.

Sequencing counts are compositional: each feature is observed relative to a library total, so a rise in one component can make unchanged components appear to fall.

Archive raw-data accessions, sample metadata, annotations, software versions, commands, filters and random seeds so another analyst can reconstruct every decision.

Contents · Previous section · Next section

Section 20 of 36

20. Plot every biological sample

Plot each sample after the chosen log-ratio transformation. Group means alone can conceal outliers, zero-heavy features and inconsistent directions.

ALDEx2 samples Dirichlet Monte Carlo instances and transforms each instance into log-ratios. The output summarises uncertainty across many plausible compositions.

Count independently collected organisms, people, cultures or specimens as the biological replicates; computational resamples and repeated reads do not create new experimental units.

Contents · Previous section · Next section

Section 21 of 36

21. Inspect the method-specific diagnostic

Inspect Monte Carlo stability and denominator behaviour. If a feature changes classification across instances, that uncertainty belongs in the narrative.

The denominator defines the reference for log-ratios. Using all features, robust subsets or external references expresses different assumptions and must be documented.

Pair adjusted evidence with effect direction, uncertainty, sample-level plots, diagnostics and at least one sensitivity analysis capable of contradicting the headline.

Contents · Previous section · Next section

Section 22 of 36

22. Challenge the key assumption

Ask whether the reference set is plausibly stable and whether absolute scale could differ between groups. Neither assumption is guaranteed by a large table.

The current vignette cautions that centred log-ratios are scale invariant but can mislead when scale differs; gamma-based scale modelling makes that uncertainty explicit.

Declare the testing family, comparisons and exclusions before selecting examples. Multiple testing belongs to the complete decision set, not merely the displayed rows.

Contents · Previous section · Next section

Section 23 of 36

23. Test filtering sensitivity

Repeat reasonable prevalence thresholds and note how the denominator and retained feature universe change.

The ALDEx2 effect statistic combines a between-group difference with within-group dispersion across Monte Carlo instances; it complements, rather than duplicates, an adjusted test result.

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

Contents · Previous section · Next section

Section 24 of 36

24. Test normalisation sensitivity

Compare justified denominators, such as all features and a robust reference subset, while avoiding a post-hoc search for desired answers.

Zero counts need probabilistic handling because logarithms of zero are undefined. A Dirichlet model represents uncertainty; it does not prove a feature is biologically present or absent.

Use association, differential abundance or differential expression when that is what was estimated; reserve cause, mechanism, diagnosis and benefit for stronger evidence.

Contents · Previous section · Next section

Section 25 of 36

25. Test sample influence

Leave out one independent sample at a time where possible and record effect-direction and ranking changes.

Sequencing counts are compositional: each feature is observed relative to a library total, so a rise in one component can make unchanged components appear to fall.

Archive raw-data accessions, sample metadata, annotations, software versions, commands, filters and random seeds so another analyst can reconstruct every decision.

Contents · Previous section · Next section

Section 26 of 36

26. Test parameter sensitivity

Increase Monte Carlo instances and explore the vignette’s modest gamma scale settings when scientifically warranted; instability is a result to report.

ALDEx2 samples Dirichlet Monte Carlo instances and transforms each instance into log-ratios. The output summarises uncertainty across many plausible compositions.

Count independently collected organisms, people, cultures or specimens as the biological replicates; computational resamples and repeated reads do not create new experimental units.

Contents · Previous section · Next section

Section 27 of 36

27. Compare an adjacent method

Contrast ALDEx2 with count-model tools only after acknowledging that absolute-count coefficients and compositional log-ratios answer different questions.

The denominator defines the reference for log-ratios. Using all features, robust subsets or external references expresses different assumptions and must be documented.

Pair adjusted evidence with effect direction, uncertainty, sample-level plots, diagnostics and at least one sensitivity analysis capable of contradicting the headline.

Contents · Previous section · Next section

Section 28 of 36

28. Build an evidence card

Record count provenance, denominator, Monte Carlo count, gamma or scale choice, tests, effects, plots, sensitivity, versions and validation.

The current vignette cautions that centred log-ratios are scale invariant but can mislead when scale differs; gamma-based scale modelling makes that uncertainty explicit.

Declare the testing family, comparisons and exclusions before selecting examples. Multiple testing belongs to the complete decision set, not merely the displayed rows.

Contents · Previous section · Next section

Section 29 of 36

29. Connect to fair testing

A fair comparison defines its reference and measurement scale before looking at results. That is the compositional counterpart of controlling variables.

The ALDEx2 effect statistic combines a between-group difference with within-group dispersion across Monte Carlo instances; it complements, rather than duplicates, an adjusted test result.

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

Contents · Previous section · Next section

Section 30 of 36

30. Try a classroom activity

Give students coloured counters in fixed-size jars. Increase one colour and observe the apparent proportion decrease in every unchanged colour.

Zero counts need probabilistic handling because logarithms of zero are undefined. A Dirichlet model represents uncertainty; it does not prove a feature is biologically present or absent.

Use association, differential abundance or differential expression when that is what was estimated; reserve cause, mechanism, diagnosis and benefit for stronger evidence.

Contents · Previous section · Next section

Section 31 of 36

31. Connect to mathematics

Logarithms convert ratios into differences; probability distributions represent uncertain compositions; medians and dispersion summarise repeated plausible values.

Sequencing counts are compositional: each feature is observed relative to a library total, so a rise in one component can make unchanged components appear to fall.

Archive raw-data accessions, sample metadata, annotations, software versions, commands, filters and random seeds so another analyst can reconstruct every decision.

Contents · Previous section · Next section

Section 32 of 36

32. Connect to computing literacy

Seeds, parallel settings and package versions affect reproducibility. Scripts should expose the denominator and scale choices as first-class parameters.

ALDEx2 samples Dirichlet Monte Carlo instances and transforms each instance into log-ratios. The output summarises uncertainty across many plausible compositions.

Count independently collected organisms, people, cultures or specimens as the biological replicates; computational resamples and repeated reads do not create new experimental units.

Contents · Previous section · Next section

Section 33 of 36

33. Connect to education pathways

Compositional data appear in microbiomes, ecology and sequencing, connecting school biology and mathematics with biostatistics and health research.

The denominator defines the reference for log-ratios. Using all features, robust subsets or external references expresses different assumptions and must be documented.

Pair adjusted evidence with effect direction, uncertainty, sample-level plots, diagnostics and at least one sensitivity analysis capable of contradicting the headline.

Contents · Previous section · Next section

Section 34 of 36

34. Create a family question habit

When a proportion changes, ask whether the component changed absolutely, the total changed, or another component moved the denominator.

The current vignette cautions that centred log-ratios are scale invariant but can mislead when scale differs; gamma-based scale modelling makes that uncertainty explicit.

Declare the testing family, comparisons and exclusions before selecting examples. Multiple testing belongs to the complete decision set, not merely the displayed rows.

Contents · Previous section · Next section

Section 35 of 36

35. Use precise scientific language

Say that ALDEx2 supports a difference in relative abundance under a stated log-ratio reference and scale model. Do not imply an absolute count change.

The ALDEx2 effect statistic combines a between-group difference with within-group dispersion across Monte Carlo instances; it complements, rather than duplicates, an adjusted test result.

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

Contents · Previous section · Next section

Section 36 of 36

36. Finish with an evidence bundle

Deliver counts, metadata, reference choice, Monte Carlo settings, complete tables, transformed sample plots, sensitivity analyses, versions and bounded claims.

Zero counts need probabilistic handling because logarithms of zero are undefined. A Dirichlet model represents uncertainty; it does not prove a feature is biologically present or absent.

Use association, differential abundance or differential expression when that is what was estimated; reserve cause, mechanism, diagnosis and benefit for stronger evidence.

## Source date and scope note Primary studies and current official software documentation were checked for this article on 11 October 2026. Package interfaces, recommendations and dependencies can change; readers should consult the linked official package page and vignette before reproduction.

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

Contents · Previous section · Continue to the Science Learning Hub