The review 05: the Nature journal-editor perspective (an independent assessment, 2026-09-28) • msuiter

The reviewer persona: a senior editor responsible for computational-biology method papers (a Nature/Nature Methods perspective), desk assessment, honesty first. The one-sentence overall assessment: this project’s research quality and engineering hygiene are in the top 5%, but packaged as it currently is amounts to playing a hand of good cards at Genome Biology odds.

1. Venue: the adjudication (the conclusion first)

Executed as currently planned, this is not a Nature paper, and most likely not a Nature Methods paper either. It is (if written well) a Genome Biology or Nature Communications paper.

The field’s own publication record (more convincing than any argument):

The flagship tool/method	The actual publication destination
MuSiCal	Nat Genet 2024
SigProfilerExtractor	Cell Genomics 2022
SigProfilerMatrixGenerator	BMC Genomics 2023
PASA/mSigAct	Brief Bioinformatics
SUITOR	PLoS Comput Biol
BASCULE	Genome Biology 2025
The Medo benchmark	Nat Commun 2024
Cornet	bioRxiv (2026-09, still a preprint)
PCAWG/Degasperi/Everall	Nature/Science/Nat Genet (discovery papers)

The pattern: in this field, Nature/NG goes to the discovery papers; the tool and method papers land in Cell Genomics, Genome Biology, PLoS CB, Brief Bioinform, Nat Commun. The “architecture/performance/applications” three-act structure is a Bioinformatics/Genome Biology structure — aligning the three-act with “Nature-grade” is a category error.

The only statistical contribution big enough for the NM/NC narrative core: the MSU-Fit / MSU-LowCount / MSU-Infer cluster — “exposure inference’s calibrated statistics”.

2. The novelty-tier ranking

MSU-Fit: the first inference layer that performs empirical coverage-rate measurement on the exposure intervals — the open question Medo 2024 left in black and white, with no incumbent occupying it; falsiable, benchmarkable. The only contribution in the whole plan big enough for the top journal’s narrative core.
MSU-LowCount: principled inference (conformal) for panel/ctDNA — the real wilderness after SigMA, with the biggest clinical sample pool.
MSU-Infer: the formal statistics of CoDA/differential abundance — no competition, and it makes “the statistics of exposure” a complete subject.
The five-modality end-to-end + the Catalog IR + the versioned refdb — a real ecological niche but of Genome Biology grade.
The Rust performance — real but worth only one “enabling” sentence; the Alexandrov group is moving toward matrix generation/QC, and a speed moat has a shelf life.
MSU-Extract — an engineering synthesis of the SigProfiler consensus + the SUITOR CV. Reviewer’s one sentence: “engineering synthesis of SigProfiler and SUITOR, in Rust.”
MSU-Corr — the highest danger. Cornet is the Park group’s preprint from two weeks ago, and formal publication will most likely be NG/NM grade; it will be written up as “derivative of Cornet”. Unless the five-modality joint makes a discovery Cornet cannot, demote it to a component.
MSU-Tempo/Tensor — application components, out of the abstract.

The eight MSU-* names must be cut down to 2–3 — once a reviewer sees through one “wrapper”, every name depreciates together.

3. The competitive positioning

The four power centers (Alexandrov/Park/Rozen/Nik-Zainal) each guard their own wedge, and the “integrated ecosystem” road is exactly the one they structurally will not walk — no head-on competition, but also structurally low renown.
“exposure inference’s statistics” (the calibrated CIs, CoDA, the conformal low-count) is the middle ground none of the four groups has taken: Rozen has the tests but no interval calibration, Park has the sparse point estimates but no coverage verification, DCEG only handles the K selection. This is the only high ground one can take first and, once taken, hold with defensive depth.
Under the current plan, a reviewer will read msuiter as a “derivative of SigProfiler+MuSiCal engineering synthesis”. The defense is not better engineering but having one thesis sentence nobody else has.
The author is the sigminer author — a double-edged sword: credibility credit, but it will be read as “sigminer v3 redone in Rust”. The paper narrative must make the statistics of MSU-Fit stand on its own independent of the author’s prior work.

4. What the three-act framing lacks

A tool-enabled discovery as the headline — the Cornet preprint is itself the template (a method + “SBS92a tracks the ERCC2 platinum status”). The current application cases are all “run a workflow” on public data — decorative.
Adoption/community evidence — “will a domain PI switch pipelines tomorrow”.
The framing opportunity: speed is not the product, speed is the precondition that “calibrated statistics are affordable” — the bootstrap/conformal/CV are all computation-hungry, and SigProfiler’s day-scale runtimes make coverage-calibration studies de facto unaffordable. “We are fast so that we can afford statistical honesty” is a sentence no competitor has said and that is intrinsically self-consistent — write it into the abstract.

5. The top-journal evidence standards

The benchmarking surface already meets the bar (the three-protocol bit-for-bit replication is rare). Note that PASA has already shown DBS/ID to be MuSiCal’s soft ribs — attack there, and do not fight a draw against MuSiCal on SBS.
The calibration surface (what the headline needs): the nominal-vs-actual coverage-rate heat map of 13 tools × a TMB grid — if the mainstream tools’ “95% CI” really covers only 60–80% under clinical TMB, this figure is Nature Methods’ Figure 1. It must be made real, otherwise the narrative collapses.
The real-data consequence (currently wholly missing): the calibrated CI must change one published real conclusion (e.g.: the Nacer 2026 disagreement of the 7 HRD methods vs the FDA assay, re-adjudicated with calibrated uncertainty).
Three-one choice to turn the case studies from decorative to forceful: (a) a treatment-history cohort + real treatment-record concordance; (b) an RNA-seq-only cohort (the matched DNA-RNA subset giving the concordance, the unmatched majority giving the new information); (c) the conformal validity on low-count panels + MUTYH/MSI diagnostic-grade detection.

6. The strategic risks, Top 8

#	The risk	The severity	The mitigation
1	MSU-Corr gets crowded out by Cornet’s formal publication	high	The five modalities are what Cornet lacks — if a multi-modality-unique result is produced within 12 months it stays, otherwise demote; promise nothing in the abstract
2	No discovery / no clinical consequence in the cases = a top-journal one-vote veto	high	Start the cohort collaboration (treatment records/RNA-only) at once, ahead of any code
3	Timeline fantasy (solo, realistically 12–18 months to submission)	high	Cut the scope; re-review the competitive landscape quarterly
4	The 8 MSU-* dilute the single true contribution	medium-high	Name only 2–3; demote the rest to features
5	bus factor = 1; without a statistics co-author the conformal/CoDA theory will be attacked	medium-high	Find a statistics co-author + a clinical/cohort collaborator + a co-maintainer
6	The “sigminer rewrite” perception	medium	Route the narrative around the sigminer lineage, standing on its own via the MSU-Fit statistics
7	The coverage calibration may falsify itself (the competitors’ CIs may not be that bad)	medium	kill criterion — spend two weeks first on the minimal coverage experiment (2 tools × a TMB grid) before writing any narrative
8	The R-only audience is bounded; the COSMIC licensing	medium	A CLI + a parquet/arrow exchange layer makes it consumable by Python; the licensing is already prudent

7. The candidate paper framings

A (the recommended main thrust) “Calibrated inference of mutational signature exposures”: never one of the field’s exposure intervals has been verified; the first complete inference layer of the LRT + the support-conditional bootstrap + the coverage calibration + the conformal low-count; the Rust performance makes the calibration feasible at cohort scale in minutes. Target: Nature Methods (the sprint) / Nat Commun (the realistic landing). Figure 1 = the 13-tool × TMB coverage-rate heat map (a red wilderness) + the calibration curves.
B (the upgraded fallback) “A unified, bit-compatible ecosystem for five classes of mutational signatures”: Target Genome Biology (high probability) / NC. Figure 1 = the capability-coverage matrix vs all competitors + the speed wall + the 12k WGS end-to-end.
C (the highest ceiling, needing collaborators) “Signature surveillance beyond WGS”: principled inference for panel/ctDNA/RNA-only + real clinical cohort validation. Target: NC; if the clinical validation is strong, NM is in play. The only path that can grow into a discovery story.
A+B can merge; A+C merging is the combination that maximizes the top-journal probability.

8. The probability estimates (the honest version)

The journal	The current plan executed smoothly	The Framing A/C restructure + a real cohort
Nature Methods	~5%	15–25%
Nature Communications	20–30%	35–45%
Genome Biology	50–60%	60–70%
Bioinformatics (the floor)	75–85%	same grade

Nature/NG: only if Framing C grows into a real discovery paper does it enter the conversation.

9. Kill criteria / the pivot triggers

The extraction benchmark shows parity and not advantage → switch the headline to the Framing A calibrated statistics, with GB as the floor.
The pilot coverage experiment shows the competitors’ CIs are basically calibrated → the MSU-Fit headline is falsified → accept the Framing B, do not force-manufacture a difference.
Cornet is accepted at NG/NM and extends to DBS/ID (before M5) → demote MSU-Corr.
Before the end of M3, no real cohort/clinical collaboration is locked → the case studies are destined decorative → proactively accept the GB/NC ceiling.
SPA/MuSiCal publish a calibrated CI/conformal → the gap closes → pivot to the application/discovery papers.
The M3 schedule slips >100% → cut CN/SV/Tensor/RNA, ship only the SBS/DBS/ID + the MSU-Fit core.
Persistently solo with no co-maintainer → find the co-maintainer before M1.