Angiospermae · an interactive demonstration built on Kew's Tree of Life · release 4.0

A Tree of Life
you can compute on

Kew's Tree of Life places 353 nuclear genes across 20,475 flowering-plant samples spanning every order and family, with the gene trees and alignments all published. That is target capture, not whole genomes — about 260 kbp of sequence, under 0.1% of a genome. Those same loci are a coordinate system the whole tree already shares — and this demonstration reads them three ways. Pick a clade and see, across all 353 loci, which are recovered, which carry signal at that depth, which hold the clade together against the rest of the tree, and how divergent they are. Pin them to a sequenced genome and they land dispersed across its chromosomes, each inside a gene — broad genomic dispersion, consistent with the reduced linkage the coalescent tree assumes. Thread them through several genomes in the tree's own order and they become synteny anchors, showing where chromosomes keep their shape across a clade, and where a lineage has dealt them into a different hand.

The locus lens

Three orders carry real computed results: , and — click one here, or a green-highlighted tip in the tree below. For them, Recovery, Informativeness, Clade cohesion and Divergence are all measured from Kew's data; the rest of the tree is drawn for context, not clickable — the other orders carry no clade-specific data to show. (Node concordance is the one exception among the measures: a proposed placeholder, even for these three.) The panel opens on Poales with Clade cohesion showing, so the first thing you see is a real result.

Click a highlighted order, then choose a measure — Recovery, Informativeness, Clade cohesion, Node concordance or Divergence. The 353-bar panel recomputes for that clade — every bar is one Angiosperms353 locus, in Kew's gene order. For Poales and Fagales you can also drop a level to the family nested inside — Poaceae or Fagaceae — with the grain toggle above the measures, which reads the same 353 loci one level deeper. The note under the buttons says what the measure means and what to look for. And any single locus is a thread you can pull through the whole page: click one — a bar here, a table row, or a chromosome tick or ribbon in the genome panels below — and it stays lit in all of them at once, and stays lit as you switch clades. To reach a locus you can't see among the 353 bars — say PSAG (6498) — type its id or gene name into the find box above the chart. Shift-click bars or rows to pin several at once; each pinned locus shows as a chip you can release.

Measured — Kew release 4.0 Illustrative — Node concordance, not yet computed

Order-level tree

Orders highlighted in green — · — are the three with measured data, and the only orders you can select. The rest of the tree is drawn for context.

Branch colour is Kew's own ASTRAL-IV local posterior probability (LPP) — its branch-support score, from 0 to 1. The Flag LPP < 0.5 button above marks the weakly supported nodes: 13 of 67 fall below 0.5; five sit at exactly 0.0. Those are the deep nodes where the gene trees disagree. A coalescent tree shows its uncertainty rather than hiding it, and they are exactly the nodes a locus lens is for. Methods

353 loci · Poalesmeasured

locus 1quartet concordance factor for this cladelocus 353

Hover any locus.

Locus	Exemplar	Species	Genera	Cohesion · QCF

Show

What comes from Kew, and what this demonstration computes. The tree, its support values, the 353 locus identifiers, their exemplar Arabidopsis gene names, and per-locus species, genus and recovery figures are all from Kew's data. For Fagales, Chloranthales and Poales — and, one level down, the families Fagaceae and Poaceae — this demonstration computes three measures: Clade cohesion is a quartet concordance factor on each of the 353 gene trees — the fraction of quartets, drawn as two tips inside the clade against two from anywhere outside it, in which the clade's pair stays together; Informativeness is the observed fraction of parsimony-informative sites within the clade; and Divergence is the mean pairwise sequence divergence within the clade. Cohesion answers the clade's own premise: Poales is non-monophyletic on almost every gene tree, yet its cohesion stays above 0.99 — the size artefact that disqualifies monophyly counts, shown directly. Because "outside" spans the whole tree, it measures the clade's boundary against everything else, not resolution among its internal branches. Divergence ships the observed value; the Jukes-Cantor-corrected value is shown labelled beside it, never in its place. Node concordance remains simulated for every clade. The measured values are observed counts — no rate model, no assumed timescale.

The same loci, one level down — and the two families disagree about how much that matters. Drop Fagales to Fagaceae and every locus carries about a third as much signal: median parsimony-informative fraction falls from 0.36 to 0.12, median divergence from 0.084 to 0.025 — within the family the sampled tips are shallower and closer, so less varies. Drop Poales to Poaceae and the budget barely moves: informativeness 0.82 to 0.76, divergence 0.17 to 0.12, the ranking almost unchanged (rank correlation 0.93). Poaceae is a deep radiation that also makes up much of the sampled Poales, so the order and family views overlap heavily and dropping a level costs little — a contrast that reflects sampling depth and composition as much as absolute clade age. In both, the ranking of loci is broadly preserved, and Clade cohesion differs only in the tail: loci that fight the deep order tree resolve cleanly within the family — PSAG (6498) does it in both, rising from QCF 0.61 to 0.93 in the oaks and 0.75 to 1.00 in the grasses, because its conflict is with deep relationships, not the family. How much depth matters depends on which clade you are standing in.

What you can find here

Four questions a working systematist asks, and how to answer each by clicking. Every named locus is a measured result — select the clade in the tree, then the measure, and read it off.

1 · Which loci should I sequence or align for my group?

Select your clade, open Informativeness to see which loci carry signal at that depth, then Clade cohesion to see which of those hold the clade together against the rest of the tree. Loci tall on both are the workhorses. For Poales, genes 5335, 7021 and 5599 reach QCF 1.00 with ~84–90% informative sites — start a dataset there, not from a locus you happened to have a primer for.

2 · Where is the interesting conflict — introgression, sorting, paralogy?

On Clade cohesion, hunt the rust bars: loci whose gene tree places your taxa outside the clade. Cross-check Informativeness — if the same locus is also tall there, the disagreement is real signal, not noise, and worth investigating. Gene 6376 (INDL) is the deepest dip in Fagales (QCF 0.62), Poales (0.72) and Chloranthales (0.64) alike — a locus that fights the species tree across unrelated clades is a paralogy suspect you would want flagged before it poisons a concatenated analysis.

3 · Which of my loci diverge fastest — and might mislead?

Open Divergence: warmer bars are more divergent inside your clade. High divergence can erode historical signal, but on its own it does not diagnose saturation — repeated substitutions, paralogy, alignment error and compositional bias all look alike in a raw p-distance. What is worth a second look is a locus that is both highly divergent and rust on Clade cohesion. Gene 6376 (INDL) is the single most divergent locus in all three clades — about 23–27% — and also the most discordant. That combination makes it a quality-control priority, not a settled verdict: check for saturation, paralogy, alignment error or compositional bias before a concatenated tree averages it in.

4 · Which loci come back well tree-wide, and which barely do?

Recovery shows how much of each locus comes back, averaged tree-wide across mixed source types (target capture, genome- and transcriptome-mined) — a coarse guide, not a per-clade or per-collection prediction. Genes 6507 and 6406 are the two lowest-recovery loci tree-wide, at 43.8% and 46.4%; a locus that recovers poorly across the tree is a weak default choice, though it may still work in a particular lineage. Read it beside Informativeness before dropping anything from a panel.

The habit this builds. Notice that Poales looks absent under a monophyly check — non-monophyletic on all 353 gene trees — yet Clade cohesion shows it is 99.7% concordant. Reading cohesion instead of counting monophyly is the difference between "this clade falls apart" and "this clade is solid; three loci disagree, and here they are."

A genome layer, on one branch

Kew's tree resolves every flowering-plant lineage from the same 353 nuclear markers — about 260 kbp, under 0.1% of a genome. Attach one whole genome to a single tip — Quercus robur — and this is what appears: every one of the Angiosperms353 loci at its real position on the twelve chromosomes, in genomic context. Hover a locus in the panel above, or a tick here, and it lights up in both. Do this wherever the tree meets a sequenced genome and the 353 markers stop being a list and become a map. Methods

Positions, chromosomes, gene density — measured hover a locus

Hover a tick, or a locus in the panel above, to place it in the genome. Click to pin it.

What a genome layer adds. In the panel above a locus is an identifier and a number; anchored to a genome it gains an address — locus 5599 (NERD1) sits on chromosome 6, inside gene LOC126733144, beside its neighbours. And the pattern across all 353 is the assumption the tree rests on: they occupy just · of the genome yet scatter across all twelve chromosomes, and they land where the genes are — the gene-poorest quarter (about · Mb, mostly pericentromeric) holds · of them, the gene-richest holds ·, and · of 353 fall inside an annotated gene. Loci this dispersed and this genic are consistent with the reduced linkage the coalescent species tree assumes of them — dispersion lowers linkage without proving the loci independent. That is one branch's worth of a genome layer. Run the same alignment on the other sequenced Fagaceae genomes and the shared loci become synteny anchors, tying one genome to the next — which is the next panel.

Twelve chromosomes — but not the same twelve

Now thread the same 353 loci through five Fagaceae genomes, stacked in the order Kew's own tree gives them. All five assemblies report twelve chromosomes. Across chestnut, stone oak and the two oaks, those twelve are the same twelve — each locus stays on the matching chromosome, while the genome runs from 716 Mb to 964 Mb. The anchors keep their chromosome assignment as the genome grows; how much finer structure shifts between them is beyond what 353 markers resolve. Now beech, sister to the other four: still twelve chromosomes, but the anchors have been dealt into different hands. Methods & what's unresolved

Measured — five chromosome-level assemblies Chromosome homology computed, not assumed hover a locus

Hover an anchor to follow one locus down the tree, or click to pin it.

How to read it. Each row is one genome, drawn to scale, so the bars lengthen as the genome grows. Each tick is one of the 353 loci at its real position; each ribbon follows one locus from the genome above to the genome below. Click any anchor to pin it — the locus stays lit here, on the oak chromosomes and in the 353-locus chart until you click it again. Colour marks the linkage group it occupies in Q. robur — six hues, hatched bars and dashed ribbons for groups 7–12, and every bar is labelled, so no group is told apart by colour alone. Between the lower four genomes the ribbons run in twelve clean bundles. Between beech and chestnut they cross.

What was measured, and what was not. Chromosome numbering is arbitrary in every assembly, so homology was computed from the shared loci themselves by maximum-weight matching — a one-to-one pairing by construction, so its credibility rests on the match purity below, not on the pairing itself. Measured against Q. robur, the other three of the lower four put · of shared loci on the matching chromosome, and each of the four resolves into exactly twelve syntenic blocks. Beech scores · and resolves into · blocks. That is not a broken assembly: the same mosaic appears in three independently assembled beech genomes — Sanger, Senckenberg and Genoscope — which agree with each other 98.0–99.4% and all disagree with the oaks. Which lineage did the rearranging, we cannot say. Beech is the sister lineage, so the change could sit on either stem branch; six Fagales outgroups all lean toward the oak arrangement, but only the two Juglandaceae clear a bootstrap interval and they are a single family, not independent replicates. Reported unresolved. These 353 anchors resolve which chromosome a locus sits on — not fine-scale order, inversions, or block boundaries. Chromosomes are drawn in whichever orientation agrees with the reference, since assembly orientation carries no meaning. Loci with a strong second hit are drawn at their best hit and counted in the table.

What these two panels could become across the tree. Both run on one fact: the same 353 loci are recovered at every tip, so they are a coordinate system every angiosperm genome already shares. Through them a newly sequenced genome pins to the tree with no whole-genome alignment. Stack any clade's genomes in Kew's topology and the beech mosaic stops being a single finding and becomes a question you can ask anywhere — where do the linkage groups hold, and where has a lineage dealt them into a different hand? Count how many times each anchor appears in a genome and repeated hits flag a locus for duplication follow-up — though 353 deliberately low-copy markers can only raise the question, not diagnose whole-genome duplication, which needs genome-wide syntenic evidence. The empty branches carry information too: only about 1,222 of some 13,600 flowering-plant genera have any genome at all, so the map shows which lineages, if sequenced, would tie the most of the tree together. Each of these is a direction these tools could open across the full tree — an illustration of what Kew's data could power, not a result claimed here.

Methods & data

Where each figure's numbers come from, how they are computed, and what to distrust. Every measured value is observed, not modelled, and nothing is computed in the browser — the page embeds a precomputed table and fetches nothing. Each entry names the script that reproduces it. Informativeness, Clade cohesion and Divergence are measured from Kew's published alignments and gene trees for three orders — Fagales, Chloranthales, Poales — and two families within them, Fagaceae and Poaceae. For every other clade, and for Node concordance throughout, the bars are a placeholder.

What this is for, how it was made, and by whom

Purpose: A demonstration of what AI coding tools now put within reach: interactive graphics a working scientist can build, refine, and iterate on directly — without a separate engineering team — to explore a dataset and sharpen the questions worth asking of it. Nothing here is claiming a novel scientific result. Directed by a phylogeneticist the same approach would go much further, and anything bound for production would be rebuilt in whatever system hosts it; the point is the tool, and how quickly the researcher doing the science can now make one.
Direction & authorship: Peter Repetti set the scientific direction: the questions, the clades to measure, the framing, and where a claim ends and an open question begins. Claude (Anthropic) wrote the analysis scripts, the computed tables, and this page under that direction and review.
What is checked, and what isn't: The claim is reproducibility, not infallibility. Every measured value comes from a script in a private repository, run over Kew's published files — none hand-entered or AI-guessed — and it reruns identically. But rerunning identically only proves the code is deterministic, not correct: a wrong assumption returns the same wrong answer every time. The arithmetic is checked against independent controls (quartet counts against brute force; family QCF against a second implementation), and the pipeline asserts its own consistency — but none of that catches a wrong assumption (a mis-parsed label, a debatable threshold, a subtle bug) that yields numbers which are self-consistent and plausible but off, a failure mode AI-written code is prone to. So: a demonstration, not a published result. The script behind every measure is named in its entry below and can be shared for audit.
Data & tools: The data are Kew's. Standard third-party tools the scripts call — minimap2, dendropy, scipy — are cited where used. The scripts that tie them together were written for this demonstration.

The data underneath

Primary source: Kew Tree of Life Explorer, data release 4.0 (April 2026) — the analysis is locked to this release, not Kew's rolling latest. 64 orders, 417 families, 12,239 genera, 20,375 species across 20,475 sequenced angiosperm samples; the species tree has 20,485 tips, the extra ~10 being the gymnosperm outgroup used to root it. API at treeoflife.kew.org/api; bulk files at sftp.kew.org/pub/paftol/current_release.
The 353 loci: The Angiosperms353 universal probe set (Johnson et al. 2019): 353 low-copy nuclear genes, ~260 kbp of target — under 0.1% of an angiosperm genome.
Licence & norm: CC BY 4.0, released pre-publication under the Toronto guidelines, which ask users to let the data producers publish the first global analyses. Everything here is a bounded prototype on a few named clades; no tree-wide surface is built.
Cite: Baker et al. 2022, Systematic Biology 71:301–319, plus release 4.0.

The order-level tree measured

Source: Release 4.0 order-level species tree, 64 orders + 5 gymnosperm outgroups.
Branch colour: Kew's own branch support. The topology is weighted ASTRAL, but the support values shown are computed with ASTRAL-IV on that same topology — Kew switched to ASTRAL-IV because weighted ASTRAL emitted -nan for some values. They are local posterior probabilities (0–1).
What it shows: 13 of 67 internal nodes fall below LPP 0.5; five sit at exactly 0.0 — the deep nodes where the gene trees disagree. Flagged by default.
Read as: A coalescent tree shows its uncertainty; low support is a real property of the data, not an error, and is not a claim that any lineage rearranged or came first.

Recovery measured, every clade

Source: Kew's per-gene metadata (API), tree-wide.
Shows: Mean recovered contig length as a percentage of the target — how much of each locus you actually get. Tree-wide, so it is measured for every clade.
Caveat: Recovery is not signal: a fully recovered locus can still be uninformative or discordant for your clade. Read it alongside the other three.

Informativeness measured · 5 clades

Source: The 353 aligned FASTA files, fasta/alignments/{gene}.dna.aln.fasta (~8.1 GB, streamed, never stored), restricted to the taxa in the selected clade.
Shows: Observed parsimony-informative sites per scored column within the clade — how many characters actually distinguish your taxa. A column is scored when at least 4 clade taxa carry an unambiguous base (MIN_SCORED = 4).
Not: Not Townsend phylogenetic informativeness — that assumes a rate model and a time-calibrated tree, and angiosperm dating is unsettled. This is an observed count, no model.
Read as: The denominator is scored columns, not target length, and a column needs only 4 clade taxa present (MIN_SCORED = 4). In a densely sampled clade a sparse or short locus can therefore post a high fraction on few taxa; read it beside occupancy and total aligned length, and treat it as a within-clade ranking, not an absolute yield.
Reproduce: analysis/clade_locus.py (orders), analysis/family_clade.py (families), via informativeness.py.

Clade cohesion measured · 5 clades

Source: The 353 gene trees, tree/gene/{gene}.tree (Newick, internal labels are support percentages 0–100).
Shows: An exact quartet concordance factor: of the quartets that take two tips from your clade and two from outside, the fraction the gene tree resolves as keeping your two clade tips together — concordant / (concordant + discordant), over the decisive quartets only. Counted in one traversal, not enumerated.
Why quartets: Strict monophyly on a gene tree is destroyed by one misplaced tip, and larger clades have more tips to misplace (log clade size vs monophyly r = −0.650). Quartets degrade gracefully, so Poales scores > 0.99 despite being non-monophyletic on almost every gene tree. It is not size-invariant — still sensitive to gene-tree error and paralogy.
Read as: Because "outside" spans the whole tree — very distant taxa included, gymnosperms among them — most quartets are easy, so this measures the clade's boundary against the rest of the tree, not resolution among its internal branches. It is a clade-cohesion statistic, not a node-level support value.
Threshold: Gene-tree branches below support 30 are collapsed before counting (also reported at 0 and 50 for sensitivity). This threshold is our defined choice; the 4.0 release notes do not specify an in-pipeline collapse.
Reproduce: analysis/qcf.py (self-tested against brute force).

Divergence measured · 5 clades

Source: The clade-restricted alignments, as Informativeness.
Shows: Mean pairwise p-distance within the clade — the observed fraction of positions at which sequences differ. The Jukes-Cantor-corrected value is shown labelled beside it, never in its place.
Scale: Bar height is a within-clade ranking only, scaled per clade: median divergence differs between clades (Poales 0.17 > Fagales 0.08), a gap that reflects clade depth, taxon sampling and lineage-specific rate together — not age alone — so a shared scale would mislead. Do not compare bar heights between clades. Observed divergence, not an evolutionary rate: no elapsed-time denominator is assumed.
Reproduce: analysis/clade_locus.py (orders), analysis/family_clade.py (families), via informativeness.py — the mean pairwise divergence is computed alongside the informative-site count.

Node concordance proposed — not computed

Status: The bars are placeholders and their ranking is not a result. It would measure how well each locus's gene tree agrees with the species tree, node by node.
Why not yet: Done over the whole tree it would be a first global analysis on the full data — beyond what this demonstration sets out to claim. A bounded version could ask it of the family-level nodes inside one order (e.g. Poaceae, Cyperaceae, Bromeliaceae), where enough taxa sit either side of a node that a short locus's gene-tree error does not read as discordance.

The grain toggle: order vs family measured

What: Reads the same 353 loci one level deeper — an order (Fagales, Poales) or a family nested inside it (Fagaceae, Poaceae). Clade membership keys on the family field of Kew's specimen metadata instead of order; the computation is otherwise identical.
Measured clades: Three orders — Fagales, Chloranthales, Poales — and two families, Fagaceae and Poaceae. Chloranthales has no family grain because it is monofamilial (the order is Chloranthaceae).
What it shows: How much depth matters depends on the family. Fagaceae, small and young, loses ~3× its signal one level down; Poaceae, an old deep radiation, barely loses any (rank correlation 0.93). A few loci discordant across an order resolve cleanly within the family.
Reproduce: family_clade.py --family Poaceae → merge_family.py → embed.py.

The oak chromosome panel positions measured

Genome: Quercus robur GCF_932294415.1 (dhQueRobu3.1), 12 chromosomes.
Shows: The real position of each of the 353 loci on the twelve chromosomes, mapped with minimap2 (spliced, best hit per locus), plus gene density along each chromosome.
Circularity: Kew's oak sequence was mined from this assembly, so it maps back by construction — no recovery is claimed. What the positions describe is where each locus sits on the genome, which holds regardless of how the query was obtained.

The cross-oak synteny panel measured

Genomes: Five chromosome-level Fagaceae assemblies: Fagus sylvatica GCA_964211995.1, Castanea sativa GCF_040712315.1, Lithocarpus litseifolius GCA_040182985.1, Q. lobata GCF_001633185.2, Q. robur GCF_932294415.1.
Method: Each genome's own Kew-recovered A353 sequence mapped with minimap2; native chromosome → reference linkage group computed from the shared loci by maximum-weight bipartite matching (one-to-one by construction; its credibility is the match purity below, not the pairing). Stacked in Kew's own topology, (Fagus,(Castanea,(Lithocarpus,(Q. lobata, Q. robur)))).
Result: The four Quercoideae share the same twelve linkage groups (97.7–98.3% of loci on the matching chromosome) while the genome runs 716→964 Mb — the anchors keep their chromosome assignment as the genome grows. Beech's twelve carry a reorganised set of the same linkage groups (46.4% on the matching chromosome, 30 syntenic blocks) — homologous chromosomes whose composition has shifted, not a different set.
Unresolved: The beech mosaic is not an assembly artefact — three independently assembled Fagus genomes agree with each other and all disagree with the oaks. But which lineage rearranged is not settled: Fagus is sister to the other four, so the change could sit on either stem branch, and the outgroup test could not polarise it. Reported open. The anchors resolve a locus to a chromosome, not to finer order.
Reproduce: analysis/synteny.py, checks in analysis/synteny_checks.py.

Reading the code, and steering it, without writing any

You do not have to program to check whether the analysis is sound. What matters in a script is not the machinery but a handful of decisions — thresholds, what counts as data, what gets measured — and those read in plain terms. Two examples, one you can read and one you can't.

A script you can read. This is the core of informativeness.py, behind the Informativeness bar — lightly simplified, the array bookkeeping trimmed, the decisions left as they run:

MIN_SCORED = 4   # a column needs ≥4 of your clade's species present to count at all

# For each alignment column, count how many A, C, G, T appear across your clade's
# sequences. Only real bases count — gaps (-), N and ambiguity codes are treated as
# missing. `present` = number of species with a real base in that column.

scored       = present >= MIN_SCORED
n_states     = bases_present(column)          # how many of A,C,G,T appear at all: 1 = constant
n_states_ge2 = bases_in_two_or_more(column)   # how many appear in ≥2 species

variable = count_columns( scored & (n_states     >= 2) )   # species differ here
pis      = count_columns( scored & (n_states_ge2 >= 2) )   # ...differ usefully for a tree

pis_frac = pis / cols_scored                  # ← the Informativeness bar

Four decisions, all legible without reading the machinery: the ≥4 floor (score a locus only where at least four of your clade's species have real sequence — raise it for a sparse clade and rerun); what counts as data (only A, C, G, T, so a gappy column can't masquerade as variable); the variable-vs-informative distinction (a lone odd base varies, but a tree can't use it); and the denominator (scored columns, not full length, so a locus isn't charged twice for missing data). A systematist can accept, reject, or change any of them without touching the array code.

A script you can't — in plain words. Clade cohesion is computed by qcf.py, which is genuinely dense: it counts quartets, and there are far too many to list (Poales against the rest of the tree is about 260 trillion), so it totals them by a bookkeeping trick in a single pass over the tree. You do not need to follow that. What it does — and what you can question — is this. It takes one gene tree and one clade, considers every way of picking two species from your clade and two from outside it, and asks whether the gene tree keeps your two together; cohesion is the fraction where it does, over the quartets the tree actually resolves. The choices inside are legible even though the code isn't: "outside" means every other tip on the tree, gymnosperms included; weakly supported branches (below bootstrap 30) are collapsed first, so a shaky branch gets no vote; and unresolved quartets are excluded from both the count and the total. The fast counting is checked against a slow, obviously-correct version that really does enumerate every quartet, on small trees, and the two must agree. Those choices — what "outside" is, the collapse threshold, dropping ties — are exactly what a systematist would weigh, and none of them requires reading the counting code.