Statistical Genomics Foundry
Molds Mold
eval ✓ · scenarios ✓ family/brole/critique
contract

Reference Loading

Typed references describe what casting consumes, and when the generated skill should load each artifact.

Research turakhia-2020-gene-loss

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeUpfrontVerbatimCorpus Observed
Purpose
The validity-axis anchor: the enumerated technical causes of spurious loss (assembly gaps, annotation-tracks-N50, paralog-as-ortholog, private mutations, reference bias) and the guard for each (gap masking, raw-assembly not annotation, >=20x ortholog/paralog separation, gene-in-synteny >=20, unique 1:1, first/last-2-AA exclusion, phylogenetic filter), plus the >30%-vs-<10% FPR ablation that proves absence-of-annotation alone is not loss.
Research kirilenko-2023-toga

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeUpfrontVerbatimCorpus Observed
Purpose
The intactness/loss classification substrate: five transcript states (intact / partially intact / missing / uncertain loss / lost), the M-vs-D assembly-gap safeguard (>=10 N), the middle-80% rule, the >=2-exon / <60%-intact lost criterion, orthology-score >=0.5, and the coding-capacity-not-expression scope.
Research shumate-salzberg-2021-liftoff

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeOn DemandVerbatimCorpus Observed
Purpose
The mis-use guard: Liftoff is a same/closely-related-species annotation-transfer tool, not a loss detector; absence in its output is a failed lift-over reported to the -u unmapped file, and success is limited by divergence. Defaults -a 0.5 / -s 0.5.
Trigger
when the claim rests on a Liftoff / minimap2 pairwise annotation-transfer run
Research sharma-hiller-2017-cesar2

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeOn DemandVerbatimCorpus Observed
Purpose
Establishes exon projection as an HMM codon/frame- and splice-site-aware step whose known hard case is shifted splice sites; grounds 'exon-projection failure' as a distinct artifact class separate from true deletion.
Trigger
when an exon-projection or splice-site-shift artifact is at issue
Research fiddes-2018-cat

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeOn DemandVerbatimCorpus Observed
Purpose
Multi-method integration substrate (TransMap + AUGUSTUS variants + homGeneMapping over a Cactus/HAL alignment); grounds that projection is coverage-filtered and that input-annotation errors propagate across a clade.
Trigger
when the claim comes from a multi-method / consensus annotation set (CAT)
Research armstrong-2020-cactus

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeOn DemandVerbatimCorpus Observed
Purpose
The reference-free WGA substrate (HAL); grounds the reference-choice-bias check and that alignment accuracy degrades with divergence (F1 0.989 primate vs 0.795 mammal).
Trigger
when reference-choice bias or the upstream whole-genome-alignment substrate is at issue

audit-gene-loss-call

Referee a claim that a gene is lost / absent in a query genome, or that an annotation-projection result establishes loss. Decide whether the claim is an established loss or a technical artifact (assembly gap, missing/absent annotation, paralog-as-ortholog, exon-projection failure, out-of-range divergence, single-species/private mutation, or intactness-vs-expression conflation).

Work the gates in order; each gate can down-grade the verdict, and any gate that fails means the loss is not established. Do not certify a loss that no gate has positively cleared — silence is not evidence.

Gate 0 — What is being claimed, and by what tool?

Establish (a) the exact claim (this gene is lost / absent / non-functional?), and (b) the producing tool/method. This routes every later gate.

  • If the tool is a loss detector (TOGA transcript-status classification per kirilenko-2023-toga; aggregate-erosion + phylogenetic filter per turakhia-2020-gene-loss), continue.
  • If the tool is an annotation-transfer / lift-over tool being read as a loss detector, go straight to Gate 1a and expect a flag.

Gate 1a — Tool-appropriateness (invalid-tool-use)

A gene missing from an annotation-transfer output is a failed transfer, not an established loss. Per shumate-salzberg-2021-liftoff, Liftoff maps annotations for the same or a closely related species, writes genes it cannot map to a separate unmapped-IDs file (-u), and its success is explicitly limited by reference↔target divergence; it makes no loss claim. Plain armstrong-2020-cactus halLiftover projection and fiddes-2018-cat coverage-filtered TransMap are likewise transfer, not loss detection.

  • Verdict if the loss rests only on absence from a transfer output → invalid-tool-use; not established. Downgrade to “unmapped / needs a loss detector.”

Gate 1b — Assembly-gap / missing-data (not-established vs true absence)

Does the claimed-lost locus overlap an assembly gap? Both anchors treat this as the first safeguard:

  • kirilenko-2023-toga calls an absent exon M (missing) rather than D (deleted/diverged) when the expected query locus overlaps an assembly gap of >=10 consecutive N; at transcript level <50% CDS present → missing (undecided), not lost.
  • turakhia-2020-gene-loss masks chain-alignment gaps that fall in query assembly-gap regions so they are not counted as deletions.
  • Gene-count / annotation completeness tracks assembly contiguity (N50), so an absent gene in a fragmented assembly is expected without any true loss.
  • Verdict if the locus is gap-explained → not established (missing); never “true loss.” [GAP: no note gives a quantitative fraction-of-gene-in-gap threshold that flips missing→callable beyond TOGA’s <50% CDS present → missing binary and the >=10 N gap definition; a ”% of exon bases in gap” cutoff would need another source.]

Gate 2 — Absence-of-annotation is not loss

Is the loss inferred from the gene simply not appearing in the query’s own annotation set? Per turakhia-2020-gene-loss, filtering on “lacks an annotation” alone gives FPR > 30%, versus < 10% for the full chain-mapping + outlier + phylogenetic pipeline. A valid loss call must work from the raw assembly + alignment, not from a downstream annotation’s absence.

  • Verdict if loss = “not in the annotation” → not established.

Gate 3 — Ortholog vs paralog / synteny soundness

Is the compared locus the true ortholog? A paralog (or processed pseudogene) mistaken for the ortholog inflates apparent divergence/loss.

  • kirilenko-2023-toga: gene–chain pair treated as orthologous only at orthology score >= 0.5; chains with alignment score < 15,000 are unclassified; conserved intronic/intergenic context is what separates orthologs from paralogs/processed pseudogenes.
  • turakhia-2020-gene-loss: require best chain’s score >=20x the second-best (ortholog/paralog separation), gene-in-synteny >= 20 (aligning-block bases >=20x the gene’s own bases), and a unique 1:1 mapping (discard all overlapping mappings).
  • Verdict if the ortholog assignment is unsupported → not established (possible paralog).

Gate 4 — Divergence within calibrated range

Is the query within the producing tool’s calibrated divergence?

  • kirilenko-2023-toga: reliable within ~0.55 Ks (substitutions/neutral site); intronic/intergenic signal degrades beyond ~0.6 Ks; clade-specific references recommended for distant queries.
  • armstrong-2020-cactus: alignment accuracy itself falls with divergence (F1 0.989 primate → 0.795 mammal), so the upstream substrate is weaker at distance.
  • A projection/loss call beyond the calibrated range must be marked low-confidence, not passed at face value. [GAP: no note supplies a divergence-in-Myr limit for any tool — TOGA is stated in Ks, CESAR and Liftoff give no numeric divergence or percent-identity floor. Any “~N Myr” or “~X% identity” bound must be flagged as unsourced, not invented.]

Gate 5 — Exon-projection / single-artifact robustness

Does the loss rest on a single frame-disrupting event, or one near an exon boundary?

  • sharma-hiller-2017-cesar2: exon projection is codon/frame- and splice-site-aware; its named hard case is a shifted splice site — a projection failure that can mimic a disrupted exon rather than a true deletion.
  • kirilenko-2023-toga: only inactivating mutations in the middle 80% of the CDS drive uncertain-loss/lost calls (first/last 10% disregarded); “lost” requires max percent intact reading frame < 60% AND inactivating mutations in >=2 coding exons (scaled rules for many-exon / single-exon / large-single-exon). Base-call errors producing frameshifts/stops are counted as inactivating and can inflate loss.
  • turakhia-2020-gene-loss: excludes the first and last two amino acids of each exon (boundary-shift guard) and uses aggregate erosion (fraction AAs deleted + fraction substituted) rather than single events, explicitly to avoid spurious calls from e.g. a stop codon near a gene’s end.
  • Verdict if the loss is a single boundary-proximal or single-exon event without corroboration → uncertain / not established.

Gate 6 — Phylogenetic support (fixed & ancestral, not private)

Is the loss confirmed across lineages, or seen in one individual/species?

  • turakhia-2020-gene-loss: a high-confidence loss requires >=2 species in the group marked eroded (all transcripts MD > 15) AND no leaf in the group sequence-conserved (MD <= 5), with an outgroup lineage retaining the intact ortholog. This guards against private (individual-specific) mutations and reference-genome bias. The undetermined zone 5 < MD <= 15 is left uncalled, not forced.
  • Verdict for a single-species loss with no outgroup control → low-confidence / not established. [GAP: MD is Turakhia-specific (aggregate-erosion) and its per-species covariance fitting (mu, S) is not recovered in the note; the notes give no general cross-tool rule for minimum lineage support when the caller is TOGA-style single-event rather than aggregate-erosion — treat ”>=2 supporting species + intact outgroup” as the anchored convention, flag if applied outside Turakhia’s method.]

Gate 7 — Intactness / coding-capacity vs function (scope overreach)

What is the claim actually licensed to say?

  • Both kirilenko-2023-toga and turakhia-2020-gene-loss detect coding-capacity / sequence loss — frame-disrupting mutations, exon/gene deletion, aggregate erosion — NOT expression loss or pseudogenization without sequence change, and not regulatory/non-coding loss. Turakhia concedes it is “nearly impossible” to prove a locus emits no transcript.
  • Therefore: an “intact” call presented as evidence of retained protein function must be flagged (intactness ≠ expression ≠ function). A “lost”/eroded call presented as proven expression silencing must be flagged (sequence erosion ≠ demonstrated expression loss).
  • Verdict: downgrade any claim that steps from sequence status to functional / expression status → coding-capacity-only; overreach flagged. [GAP: none of the six notes provides a method for establishing expression loss (transcriptomic/proteomic evidence); a function-level claim needs a source outside this set — do not invent an expression criterion.]

Gate 8 — Reference-choice bias

Was a single, distant reference used where a clade-specific one is indicated?

  • kirilenko-2023-toga recommends clade-specific references for distant queries; a lost call from a distant reference may reflect reference bias, not query loss.
  • armstrong-2020-cactus is reference-free (reconstructs ancestors), so a Cactus/HAL substrate avoids single-reference bias — a point in favor when present.
  • Verdict: flag single-distant-reference loss calls as reference-choice-bias risk.

Verdict synthesis

  • Established loss — cleared Gates 1a–8: a loss detector, not gap-/annotation- explained, ortholog-sound, in calibrated range, multi-exon/aggregate signal, phylogenetically supported, claimed only at coding-capacity level.
  • Not established (missing / artifact) — failed Gate 1b, 2, or 3.
  • Invalid-tool-use — failed Gate 1a.
  • Low-confidence — failed Gate 4, 5, or 6.
  • Overreach flagged — passed as a loss but the claim exceeds coding-capacity (Gate 7) or rests on a biased reference (Gate 8). [GAP: the notes supply per-gate thresholds but no integration rubric for weighing gates against each other into a single confidence score — the ordering above is a design synthesis across sources, not a rule any one note states. Flag as convention pending a source that defines combined-confidence scoring.]