Sampling
One regeneratively managed pasture, sampled once. Soil was taken from beneath five kinds of plant growing in that same field — chicory, fescue, ragweed, red clover and white clover — giving samples in all.
The five plants are the comparison. Because every sample comes from one field under one management history, nothing here can be attributed to regenerative management: there is no differently managed field to compare against. What the design can see is whether the plant growing above changes the community below.
One of the white-clover samples came from a compacted vehicle track. It is kept and reported as an incidental observation, not treated as a designed contrast — compaction changes soil profoundly, and folding it in silently would be smuggling a second experiment into the first.
In the lab
DNA was extracted from each sample as a mixture — everything present, plant and animal and microbe together. Two short marker genes were then copied (amplified) out of that mixture and sequenced. A marker gene is a stretch of DNA that every organism in a group carries and that differs just enough between them to tell species apart, so copying one is how you make a census of one group and ignore everything else.
| Marker | What it reads | Region | Forward primer | Reverse primer | Reference database |
|---|
samples were sequenced for the bacterial marker and for the eukaryotic one, which is why the food-web results on the home page cover fewer patches than the bacterial ones.
The letters in those primer sequences that are not A, C, G or T are deliberate:
M means "A or C", H means "anything but G". A primer
is made degenerate on purpose so that one ordered molecule matches the slightly
different versions of the same gene carried by organisms whose last common
ancestor lived billions of years ago. It is also why a primer cannot be found
in a read with a plain text search, which matters for
verifying public data further down.
No fungal ITS marker was ever sequenced. A folder delivered under that name contains the same sequences as the 18S run, matched against a fungus-only database — which forces every read, including protists and animals, onto a fungal name. Those outputs are artifacts and are not used. Fungi are described here only at the coarse level 18S actually supports.
The pipeline
Raw sequencing output is millions of short, error-prone reads. Turning them into a table of organisms and counts is done in , the standard toolkit for this kind of survey, in three steps:
- Denoising. models the machine's error pattern and corrects reads back to the exact sequences that were really there, rather than lumping similar reads into approximate groups. The result is called an ASV — one distinct sequence, treated as one kind of organism. The older convention was to cluster reads at 97% similarity and call each cluster a species, which both merges organisms that differ and invents ones that do not exist; it also produces clusters that cannot be compared between studies, where an exact sequence can.
- Classifying. Each ASV is compared against a curated reference library with a classifier, which reports the most specific name the evidence supports — often a genus, sometimes only a phylum, sometimes nothing. Two consequences run through everything below: a fragment this short frequently cannot resolve a species, so genus is the working unit of this study, and the unnameable fraction is not noise to be dropped but soil we cannot identify.
- Analysis. Everything after that — diversity, ordination, the guild abundances, every figure and every number on this site — runs in .
Every tool named here, both reference databases, and each statistical test used below are listed with their references on the sources page, alongside what each one is used for in this study.
What came out
distinct bacterial and archaeal sequences across samples, at between and reads per sample.
A table like that is compositional, and everything downstream is shaped by it. A sequencing run returns a roughly fixed budget of reads, so what comes out are shares of a total that has nothing to do with how much life was in the ground. If one group doubles, every other group must appear to shrink even though nothing happened to it. This is not a technicality: it is the reason the clover result on the home page needed a section of stress tests rather than a correlation coefficient, and the reason one of those tests re-runs the whole thing on log-ratios instead of percentages.
Equalising sequencing depth
The deepest sample was read times as deeply as the shallowest. A sample read more deeply shows more kinds of organism for no biological reason at all, so comparing diversity across samples of unequal depth partly measures the sequencing run rather than the soil.
Diversity is therefore computed on tables subsampled down to a common depth of reads — the shallowest sample's own count, which is the deepest level at which no sample has to be discarded. The subsampling uses the multivariate hypergeometric distribution, which is exact sampling without replacement rather than the multinomial approximation usually used for it, and the whole draw is repeated times and averaged so that no reported value depends on one lucky shuffle.
The check that did not come out the way it usually does
Rarefaction is meant to break the link between how deeply a sample was read and how rich it looks. Here it barely moves it: richness tracks the original read depth at before subsampling and still after. The honest reading is that the deeply sequenced samples were genuinely richer rather than merely better sampled — depth and richness are correlated in this soil, not only in the machine. It is reported because a correction that changes nothing is itself evidence, and leaving it out would imply the confound had been removed.
Is there as much life under one plant as another?
Three metrics, because "how much life" has three defensible meanings and a result that holds under one but not the others is a result about the metric. Each is tested across the five plants with a Kruskal–Wallis test, the rank-based alternative to ANOVA — chosen because with two or three cores per plant there is no way to establish that anything is normally distributed.
| Metric | What it counts | H | df | p |
|---|
Reporting all three matters in a specific way. A community can hold a great many kinds of organism that are all close relatives; richness and Shannon would call that diverse, and Faith's phylogenetic diversity — which measures the total length of evolutionary branch present in a sample — would not. Finding every plant equal under all three is a stronger statement than finding them equal under any one of them.
A flat result is not proof of equality. With cores across five groups these tests could only have detected a large difference. "No difference was detected" is what they say; "the plants host equal diversity" is more than this design can support.
Is it the same life?
A different question, and the one carrying the home page's result. It is answered by measuring how different every pair of samples is, then asking whether cores under the same plant are more alike than chance would make them — a test called PERMANOVA, run over reshufflings of the plant labels with the random seed fixed at so that a reviewer reproduces the exact p-value rather than one near it.
Averaged over every pair, two cores under the same plant differ by and two cores under different plants by (Bray–Curtis, where 0 is identical and 1 shares nothing at all). That is same-plant pairs against different-plant pairs: a real gap, and a narrow one, which is exactly why it needs a test rather than a sentence.
| Distance measure | What it counts | pseudo-F | R² | null | excess | p |
|---|
Why the raw R² is not the number
PERMANOVA's R² is routinely read as "the fraction of variation the grouping explains", which invites reading as most of it. That is wrong. R² here is a ratio of sums of squares whose expected value under a meaningless grouping is not zero but roughly (k − 1) / (n − 1) — with five plants and cores, — because within-group sums run over far fewer pairs than the total does.
So it is compared against itself. Shuffling the plant labels times puts the empirical null mean at , within a thousandth of that analytic floor, and its 95th percentile at . The observed value sits above that percentile, and what gets reported is the of it chance does not account for (p = by that effect-size test, agreeing closely with the pseudo-F permutation p-value of ). The gap between the raw figure and the excess is the single most common way a study of this size is oversold.
A significant PERMANOVA has one notorious alternative explanation: groups can come out "different" merely because one is more internally variable than another, and the test cannot tell that from a genuine difference in location. A separate test — PERMDISP, over permutations — compares each sample's distance to its own group's centre (F = , p = ) and finds no evidence of it. So the plants differ in where their communities sit and not only in how scattered they are. This test is weak at this sample size as well: a null result is reassurance, not proof.
The phylogeny, and why two of those rows exist
Bray–Curtis treats every ASV as equally unrelated to every other. Two samples that differ only in which strain of Nitrososphaera they carry look exactly as different, under that measure, as two samples that differ by an entire phylum. UniFrac repairs this by weighting each difference by the length of evolutionary branch separating the taxa, and Faith's phylogenetic diversity does the same for the single-sample metrics. Agreement between the two families is evidence that a result is about community biology rather than about a choice of arithmetic.
That needs a tree, and the 16S run shipped one — which no analysis in
this project had ever opened, for a mundane reason. The exported count table
had been renamed to ASV00001-style identifiers while the tree kept
the pipeline's original sequence hashes, so the two files shared not one label
and any straightforward join produced an empty table with no error raised. A
stored mapping between the two naming schemes is the bridge; the loader applies
it, refuses to continue if the join comes back empty rather than reporting
zeros, rarefies to reads, then trims the
tree down to exactly the sequences that survived.
Rarefying first is not optional here. Unweighted UniFrac and Faith's PD both count rare branches, and a more deeply read sample is simply more likely to have stumbled onto one, so an uncorrected comparison across a × depth range would partly measure the sequencing run. The two UniFrac variants answer different questions, which is why both are in the table above: the weighted one is dominated by which organisms are abundant, the unweighted one by which rare lineages are present at all. The weighted variant gives the largest excess R² of the three measures and the unweighted the smallest — the honest shape of a result driven by shifting abundances among common taxa rather than by wholesale replacement of rare ones.
Putting names to jobs
Sequencing says who is present, not what they are doing. To get from names to nitrogen, genera known to perform a given job are listed in advance and their combined share of the reads is used as a proxy for that job's capacity: genera of ammonia-oxidising archaea, of ammonia-oxidising bacteria, nitrite-oxidisers and nitrogen-fixing genera.
Matching is done case-insensitively against the genus field of the reference lineage, as a substring, so one list entry absorbs species-level suffixes and the reference database's own naming variants. The nitrogen-fixer list is deliberately the strict one: canonical fixers only. An earlier version included the broad facultative candidates — Methylobacterium, cyanobacteria, Herbaspirillum and similar — which inflated the fixer estimate several-fold. Those are excluded, and the figure on the home page is the conservative one.
These lists live in one file in the analysis code, and adding or removing a genus changes a published number, so the test suite pins the results they produce. Two limits come with the approach and are stated wherever the result appears: a gene being present is not the same as a gene being switched on, so this measures potential; and read abundance is not cell count, which gets a check of its own below.
One list entry was wrong for a year and nobody noticed, which is worth recording rather than quietly fixing. The reference database spells one genuine ammonia-oxidising genus Nitrocosmicus; the list looked for Nitrosocosmicus, with an extra syllable. The two never matched, and a real ammonia oxidiser worth 0.27% of all reads was dropped from every archaeal count. Correcting it moved the home page's headline ratio from 143 to 148. The lesson is not that substring matching is bad — it is that a list which silently returns nothing looks exactly like a list which correctly returns nothing. The independent database described just below agrees the correction was right: our count moved from 8.53% to 8.79% of reads, and that database, which knows nothing of our lists, says 8.97%. The fix moved us towards it.
The same table, annotated by strangers
As an independent check the table is re-annotated with FAPROTAX , a published database built by other people, from other literature, that knows nothing of our lists. It puts ammonia oxidation at % of reads against our hand-curated %, and nitrogen fixation at % against our %.
Coverage is the limit of that agreement, so it is stated too: the database can say something about % of reads on average, ranging from % to % across samples, and recovers of its functional groups in this soil. Two unrelated methods agreeing about two-thirds of the reads is corroboration; it is not corroboration about the other third. Separately, of functional groups compared between microsites, survived correction for having asked that many questions at once.
Sorting the archaea by the acidity they can tolerate
Ammonia-oxidising archaea are not one kind of organism with one preference. One lineage, Nitrosotalea and its relatives, is an obligate acidophile: it grows between about pH 4.0 and 5.5 and not outside that. Another, Nitrososphaera and Nitrosocosmicus, prefers near-neutral soil and is the group usually found around plant roots. Since public soil maps place this site near pH 5.5, which lineage dominates is a test of whether acidity explains the headline — and it is a test the existing reads can run.
The split is counted at order rank, not genus, and that choice is doing real work. About 92% of the archaeal reads here resolve no further than the family Nitrososphaeraceae: at genus rank most of them are simply blank, and a genus-level count would discard them. The acidophile boundary happens to fall one rank higher — Nitrosotaleales against Nitrososphaerales — which the reads do resolve. The question is answerable precisely because the distinction that matters is coarser than the distinction the data cannot make.
Both lineages are counted the same way as each other, which is what a comparison requires. Their combined total is slightly larger than the genus-strict archaeal figure quoted elsewhere, because counting at order rank also picks up the family-only reads that have no genus for the guild list to match. The two numbers are not interchangeable and are not presented as though they were.
Two limits travel with the result wherever it is quoted. Which organisms are present is not how active they are, and a neutral-soil lineage can persist in acid soil while doing less work. And the pH itself is modelled from public maps, not measured in this field.
The third ammonia oxidiser this study cannot see
The headline ratio counts archaea against bacteria, which was for a long time the whole story. It is not: some Nitrospira bacteria carry out the complete oxidation of ammonia to nitrate on their own, and they have a higher affinity for ammonia than either of the two groups in the ratio — meaning they compete best in exactly the low-ammonia soil this study infers.
They cannot be separated here. Telling a complete oxidiser from an ordinary nitrite-oxidising Nitrospira requires reading the ammonia-processing gene itself, and this study reads a different gene. So the total Nitrospira share is reported as an upper bound on the complete oxidisers rather than a measurement of them, and the headline is described as a comparison between the two canonical ammonia-oxidising guilds rather than between everything that oxidises ammonia.
Reads are not cells
The home page's headline is a ratio of read shares, and there is a dull reason a ratio of read shares can mislead: a genome carries anywhere from one to roughly fifteen copies of the 16S gene, and fast-growing bacteria tend to carry more of them. A group with more copies per cell contributes more reads at the same number of cells. If the archaea-to-bacteria ratio were a copy-number artifact, this is the shape it would take.
So each sample's community-weighted mean copy number was computed from , a published per-taxon copy-number database, covering % of reads. It ranges from to copies per genome across the thirteen cores, so there was real variation for the check to find.
It found nothing. Copy number does not track the archaeal share (ρ = ), does not differ between legume and non-legume cores (p = ), and shows no gradient along the community's dominant axis. The ratio is not a copy-number effect. Reads still are not cells, and the page says so wherever the number appears.
Testing whether the clover result is really about nitrogen
Ammonia-oxidising archaea are scarcer under clover, and clover fixes nitrogen, so the tidy story is that leaked nitrogen leaves the archaea less work to do. The correlation is strong (ρ = , Spearman's rank correlation, as everywhere ρ appears on these pages: it ranks the cores rather than using their values, so one unusual core cannot manufacture a slope). Three cheap explanations can produce a number like that with no mechanism behind it, and each one gets a test.
- In a fixed total, everything correlates with everything. The archaeal share spans % to % of reads here, so it is a large and highly variable slice of that fixed total, and almost any group that does not rise with it is pushed the other way by arithmetic alone. The fixers' correlation is therefore ranked against the same correlation computed for every reasonably abundant genus: all that average at least % of reads and appear in at least of the cores. The nitrogen-fixers rank , and unrelated genera track the archaea at least as strongly as . This is the test the tidy story fails, and it is why the home page says we cannot yet say why.
- The fixed-total artifact specifically. The whole comparison is re-run on centred log-ratio abundances, which are free of the sum-to-a-constant constraint (zeros replaced by half the smallest observed non-zero value before taking logarithms). The association survives (ρ = , p = ), with genera still past the threshold. Compositional closure did not manufacture it, then: the association is real. It is its specificity that fails, which is a different and more interesting problem.
- A two-group difference wearing a gradient's clothes. The cores split into legume and non-legume, and a correlation across all thirteen looks like a smooth slope even when it is really a single step between two clusters. So it is re-tested inside each group, where no step between the groups can contribute.
| Group | Cores | ρ | p |
|---|
The two-group contrast itself is significant (p = , Mann–Whitney, which like Kruskal–Wallis assumes nothing about the shape of the distribution). Wherever many things are tested at once — every genus against an axis, every functional group between microsites — p-values are adjusted by the false-discovery procedure, implemented in this repository rather than imported, and pinned by its own tests.
What the archaea are actually tracking
Turning that negative result over yields a positive one. If many unrelated genera all follow the archaeal share, the likely explanation is that the whole community varies along one dominant direction and the archaeal share is a readout of position along it. That is directly checkable: correlate the archaeal share against each of the leading axes of the ordination — the same principal-coordinates projection the home page plots, which turns the table of pairwise dissimilarities into positions.
| Axis | Share of all variation | ρ | p |
|---|
The archaea ride the community's single dominant gradient, which by itself carries % of the variation between cores. They are not responding to one neighbour; they move with everything at once. What sets that gradient is the open question this study ends on.
Other people's data
The cross-farm comparison on the home page reprocesses a public dataset from raw reads through this same pipeline — the same denoiser, the same reference database, the same guild definitions — rather than comparing published percentages. Numbers from two different pipelines are not comparable, and quoting them side by side is a mistake this project has had to correct in inherited work.
The two studies' tables are then aligned by collapsing both to genus, because reference lineages still differ above genus even after identical reprocessing. Phylum-level figures have to be summed from the uncollapsed lineages instead, since collapsing to genus discards the phylum prefix. The comparison's control is internal: organic versus conventional management within the published study, where the wet lab and the sequencing run are shared and no cross-study batch effect can contribute to the difference.
Checking what a public file actually contains
Public metadata records only "amplicon". Accessions are therefore verified before use by reading the primer still attached to the submitted reads: stream the head of several runs off the archive mirror — a bounded read, so whole files never download — and match the starts of reads against known primer sequences. Three complications each caused a false negative in an earlier version and are now handled explicitly. Primers are searched anywhere within the leading bases, because libraries carry barcodes and spacers ahead of the primer. Both members of a pair are searched in both read files, because single-end submissions are sometimes the reverse read. And where a submitter stripped the primers off, 16S V4 is still recognisable from the conserved motif that opens the region, reported as a weaker call.
Sampling one run is not enough, which is the lesson of the accession that forced all this. It was registered as samples of bacterial 16S and ranked as the best available public test of the archaea-to-bacteria ratio. Read by their primers, those runs hold different markers:
| Marker on the reads | Runs | Share |
|---|
Only four runs are bacterial, and they are the wrong four: every unfertilised control was deposited under a different marker, so the nitrogen gradient that made the study attractive cannot be tested there at all. Sampling a single run would have called the whole project fungal, which is wrong in a way just as costly as trusting the registry was. The per-run table is committed to the repository, and the verifier exits non-zero on a mismatch so that it can gate the pipeline rather than merely warn.
Where this soil is, according to public maps
None of this is a measurement of our thirteen cores. It is what published map products say about this location, which is a prior about the neighbourhood. Our own laboratory pH is still pending. The distinction matters, because "pH was not measured" and "pH was not measured, and two independent surveys put it near " are different states of knowledge, and the second is the one we are in.
Soil pH drives the archaea-to-bacteria balance nearly as strongly as nitrogen does, which makes an unmeasured pH the largest single alternative explanation for this study's headline. Two independent sources were queried at the farm point and agree with each other: a global machine-learned soil model puts surface pH at (±), and the national soil survey's mapped components give a depth-weighted — a consensus of , which is classed . Acidic soil favours archaeal ammonia oxidisers, so this is a live competing explanation for the ratio, not a footnote to it.
The same sources put the surface at % clay, % silt and % sand, with g/kg organic carbon, g/kg total nitrogen and a cation-exchange capacity of . The survey does not map this ground as one uniform soil either: it maps components with different properties, including different pH.
| Component | Share | Drainage | Surface pH | Clay % | Organic matter % |
|---|
Two things are deliberately not on this page. The farm's coordinates, and the survey's names for the map unit and its component soil series. A named series is mapped over a small area, so publishing one would narrow the location to roughly a county — and this is an explainer written for the farmer whose field it is, not a record of where the field is. The pH estimates above are the part that bears on the science, and they are all here. The full survey record, names included, is in the repository for the analysis to use.
Reproducibility
Every random draw in the analysis — the rarefaction subsampling, both PERMANOVA permutation tests, the R² null, PERMDISP — is seeded at . That is deliberate, and it has a cost worth naming: a permutation p-value genuinely does vary from run to run, and fixing the seed hides that variation in exchange for a number a reviewer can reproduce exactly. The unseeded spread was checked, and for the main test it is roughly 0.015–0.025 — comfortably inside the same conclusion.
Statistics that do not depend on a random draw — pseudo-F, every R², every correlation — are deterministic regardless of the seed, which is why those are the ones the test suite pins. The suite recomputes every headline number on this site from the raw tables on every commit; the badge in the top bar is that suite's live status, and a red badge means a number on these pages is in doubt.
| Package | Version |
|---|
What isn't documented
Three gaps, stated because a methods page that lists only what went well is not a methods page.
Parts of the wet lab are not recorded. The extraction kit, library preparation, sequencing platform and read length were not passed on with the data, and are being requested. They do not change any result already reported, but a reader cannot fully reproduce the lab work without them.
The 18S taxonomy cannot be regenerated. It arrived as a finished object with no script, log, or reference file, so the exact database release cannot be pinned down. It is used for coarse description only, and no conclusion on this site rests on it alone.
This soil's own chemistry has not been measured. No pH, no nitrate, no ammonium, no organic matter from a laboratory — only the mapped priors above. Every mechanism this study proposes would be settled or refuted faster by one round of soil chemistry than by more sequencing. The acidity test above works around the gap by asking which kind of archaea are present rather than what the pH is, and it argues that acidity is not the explanation — but an argument from which organisms turned up is not a measurement, and should not be mistaken for one.
Nothing here measures activity. Every functional statement on both pages is about which organisms are present and in what proportion. A gene that is present may be switched off; an organism that is abundant may be dormant. Nitrogen cycling was never watched happening in this soil, and no number on this site is a rate.