Audit a whole-genome-alignment claim

Referee a whole-genome-alignment (WGA) result plus the biology drawn from it (synteny, SV, “region absent/unaligned”, ancestral sequence, pangenome variant). The cardinal sin: an alignment artifact read as biology. Walk the gates in order; a claim passes only if every triggered sub-form check clears. If a check cannot be cleared from the evidence supplied, the verdict is flag (possible artifact), never pass.

Gate 0 — restate the claim and capture the alignment provenance

Extract and record:

• The biological conclusion and its logical form (presence/absence, SV, synteny, ancestral base, variant).
• Alignment type: (a) pairwise reference-anchored (chain/net, minimap2, nucmer); (b) reference-free multiple (Progressive Cactus -> HAL); (c) WGD/ploidy-aware (AnchorWave).
• Tool + version + preset/params: minimap2 preset (asm5/asm10/asm20) and -k/-w; nucmer mode (default MUM vs --maxmatch); Cactus guide tree + masking state; AnchorWave -R/-Q; HAL toolkit versions.
• Reference direction (which genome is the axis) and whether the claim was tested under the opposite direction.
• Masking state of inputs. If any of these is unstated, note it — an absent provenance field is itself a flag input.

Gate 1 — route by alignment type

• Pairwise reference-anchored -> run Gate 2A (reference bias) + Gate 2B (sensitivity/false-absence) + Gate 2C (masking).
• Reference-free multiple (Cactus/HAL) -> run Gate 2C (masking) + Gate 2D (guide-tree) + Gate 2E (HAL version/tooling).
• WGD/ploidy-aware (AnchorWave) -> run Gate 2F (alignment-depth) + Gate 2B (sensitivity) + Gate 2C (masking).
• Any pipeline that extracts from HAL or hands alignment to a downstream consumer -> also run Gate 2E.

Gate 2A — reference bias (directionality of chains/nets)

A reference-anchored net is asymmetric: it keeps only the single best-aligning target DNA per position on the reference side; the opposite-direction net is built oppositely, so target-net != query-net (Kent 2003). Therefore:

• A presence/absence or SV conclusion must be checked in both reference directions.
• If the conclusion flips (or is untested) between reference choices -> flag as reference-bias, do not report as biology.
• No numeric chain-length/score cutoff is defensible: the often-quoted “5000 bp” minimum chain is UCSC convention, NOT in Kent 2003 — if the audit sees such a cutoff cited to the paper, flag it as convention-mislabeled-as-cited. [GAP: no numeric minimum chain length/score in the notes; treat any such floor as convention.]

Gate 2B — sensitivity below threshold -> false absence

Absence of alignment != absence of homology. An “unaligned/absent/lost” call is only valid if the aligner could reach the divergence of the region:

• minimap2: asm5 “will not extend to regions with 5% or higher sequence divergence” (recommended when avg divergence ~0.1%); asm10 for ~1% avg; asm20 for ~several % avg. An absence call from a preset whose ceiling is below the expected divergence is a possible false-absence -> flag; remedy = move up the preset ladder before concluding absence.
• nucmer: seeds on long exact matches with default reference-uniqueness -> less sensitive (3-5% fewer reads than BWA/Bowtie2); divergent or repeat-heavy regions drop out. A nucmer absence call over a divergent/repetitive locus is a possible false-absence -> flag; remedy = --maxmatch (at run-time cost) and/or a more sensitive aligner.
• The minimap2 divergence estimate is “only approximate and can be biased” — do not treat a reported per-base divergence as a calibrated identity readout.
• No numeric identity floor exists in either source (the commonly cited ~70% identity floor is NOT in minimap2 or nucmer notes). [GAP: no minimum-identity threshold in the references; do not assert a numeric identity floor — label ~70% as convention if cited.]

Gate 2C — repeat masking sufficiency

Unmasked/poorly-masked repeats drive spurious alignment and, for Cactus, runtime/memory blowup (alignments enumerated to/from every copy of a repeat).

• Progressive Cactus requires soft-masked input (stated precondition); its self-alignment preprocessing masks regions with coverage > 50 (configurable), sampling 20% of chunks.
• Chain/net pipelines mask before alignment (RepeatMasker for transposons; Tandem Repeat Finder for simple repeats, period <= 12); skipping masking inflates spurious chains.
• If inputs are unmasked (or masking state unstated) and the claim is a duplication/SV/ spurious-synteny call in repeat-rich sequence -> flag as possible masking artifact.
• No numeric percent-of-TE (or %-of-genome) masking target is given in the notes. [GAP: no %-masked target; cannot assert "≥X% masked" — the only in-source numbers are Cactus's internal coverage>50 / 20%-sampling preprocessing, not a user-facing masking %.]

Gate 2D — guide-tree topology (reference-free / ancestral claims)

Progressive Cactus is progressive up a guide phylogenetic tree; ancestral assemblies are reconstructed at internal nodes by ordering blocks and taking a most-likely base per column. A wrong guide-tree topology corrupts ancestral-sequence inference. Therefore:

• Any ancestral-sequence or node-level claim must state the guide tree used and that it is the correct/rooted topology (HAL likewise presumes a rooted tree).
• If the guide tree is wrong, unrooted, or unstated for an ancestral claim -> flag.
• Accuracy also degrades with divergence (F1 0.989 primate vs 0.795 mammal; ~86% human coding covered in the placental ancestor) — deep-node ancestral claims carry lower confidence.
• [GAP: no guide-tree QUALITY criteria in the notes — no branch-length/support/rooting metric, no branchScale values (branchScale is explicitly ABSENT from Armstrong 2020). Cannot assert a numeric tree-quality bar; require the tree be stated and defensible, but flag the criterion as unsourced.]

Gate 2E — HAL version / downstream-tooling breakage

Extraction from HAL (halLiftover BED->BED; halBranchMutations per-branch BED; hal2maf) can be silently or loudly wrong across tool-version boundaries.

• Handle honestly: Hickey 2013 states no format versioning, no forward/backward compatibility policy — so there is no in-source rule to cite for HAL version safety. The audit can require that HAL-toolkit versions be pinned and that extraction was regenerated with matching tooling, but must not invent a compatibility guarantee. [GAP: HAL versioning/compatibility policy is ABSENT from the source; cannot cite a version rule — only require version disclosure and flag undisclosed version crossings.]
• Format-mismatch linkage: a downstream orthology/gene-loss conclusion (e.g. TOGA) requires pairwise chains, not a Cactus HAL — TOGA consumes chains (LASTZ->axtChain), not HAL/halSynteny. A pipeline feeding HAL directly into a chain-consuming caller is a tooling/ format mismatch -> flag (the “result” is a plumbing artifact, not biology).

Gate 2F — WGD / alignment-depth setting (polyploid under/over-matching)

For genomes with differing whole-genome-duplication histories, AnchorWave controls matching via -R (reference max alignment depth) and -Q (query max alignment depth), e.g. maize<->sorghum -R 1 -Q 2 (maize depth 1, sorghum depth 2). Anchors are reference full-length CDS (requires a GFF3 annotation).

• Check -R/-Q are set to the known ploidy/WGD ratio of the pair. A too-low depth leaves true homologous copies unmatched (under-alignment); a too-high depth admits extra paths (over-alignment). If a “copy absent” or “extra copy” claim rests on a depth setting that doesn’t match the ploidy -> flag.
• Regions with too few CDS anchors get aligned as indels, and interanchor low-homology stretches get only approximate (banded/sliding-window) alignment — an SV/absence call there is a possible artifact -> flag.
• Watch for interface drift: the notes describe -R/-Q (v1.0.0); a --ploidy flag does not appear in the source. If a claim cites proali --ploidy, flag as version-unverified.
• [GAP: the source states NO symptom/error/diagnostic for a wrong -R/-Q (under/over-matching is summarizer-inferred). No numeric anchor-identity threshold exists (the "80% default" is NOT in the source). Cannot assert a diagnostic value or identity floor.]

Gate 3 — verdict synthesis

For each triggered sub-form, record: cleared / flag(artifact sub-form) / cannot-assess.

• Pass only if every triggered check is cleared.
• If any check is flag or cannot-assess, the verdict is artifact-suspected (name the sub-form) or insufficient-provenance — never a silent pass. State the concrete remedy (opposite-direction net; higher preset / --maxmatch; re-mask; correct/rooted guide tree; version-pin + regenerate extraction; ploidy-matched -R/-Q).
• [GAP: no combined-confidence rubric in the notes — there is no source-backed way to weight or aggregate multiple partial flags into a single numeric score. Report per-sub-form verdicts; do not manufacture an overall confidence number.]