audit-wgd-inference

Referee a whole-genome-duplication (WGD) inference for method validity — whether the WGD claim survives the known artifact axes, not whether the biology is interesting. Default posture: a claim that cannot clear an axis is FLAGGED, never silently passed (the Family-B gate). A flag names the axis, the specific unmet condition, and what evidence would clear it.

Check 0 — Pin down what is actually claimed

Route the audit by the claim type; different claims trigger different axes:

(a) existence of an ancient WGD (a peak above SSD background),
(b) phylogenetic placement relative to a speciation (shared vs lineage-specific),
(c) ploidy level (duplication vs triplication / higher),
(d) a date for the event.

[GAP: the routing/aggregation logic — which checks are mandatory for which claim type, and how individual flags combine into a single PASS/FLAG verdict — is not supplied by any note; it is design synthesis. Treat the mapping below as a working default, not a sourced rule.]

Check 1 — Ks saturation boundary on the peak (cardinal; always)

Grounded in vanneste-2013-ks-saturation.

A tail peak at high Ks may be a saturation peak: an artifact where old duplicates are compressed onto smaller estimated Ks and pile up, mimicking a WGD even in pure small-scale- duplication (no-WGD) simulations. It is an artifact of Ks estimation, not a duplication burst.
Reliability bounds (note these are the numbers, but they are species-dependent): Ks < 1 is the conventional “reliable” cutoff; Ks stays linearly informative to synonymous age ~2; mixture-model peak inference is reliable only below synonymous distance 2–2.5; underestimation is visible by age 3 (mode 2.6 vs expected 3) and severe by ages 5–20.
Diagnostic that a peak IS a saturation artifact, not a real WGD: its Ks position tracks the modeled timespan / Ks cutoff, not an independently dated event (for A. thaliana the saturation- peak mode moved 2.8 → 4.4 as the modeled timespan went 5 → 20). Saturation onset is species- specific (faster in D. rerio, C. albicans, K. lactis), driven by transition bias + codon usage.
FLAG any ancient-WGD claim resting on a peak in the saturation region (roughly Ks ≳ 2–2.5, up to ~4–5 depending on species/timespan) that is not independently corroborated (rate-adjusted date, syntenic depth, or an external calibration). Also flag the mirror error: a genuine ancient WGD can fade — its peak flattens and blends into the L-shaped SSD background — so absence of a peak past ~2.5 is not evidence of no WGD.
Cross-check the analysis’s saturation cutoff against chen-2024-wgd: wgd --kscutoff default is 5 but the paper recommends “typically 2 or 3” for dating; a dating claim that kept anchors far above 2–3 inherits saturation risk.
[GAP: no turnkey rule separates a saturation peak from a real WGD peak in an empirical dataset — Vanneste diagnoses the artifact and bounds reliability but leaves the operational separator unresolved. No universal numeric Ks threshold exists; per-species saturation-onset values are illustrated for a few taxa only, not tabulated. Do not apply one fixed Ks cutoff across species.]

Check 2 — Is the peak WGD signal, or tandem/small-scale contamination?

Grounded in qiao-2019-dupgen, wang-2012-mcscanx, tang-2008-jcvi, lovell-2022-genespace.

WGD-derived signal is synteny-anchored (collinear) pairs, not the raw paranome. A Ks distribution built from all duplicate pairs mixes in tandem, proximal, transposed and dispersed duplicates, which can create or inflate a peak that is not a WGD.
Duplication modes (qiao-2019-dupgen, DupGen_finder): the five classes are exactly WGD, tandem, proximal, transposed, dispersed. WGD = intra-genome MCScanX collinear pairs; tandem = directly adjacent (0 intervening genes); proximal = ≤ -d genes apart, default 10. There is no “segmental” class in DupGen and no separate “tandem=5 / proximal 5–25” windows — a claim using either is wrong (see Check 5).
Collinear-block criteria (wang-2012-mcscanx): tandem = gene-rank difference 1; proximal = rank difference < 20; a reported block scores > 250 (≥ 5 collinear pairs), gap < 25 intervening genes, block E-value 1e-5; consecutive shared-gene matches within 5 genes are collapsed. Class priority WGD/segmental > tandem > proximal > dispersed.
Tandem-array handling: tang-2008-jcvi collapses tandems < 50 kb (residual long-distance tandems inflate multiplicity); lovell-2022-genespace collapses tandem arrays to array-representative genes before ranking (blkSize = 5 unique hits — not orthogroups — synBuff = 100, nGaps = 5).
FLAG a WGD claim from a raw-paranome Ks peak with no synteny-anchor (anchor-Ks) confirmation and no tandem/proximal removal — especially a low-Ks peak, where tandem arrays concentrate. Clearing evidence: the peak persists in the anchor-Ks (collinear-pair) distribution after tandem collapse.
[GAP: no note states a minimum anchor-pair COUNT that legitimizes calling a synteny signal a WGD (as opposed to one collinear block). Tang gives a block SCORE cutoff (>300), MCScanX >=5 pairs / score>250, GENESPACE blkSize=5 hits — these gate a block, not a genome-wide WGD verdict.]

Check 3 — Cross-lineage placement requires rate correction

Grounded in sensalari-2022-ksrates.

Synonymous substitution rates differ among lineages, so equal-age events sit at different Ks depending on the lineage measured. A focal species’ paralog Ks distribution is not directly comparable to ortholog Ks distributions involving other, differently-evolving species; naive superimposition in a “mixed plot” can misplace a WGD relative to a speciation (the oil-palm τ event was wrongly read as palm-specific in the naive plot; rate-adjustment corrected it).
Correction = RRT-style branch decomposition using an outgroup: split the ortholog Ks into per-branch contributions, rescale the divergence onto the focal Ks scale as twice the focal branch contribution (docs example: palm–rice 1.53 → palm 0.365 + rice 1.17 → adjusted 2×0.365 = 0.73). Requires ≥ 1 outgroup; ksrates docs advise 3–4 (max_number_outgroups default 4).
FLAG any before/after-speciation or shared/lineage-specific placement claim built from uncorrected cross-lineage Ks. Clearing evidence: rate-adjusted mixed plot with adequate outgroups.
[GAP: no HARD minimum outgroup count is sourced — the paper states only "an outgroup" (>=1); the >=3-4 figure is docs advice, not a requirement. And NO Ks validity/saturation ceiling for the correction is stated in ksrates (the value 1.53 is an example, not a cutoff) — do not cite this source for a Ks<1.5-type bound on rate correction.]

Check 4 — A specific ploidy level needs syntenic depth

Grounded in tang-2008-jcvi.

Syntenic depth (multiplicity) of aligned loci indexes duplication history, read against an assumed per-lineage event history: expected multiplicity Carica = 3, Vitis = 3 (γ only), Populus = 6 (γ × p), Arabidopsis = 12 (γ, β, α); empirically 88 Carica loci at depth 3, 54 Populus loci at depth 6; a 1:1:2:4 Carica:Vitis:Populus:Arabidopsis correspondence.
A duplication vs triplication distinction is a depth/multiplicity claim, not a single-peak claim. FLAG a specific-ploidy claim (e.g. “paleo-hexaploidy / triplication”) resting only on one Ks peak, with no multi-genome syntenic-depth support. Note failure modes: unpurged tandems inflate multiplicity above expected; misreading depth without the correct assumed event history misassigns ploidy.

Check 5 — Pipeline / parameter / tool-identity sanity (method reality)

Grounded in chen-2024-wgd, qiao-2019-dupgen, wang-2012-mcscanx, tang-2008-jcvi. The referee also judges whether the named method is real and correctly described (method validity per the glossary). FLAG stated-but-false tool/parameter facts:

wgd v2 detects collinearity with i-ADHoRe (v3.0.01), NOT MCScanX — a claim that wgd v2 uses MCScanX is wrong (chen-2024-wgd). wgd Ks estimator is CODEML (PAML v4.9j); redundancy removal is node-weighted by default (--node_average switches to node-averaging); mixture components default range (1, 4) selected by AIC/BIC (GMM).
DupGen classes are WGD/tandem/proximal/transposed/dispersed; “segmental” is not a class, and tandem = adjacency, proximal = single -d window (default 10) — “tandem=5, proximal 5–25” is wrong (qiao-2019-dupgen).
The 2008 MCscan (tang-2008-jcvi) controls stringency by BLASTP E ≤ 1e-5, best-5 hits, DP scoring, and score cutoffs — no cscore and no quota appear in that paper (those are the later JCVI port). A claim attributing a --cscore to Tang-2008 is wrong.

Check 6 — Independent corroboration (birth-death), with the assembly-error caveat

Grounded (as hypothesis-evidence) in de-bie-2006-cafe, mendes-2020-cafe5, han-2013-cafe3.

A WGD claim is strengthened if an independent phylogenomic signal — a burst of gene-family expansions datable to the same node — corroborates it. CAFE fits a birth-death process to family sizes and attributes rate shifts to branches (Viterbi per-branch p-value, branch-cutting, LRT; de-bie-2006-cafe), with among-family rate variation via gamma categories + empirical Bayes for fast/slow families (mendes-2020-cafe5).
Caveat that limits how much weight this evidence carries:
- Assembly/annotation error inflates gain/loss rates — “consistently overestimated,” ~2.3× at ε=0.1 and ~7.5× at ε=0.4 (han-2013-cafe3). A gene-count expansion burst from a fragmented assembly without an error model is not trustworthy WGD evidence.
- CAFE per-family p-values are uncorrected for multiple testing (de-bie-2006-cafe §10), so a handful of “significant” expanded families is weak corroboration on its own.
Use this as a corroboration/robustness check, not a primary axis.
[GAP: NO sourced note states that a CAFE expansion burst confirms a WGD or quantifies the corroboration strength — the CAFE notes are about gene-family evolution generally, not WGD detection. The link is a design inference (hence hypothesis-evidence on these three refs).]

Check 7 — Small-sample / rugged-histogram artifact