The review 01: the computational biologist (an independent review, 2026-09-28)

The reviewer persona: a senior computational biologist in the mutational-signature direction, who has led Nature/Nature Genetics-grade signature atlas projects repeatedly. The review objects: vignettes/articles/architecture.qmd, ROADMAP.md, research 00–08. The overall verdict: conditional trust — as it stands it cannot be sent to review; after the P0/P1 repairs a solid paper can be produced.

P0 (the blocking grade)

1.[P0] The “coverage-calibrated CI” of MSU-Fit defines no estimand, and the Fisher information interval is invalid at the boundary

• The attack position: ARCHITECTURE §5 MSU-Fit; ROADMAP M3; the 02 report §2.
• Why it is a scientific problem: (a) Estimand missing — is the calibrated CI that of the exposure fraction or of the absolute counts? A mixture makes the coverage uninterpretable. (b) The boundary mass — the true value is very often exactly 0 (the sparse support); the percentile bootstrap and Wald/Fisher necessarily under-cover at the boundary; the χ²₁ LRT is the right machinery for the boundary problem (Self–Liang), but the support-conditional and the unconditional are two different estimands. (c) Once the CI zeroing rule is folded into the pipeline, the composite pipeline coverage must be measured, otherwise the numbers are fake. (d) The generative model of the calibration: the field’s test framework (mSigAct) defaults to NB size=8 — the real catalogs are severely over-dispersed; a CI calibrated under the multinomial necessarily fails on the real NB data.
• The fix: nail down the three-layer estimand: ① presence = the χ²₁ LRT; ② magnitude|active = the support-conditional bootstrap CI, calibrated under both the multinomial and the NB (the Jiang dispersion protocol), reporting the empirical coverage curve by N and the signature Shannon entropy; ③ the CI zeroing gets its own separate composite-coverage measurement. The Fisher interval is demoted to an interior-only diagnostic or deleted.

2.[P0] The HRD workflow overreaches: the HRDetect 2017 coefficients are WGS-trained, yet the “ensemble uncertainty” comes with no ensemble model

• The attack position: the 03 Top-10 #1 (“supports WES/WGS”); ROADMAP M6.
• Why it is a scientific problem: (a) The six HRDetect features are computable only with WGS-grade input; applying the frozen coefficients to WES/panel is a patient-harm-grade risk. (b) LOH/LST/TAI need allele-specific CN — the input contract (the caller whitelist) is unspecified. (c) The archive cites the Nacer 2026 7-method disagreement, yet waves it past with one “ensemble + uncertainty” sentence — no weights, no calibration cohort, no pre-registration.
• The fix: the HRDetect-coefficient mode WGS-only; the allele-specific CN caller whitelist (Battenberg/ASCAT/FACETS/PURPLE/SEQUENZA); the WES mode must be marked experimental and recalibrated on an independent cohort; the ensemble rule pre-registered; the “HRDetect coefficient-validation test” is renamed the “reference-fidelity test” (it proves only that the coefficients were copied right, not that the WES features are interchangeable).

P1 (the major grade)

3.[P1] The variant-class coverage gap: the MNV split, the complex/double-indel routing and the DBS reconstruction after VCF splitting are all undefined

• SPMG explicitly separates the adjacent SNVs at distance 2–5bp (the MNV); only distance exactly 1 enters DBS78; the hts-specs normalization splits an MNV into consecutive SNVs — a naive counting will systematically inflate the SBS channels and DBS78 cannot be reconstructed from the split VCF; the complex-indel routing destination is undeclared.
• The fix: the catalog-completeness contract: enumerate the routing destination of every input-event class (SNV/DBS/MNV2-5/MNV>5/complex indel/double indel), with the skip ledger entering the provenance; the reconnection-algorithm specification for the adjacent records of a split VCF; the MNV-adjacency golden fixtures into M1.

4.[P1] The Cornet 10–40× recalibration is a model-prior temperature knob, not a data-hygiene standard (and it is [search]-grade unverified)

• For a pure multinomial/KL objective, multiplying everything by a constant changes no estimate — the rule is meaningful only under scale-sensitive regularization (the adaptive σ², the volume term); it also changes the CV deviance scale and the absolute-exposure semantics (a de-rescale is needed, unmentioned).
• The fix: demote it to an MSU-Corr-internal hyperparameter; the M7 grid sweep scale ∈ {1,10,40,100}, with the flat-signature under-estimation as the guard metric; explicitly forbidden on the NMF path.

5.[P1] The GMM hypermutator handling is misnamed “normalization”; the Degasperi common/rare two-stage has no owner in the architecture

• The GMM step of SigProfiler is classify and exclude from the de novo, then refit — it is a stratification, not a normalization; what gets excluded are exactly the APOBEC/POLE/MMR samples, and the cutoff directly decides whether these signatures can be extracted at all; the two-stage strategy has no MSU-* claimant, no milestone, no workflow.
• The fix: ms_stratify_hypermutants() (the explicit GMM cutoff + the de novo exclusion + the forced refit); promote common/rare to an atlas workflow into M6; add the 5–10% hypermutator arm (TMB 100×) to the benchmark, reporting the APOBEC/POLE P/R sensitivity to the cutoff.

6.[P1] The benchmark lacks the two most realistic failure axes — purity and caller dependence; the CN/SV count intervals are mis-scaled

• A purity of 0–70% produces mixed spectra and subclonal structures; two callers on the same BAM give different catalogs (the archive itself cites 41427300); the N of CN/SV is the segment/rearrangement count (tens–hundreds), so the TMB 10²–10⁶ does not apply.
• The fix: add purity ∈ {1.0, 0.7, 0.4, 0.2}; the caller axis (≥2 SNV + ≥2 indel callers); the class-specific count intervals (SV 10–10³, CN 10–10⁴); the joint scenario of a flat-signature group at r=0.8 co-existing with a rare signature; the subclonal exposure structure.

7.[P1] The scientific delta of MSU-Corr is the weakest: the simpler baseline “MvNMF extraction + correlation-aware refit” is not excluded; m=K is over-parameterized

• The volume regularization and the embedding are bundled together with no ablation; h=exp(α+β+l·u) at m=K carries the rotation/re-parameterization unidentifiability; the five-modality joint forces the different count scales to share the exposure.
• The fix: a pre-registered falsiable claim (simultaneously beating MvNMF+correlated refit and the original Cornet over r∈{0.8,0.9,0.95}); write down the gauge-fixing method; m<K + the selection rule (BIC/CV); reduce the five modalities to the SBS+DBS dual-modality demonstration until proven.

8.[P1] The RNA workflow is dishonest about its upstream dependencies

• The COSMIC RNA-SBS comes from a specific calling pipeline and filter chain; fitting variants from other pipelines to this reference inherits caller artifacts heavier than the DNA case; the 192 channels need stipulating which chain’s context is used; the DNA-RNA concordance needs a paired cohort and a pre-set metric.
• The fix: the minimal input requirements (RNA-Mutect2 or STAR 2-pass + documented filters; the RNA editing exclusion; the matched-normal annotation); add the RNA-artifact signatures to the reference panel; the pre-set concordance estimand; lock the case studies to a paired cohort; mark the workflow research-use.

9.[P1] The “report all the evidence” K selection has no adjudication rule — it is a dashboard, not a method

• The failure mode is exactly the evidence contradicting one another (the silhouette plateaus on the flat signatures, SUITOR biases toward parsimony, Wilcoxon biases toward lenience); the “agreement rate with SPA” is weaker yet — agreement is not correctness.
• The fix: a default adjudication rule (e.g.: within the 1-SE of the CV argmin → the maximum stability → the most parsimonious presence) as the K decision of MSU-Extract, scoring the single decision rule with the F1-vs-K over the Islam protocol, replacing the agreement rate.

P2

10.P2 The genome hygiene is missing: the chromosome whitelist, the sex opportunity correction, the alt contigs, chrM, T2T

• The opportunity vector must match the callable domain (the male chrX, the aneuploidy segments, the alt-contig double counting); the “across-build derivation” concept is muddled: the ID83/CN48/SV32/RNA-SBS labels themselves are build-irrelevant, the signature matrices are build-agnostic — the real dependency lies only in the exome/panel opportunity catalogs.
• The fix: the contig strategy + the per-sample opportunity vector (the CN segments + the sex + the callable BED); rewrite §6.2; the exome/panel opportunity-normalization reference set is the true gap.

11.P2 The “first/gap” class claims are over-claims against the 2023–2026 literature the archive itself cites

• sigfit/signeR/BayesPowerNMF all produce nominal intervals — the “first” must be narrowed to “the first empirical coverage-rate measurement of the exposure intervals”; SigNet already does the low-count refit; Morrill 2025 already is a formal compositional framework; the conformal “distribution-free” does not hold under heterogeneous N, needing a grouped/weighted stratified calibration.

12.P2 The multi-sample VCF and the germline/somatic ambiguity: the validation contract does not catch it

• The MsVariants validator does not validate the somaticness signals (the FILTER, the VAF floor, the matched-normal provenance); under ctDNA the CHIP/germline contamination is the first failure mode.
• The fix: the extended validation (the caller+normal provenance mandatory, the VAF floor, the confidence-class pass-through, the low-VAF sensitivity curve).

The overall verdict and the highlights

• Conditional trust; the benchmark paper (the panel ②) with the design + the problem-6 added axes has a real contribution; the three headline claims of the panel ① are in the “a serious reviewer will pierce it with a single page” state; writing HRD/RNA into the paper as they stand is clinical danger and scientific dishonesty.
• The annex note: M0–M8 ≈ 15–20 weeks is unrealistic; cut MSU-Tensor/MSU-Tempo down to watch.
• The three things done well: the source-level fact-check discipline; the statistical stance is correct and consistent (KL=the count MLE, the cosine demoted, the χ²₁ boundary, the precision/recall stance); the honest competitor and licensing positioning.