Community structure

Alpha and beta diversity

Diversity metrics are cheap to compute and easy to report, which is exactly why they are so often reported without a question attached. Each metric answers something specific. Two of them are dominated by sequencing depth. And the standard test for group separation cannot, by itself, tell a real difference from a difference in variability.

Alpha diversity — within a sample

MetricWhat it measuresDepth sensitivityNeeds a tree
Observed featuresHow many distinct ASVs are present. Pure richnessExtreme — sequence deeper, observe more, essentially without limitno
Chao1Richness, extrapolated using the count of rare featuresHigh, and the extrapolation is unstable at low depthno
ShannonRichness weighted by evenness; dominated by mid-abundance taxaModerate — saturates far earlier than richnessno
Simpson / inverse SimpsonProbability two reads come from different taxa; driven by the dominant fewLowno
Pielou's evennessEvenness alone, with richness divided outLow–moderateno
Faith's PDTotal branch length spanned. Richness that counts how different the taxa areHigh, like richnessyes

Report at least one richness-type and one evenness-weighted metric, and say which question each is answering. "Alpha diversity was higher in the treated group" is not a finding until you say whether that means more taxa, or a less dominated community — those have different biology behind them and can move in opposite directions.

The depth problem, and the argument about it

sequencing depth → alpha diversity observed features still climbing at the right edge 66k 174k same community, different answer Shannon — saturates McMurdie & Holmes 2014 "rarefying microbiome data is inadmissible" — and for differential abundance testing, right Schloss 2023 · reanalysis most robust control for uneven effort — for diversity and ordination split by task diversity + ordination → rarefy (repeatedly) differential abundance → model depth, don't rarefy

Choosing a depth is choosing how many samples to keep

The real 50-sample run: final depths run from 66,144 to 173,626 reads, median 125,757. Every candidate rarefaction depth trades reads against samples, and the trade is steep in the middle.

Samples retained by rarefaction depth

50 samples, non-chimeric reads. Hover any bar for the exact count.

Reading it: anywhere up to 90,000 costs you three samples. Past 100,000 the curve falls off a cliff — 120,000 costs you nearly two in five. The defensible choice here is around 90,000, and the defensible method is to look at this curve and say why, rather than to accept whatever the tutorial used. Two further points worth making explicit: the samples you drop are not a random subset — they are the ones that sequenced badly, which often correlates with biomass or inhibitor content and therefore with your groups; and a rarefaction depth chosen after seeing which depth gives a significant result is a researcher degree of freedom like any other.

Beta diversity — between samples

MetricUsesMost sensitive toPick it when
Jaccardpresence / absencerare taxa; sequencing depthyou care about membership, not amounts
Bray–Curtisabundancesabundant taxathe default abundance-weighted choice, no tree needed
Unweighted UniFracpresence / absence + treerare and deep-branching lineagesyou expect whole clades to be gained or lost
Weighted UniFracabundances + treeabundant taxa, discounted by relatednessyou expect shifts in the dominant members
AitchisonCLR-transformed abundancesratios between taxayou want a distance that treats the data as compositional
Robust AitchisonCLR with zeros handled by matrix completionas above, tolerant of sparsitythe table is very sparse, which it usually is

Running all six and reporting the one that separated is the most common way to get a false positive out of a beta diversity analysis. Pick the metric from the hypothesis — presence-based if you expect taxa to appear or vanish, abundance-based if you expect them to shift — and if you show more than one, show all of them, including the ones that did not separate.

PERMANOVA cannot tell you what you think it does

A · centroids differ shift PERMANOVA p < 0.001 · PERMDISP p = 0.6 "composition differed by group" — supported B · same centroid, different spread PERMANOVA p < 0.001 · PERMDISP p < 0.001 the same PERMANOVA output as panel A Read them as a pair PERMANOVA sig · PERMDISP ns location shift — the clean result both significant ambiguous — report both, claim neither alone PERMANOVA ns no evidence, however good the PCoA looks Unbalanced group sizes make PERMANOVA more sensitive to dispersion, not less.