# Mock Peer-Review Report **Manuscript pair:** - **Paper D** (Nature Perspective) — *From disease name to design specification: a self-correcting AI-agent campaign for eosinophilic esophagitis* - **Paper E** (Research article / bioRxiv) — *An integrated computational pipeline for eosinophilic esophagitis: from public omics to therapeutic targets, companion diagnostics, and a preclinical program* **Review model:** Three independent expert reviewers + handling-editor synthesis. All reviewers were briefed that the work is human-supervised and AI-generated (a disclosed part of the contribution, not a defect) and that wet-bench/clinical experimentation is explicitly out of scope — every requested revision is achievable in silico or by editing. **Panel:** - **Reviewer 1** — Senior EoE clinician-scientist & mucosal immunologist (translational/clinical lens) - **Reviewer 2** — Computational biologist & biostatistician (ML for genomics) - **Reviewer 3** — AI-in-science methodologist & research-ethics reviewer **Headline outcome:** All three reviewers independently recommended **Major revision** for both manuscripts. *Generated 2026-07-08 · reviews produced by independent model instances under distinct expert personas.* --- ## Editor's Decision Letter # Editor's Decision Letter **Manuscript pair:** Paper D ("From disease name to design specification: a self-correcting AI-agent campaign for EoE," Nature Perspective) and Paper E ("An integrated computational pipeline for EoE...," research article) Dear Authors, Thank you for submitting this linked pair. All three reviewers found real value here — a genuinely useful taxonomy of agentic error modes, an ethically substantive HLA-equity catch, a defensible in-silico discovery funnel, and consistent, welcome candor about limitations throughout. All three reviewers also converge, independently and from different disciplinary angles (clinical, biostatistical, AI-methodology/ethics), on the same underlying problem: several headline claims in both papers are asserted with more confidence than the evidence underneath them supports, and the two papers currently reinforce each other's unaudited claims rather than independently substantiating them. Neither paper is rejected; both need focused, entirely in-silico/editorial revision before we can proceed. --- ## Decision: Paper D — **Major Revision** ## Decision: Paper E — **Major Revision** All three reviewers independently landed on major revision for both manuscripts, with converging rationales. This is a unanimous, not a split, decision. --- ## Essential Revisions — Paper D 1. **Ground the correction ledger with a denominator (R1, R2, R3 — all three, independently).** The 8-episode ledger and 12-"principal-finding" novelty classification are currently presented without any accounting of the total pool of claims/decisions from which they were selected. Authors must state (a) how many discrete analytical claims/decisions the campaign generated in total, (b) the selection rule for what entered the ledger and the 12-finding set, and (c) whether the ledger was compiled prospectively or reconstructed post hoc for the manuscript. Without this, "self-correcting" and "calibrated" are asserted, not measured — R3 in particular notes this is a curated highlight reel unless a denominator is supplied. 2. **De-escalate the index-case as evidence (R1, R2, R3 — all three).** The blinded single-patient food-trigger test is repeatedly invoked across both papers as the campaign's strongest external validation. All three reviewers regard this as rhetorically overweighted for n=1. Required: reframe uniformly as "a single illustrative case," report the actual candidate-food denominator and a chance-level/null expectation for the observed ranking (R2's specific request for a permutation calculation), and remove or soften "strongest single piece of external validation" language (R1, R3). 3. **Scope "self-correction" to what is actually demonstrated (R1, R3, echoed by R2's point that none of the 8 episodes touch the statistical core).** Only one episode (S100A8/9→S100A4) is genuine unprompted machine self-correction; the decisive catches (MHC-I/II, HLA-equity, missed TCR paper) were human. The title/framing should be adjusted or explicitly scoped — e.g., distinguish "self-correction" from "human-supervised correction" — and the paper should state plainly that none of the ledger episodes reversed a core statistical/methodological claim (batch correction, meta-analysis, cell-type annotation), only conceptual/ethical ones. 4. **Update the therapeutic-landscape framing (R1).** Dupilumab's current FDA approval status for adult and pediatric EoE should be reflected accurately; sharpen the actual gap (mast-cell-directed therapy) rather than implying general undertreatment. 5. **Housekeeping (R1, R2 convergent):** confirm ref. 15 (Jacobse et al.) is actually cited in text; complete placeholder competing-interest/author-contribution sections (R3 rightly flags this as a live irony for a paper about disclosure and trust). --- ## Essential Revisions — Paper E 1. **Attach caveats at first mention, not after the headline claim, especially in the Abstract (R1, R2, R3 — all three).** The "treatment-refractory core" (9/10 targets) and the diagnostic-addressability figures (≥92–100%) are stated with declarative confidence in the Abstract/Results before their single-cohort or circularity caveats appear. The hedge in the Discussion must migrate to the same sentence as the claim, and the Abstract must carry it too. 2. **Report concordance statistics for mhcnuggets vs. netMHCIIpan-4.1 (R1, R2, R3 — all three, in near-identical language).** "Cross-validated" currently names two tools without reporting any actual agreement metric (rank correlation, AUC vs. IEDB-annotated epitopes, top-k overlap). This is the single most-repeated specific request across all three reviews and gates the credibility of the entire antigen-directed arm (Fig. 7). 3. **Address circularity/independence of validation cohorts (R1, R2, R3).** Diagnostic addressability (≥92%), refractory-core persistence, and endotype-assignment statistics all appear to be computed on the same discovery cohorts used to derive the targets/signature. State explicitly whether any held-out or independent data was used; where not, relabel figures as upper-bound/derivation-cohort estimates. 4. **Add a formal background/enrichment test for the GWAS-gene overlap (R2, R3 convergent).** Seven of 567 genes overlapping GWAS loci needs a hypergeometric or permutation test against a matched background gene set, not isolated per-gene odds ratios, to support any signature-level enrichment claim. 5. **Add heterogeneity/robustness diagnostics for the 9-cohort meta-signature (R2, echoed by R1's ancestry point and R3's multiple-comparisons concern).** Forest plots and I²/τ² for headline genes (CCL26, POSTN, DSG1 at minimum), plus a leave-one-cohort-out sensitivity analysis, are needed before the 567-gene signature can be called "high-confidence." 6. **Address SIGLEC6 tissue-selectivity / off-target mast-cell depletion risk (R1).** SIGLEC6 is expressed beyond esophageal mast cells (basophils, other tissue mast-cell pools); the therapeutic rationale needs an explicit in-silico caveat on tissue-selectivity, drawable from public expression atlases already within the pipeline's reach. 7. **Resolve the DQ2.2/celiac question and the DRB1-vs-DQ prediction gap (R1, R2 convergent).** State whether celiac disease was clinically excluded in the index patient given the DQ2.2 finding, and clarify how DQ-restricted (not just DRB1) presentation was actually modeled for the wheat "discordance" — Methods currently only describe DRB1 prediction, yet the wheat mechanism explanation invokes DQ2.2. 8. **Report unit-of-analysis and n for extreme p-values (R2).** The SIGLEC6 IHC statistic (P=3×10⁻³¹) needs sample size and confirmation the unit of analysis is patient-level, not cell-level (pseudoreplication risk); similarly for the refractory-core Fig. 4d result. 9. **Temper the sST2 diagnostic framing (R1).** "Repurposing an established cardiac-biomarker platform" glosses over sST2's known lack of esophagus-specificity and background population variability; this claim should not be presented as equivalent in confidence to the SIGLEC6/CCL26 assays. --- ## Notable Disagreements Requiring Author Adjudication - **None of the three reviewers substantively disagree on direction** — this is an unusually convergent set of reviews. The closest thing to a difference in emphasis: R1 focuses on clinical-plausibility gaps (SIGLEC6 tissue selectivity, sST2 specificity, DQ2.2/celiac) that R2 and R3 do not raise in the same depth, while R2 and R3 focus more heavily on statistical/methodological auditability (denominators, concordance statistics, enrichment tests) that R1 raises only in passing. **Authors should treat these as complementary, not competing** — both classes of gap need addressing, and neither reviewer's concerns supersede the other's. - **R3 raises the possibility of retitling Paper D** to remove "self-correcting" from the title outright, whereas R1 and R2 frame this more as a rhetorical-scoping fix within the existing title. Authors have latitude here: either retain the title with an explicit scoping sentence early in the paper defining what "self-correcting" refers to, or retitle. We do not mandate a specific resolution but will not accept an unscoped use of the term in the final version. --- ## Path to Acceptance — Paper D Acceptance is achievable through editorial and accounting work, not new science: supply the denominator behind the ledger and the 12-finding set, uniformly de-escalate the index-case language, explicitly scope the "self-correction" claim to what is actually demonstrated (one clean episode plus a human-supervised oversight loop), and update the therapeutic-landscape context. The taxonomy of error types and the HLA-equity exhibit are strong enough to carry a Perspective once the evidentiary overreach is corrected — this is a calibration problem, not a substance problem. ## Path to Acceptance — Paper E Acceptance requires the paper's confident headline framing to be brought into line with its own — generally honest — Discussion caveats, plus several concrete, computable additions: concordance statistics for the two epitope predictors, heterogeneity/leave-one-out diagnostics for the meta-signature, a formal GWAS enrichment test, explicit statement of validation-cohort independence (or relabeling as derivation-cohort upper bounds), and resolution of the DQ2.2/DRB1 modeling gap. None of this requires new wet-lab work or new cohorts — it requires re-running existing pipeline outputs through additional diagnostic/statistical lenses and rewriting the Abstract and Results to match what the Discussion already, commendably, concedes. --- We look forward to a revised submission that addresses these points. Given the extent of convergence across reviewers, we anticipate the essential revisions above, once implemented, will be sufficient for acceptance without need for a third round — but this will depend on the rigor with which the statistical scaffolding requests (especially items E2, E3, E5) are addressed. --- ## Full Reviews ## Reviewer 1 ### Reviewer 1: Senior EoE clinician-scientist and mucosal immunologist (translational/clinical lens) **Manuscripts assessed:** Paper D and Paper E --- ## PAPER D — "From disease name to design specification" **Brief summary:** This Perspective narrates the human–AI operating model behind an in-silico EoE discovery campaign, foregrounding a correction ledger of eight episodes (agent self-corrections, human domain catches, contained contradictions) as evidence of trustworthy agentic conduct. The disease findings are reported "compactly" as supporting evidence rather than as the primary contribution, which is the companion Paper E. **Strengths** - The explicit separation of MHC-I-vs-II framing (conceptual), the missed β-casein/TCR paper (evidential), and the HLA-burden enrollment design (ethical) is a genuinely useful taxonomy of failure modes, and the ethical catch is the strongest part of the whole pair — recognizing that HLA-stratified enrollment would encode ancestry inequity is exactly the kind of judgment that matters clinically and is rare to see surfaced this explicitly. - Refreshing candor that "presentation is a pre-test prior, not pathology" — this is clinically correct and is the single most important sentence in either manuscript. **Major concerns** 1. The index-case narrative is used as "the strongest single piece of external validation the campaign produced" — but N=1, unblinded to the analyst's foreknowledge that milk/soy are the most epitope-rich foods generally, and with wheat explained away post hoc as a "single discordance." A predictor that gets 2/2 true positives and then requires a bespoke mechanistic story to explain its one false positive is not "validated," it is "consistent with," and the paper's rhetoric ("strongest... piece of external validation") outruns what one patient can support. This needs to be tempered throughout — it reads as a triumphant case study, not calibration. 2. The claim that the agent's predictor "independently recovered the same DR7 β-casein window" after the panel was completed is presented as a validating coincidence, but the supervisor added β-casein to the panel *because* of the 2025 paper — so the "independent" recovery is confirmatory-by-construction, not blinded discovery. This should be stated more carefully; as written it risks over-crediting the model. 3. Presenting SIGLEC6/IL1RL1/CCL26 target selection, mast-cell biology, and antigen-directed strategy as "compact" background here while the real clinical reasoning lives in Paper E creates a risk that neither paper is independently checkable by a clinical reviewer — Paper D asserts biological correctness (e.g., "consistent with established EoE endotypes") without giving me the endotype definitions or effect sizes to judge it myself. 4. The framing of dupilumab and PPIs as a "narrow targeted-therapy landscape" undersells where the field actually is in 2024–25 (dupilumab is now FDA-approved for both adult and pediatric EoE with strong Phase 3 data); the Perspective should more precisely state what gap in the treatment landscape (mast-cell-directed therapy specifically) this campaign is filling, rather than implying the field is undertreated in general. **Minor concerns** 1. "Diagnosed and managed against well-established clinical criteria" cites ref 3 (Furuta 2007), which is outdated relative to current AGREE consensus criteria; should be updated or supplemented. 2. The 23-hour compute figure and 214-artifact count are presented with a precision that invites scrutiny disproportionate to their scientific relevance — this is fine for the Perspective's thesis but should not be conflated with rigor of the underlying biology. 3. Reference 15 (Jacobse et al., cited as "synthesis and subgroup analysis of tissue transcriptome") is listed but never cited in-text as far as I can find — check citation completeness. **Feasible revisions** 1. Reframe the index-case section explicitly as "n=1, hypothesis-generating, not validating" in the main text (not just Discussion-equivalent hedges), and remove or soften "strongest single piece of external validation." 2. Add a sentence disambiguating that the β-casein recovery was enabled by (not independent of) the human-supplied prior art, to avoid over-crediting agent autonomy. 3. Add 1–2 sentences updating the therapeutic-landscape context to reflect dupilumab's current approval status, sharpening the actual unmet need (mast-cell/refractory disease) the campaign targets. 4. Cross-check and complete in-text citation of all listed references (e.g., ref 15). **Recommendation:** Major revision — the operating-model narrative is a valuable and well-documented contribution for a Perspective venue, but the calibration claims built on the single index-case must be rhetorically de-escalated before this is publication-ready; the current framing overstates what an n=1 blinded test can establish. --- ## PAPER E — "An integrated computational pipeline for eosinophilic esophagitis" **Brief summary:** This research article reports a public-data-only pipeline that produces a 567-gene EoE meta-signature, validates it in single-cell data showing 14-fold mast-cell expansion, advances SIGLEC6/IL1RL1/CCL26 to structure-grounded design specs with companion diagnostics, and builds an MHC-II antigen-mapping arm calibrated against one patient's clinical food triggers, closing with a four-phase preclinical roadmap. **Strengths** - Appropriately explicit that "every candidate is a hypothesis generated in silico," repeated at multiple points including the abstract and Discussion — this hedging is well-placed and should be commended. - The HLA-agnostic enrollment design for the hypothetical Phase 4 trial is a genuinely sound piece of clinical-trial ethics reasoning, rare to see reasoned through correctly in a computational paper. - Recovery of canonical markers with correct direction (CCL26, POSTN, ALOX15, CDH26, TPSAB1, CPA3, DSG1, SPINK7, FLG) is a reasonable and appropriately modest positive-control demonstration. **Major concerns** 1. **SIGLEC6 biological plausibility is asserted, not established.** SIGLEC6 is described as "mast-cell-restricted" and a driver of "no existing EoE therapy depletes the expanded mast compartment" — but SIGLEC6 is also expressed on basophils and placental trophoblast, and critically, the mast-cell-depletion rationale needs to grapple with the fact that mast-cell-restricted depletion in EoE has an uncertain risk-benefit profile (mast cells have homeostatic/barrier roles too), and there is no discussion of why an ADC or afucosylated antibody against a receptor expressed on any mast-cell subtype elsewhere in the body (skin, gut, lung) would be tissue-selective. The paper needs to address on-target/off-tissue mast-cell depletion risk more explicitly, even as an in-silico caveat. 2. **The refractory-core finding is over-leveraged rhetorically despite the honest single-cohort caveat.** The Abstract calls it "a persistent, treatment-refractory molecular core" in confident declarative language, while the Results correctly flag "we carried it forward as a hypothesis rather than an established result" — but the abstract-level framing will be what most readers and press take away. The abstract needs the same hedge the body has. 3. **The companion diagnostic "≥92% severe-endotype addressability" figure is a definitional artifact of how "severe endotype" and "addressable" were computed from the same discovery cohorts used to derive the targets** — there is no independent validation cohort, and no discussion of sensitivity/specificity, positive predictive value, or how these thresholds would perform against clinical mimics (reflux esophagitis, other eosinophilic GI disease) that share overlapping transcriptomic/histologic features. A diagnostic claim of this specificity needs at minimum a discussion of differential-diagnosis performance, even if only addressable in silico via published comparator datasets. 4. **The sST2 companion diagnostic claim needs a stronger clinical caveat.** Repurposing "an established cardiac-biomarker platform" for serum soluble ST2 glosses over the fact that sST2 is a well-known non-specific marker of cardiac strain/inflammation broadly, with substantial background variability in the general population; no discussion is given of whether serum sST2 has adequate esophagus-specific signal-to-noise given systemic confounding — this is a materially weaker diagnostic proposition than the SIGLEC6 tissue IHC or CCL26 assays and should not be presented as equivalent in the addressability table. 5. **The index-case wheat/DQ2.2 result is clinically important but incompletely reasoned.** The manuscript correctly notes wheat gliadin is "heavily presented" via the patient's celiac-associated DQ2.2 but tolerated — yet does not address the obvious confound: was celiac disease clinically excluded in this patient? If DQ2.2 is present, that is the genotype for celiac susceptibility (much more classically DQ2.5, but DQ2.2 confers some risk), and gluten tolerance in an EoE patient with this genotype needs a sentence confirming celiac serology/biopsy status was negative, or the "instructive false positive" framing may be conflating two distinct antigen-driven diseases. 6. **MHC-II epitope prediction tool validation is thin for a clinical-facing claim.** mhcnuggets cross-validated against IEDB netMHCIIpan-4.1 is reasonable practice, but no accuracy/AUC metrics are reported for this specific application (food-allergen peptides, these particular DRB1 alleles) — given that the entire personalized-diagnostic concept rests on ranking prediction confidence, at minimum report the concordance statistics between the two predictors on this dataset. **Minor concerns** 1. Fig 4d states "nine of the ten prioritized targets... normalized in responders" while MSLN "lacked data" — worth clarifying this in the main text sentence itself, not only implicitly, since 9/9 available reads differently than 9/10 nominal. 2. GWAS odds ratios (CAPN14 1.77, DSG1 2.22) are given without confidence intervals or the source GWAS's sample size/ancestry composition — since the paper elsewhere is scrupulous about ancestry-fairness in trial design, the same lens should apply to whether these GWAS effect sizes generalize across ancestries. 3. The competing-interest statement flags that the antigen-directed concept "intersects disclosed prior art" without further elaboration — for a diagnostic/therapeutic claim this close to a specific published discovery (Dilollo et al. 2025), a clearer statement of what is and is not novel relative to that paper would help readers (and future IP counsel) immediately. 4. Dupilumab's mechanism (IL-4Rα blockade) is described as "broadly immunosuppressive," which somewhat overstates its safety profile relative to its actual clinical tolerability data; suggest softer/more precise language (e.g., "pathway-broad rather than mast-cell-selective"). **Feasible revisions** 1. Add an explicit paragraph addressing SIGLEC6 tissue-selectivity/off-target mast-cell depletion risk (skin, lung, gut mast cells), drawing on public expression atlases (e.g., Human Protein Atlas, single-cell atlases) already presumably within the pipeline's reach. 2. Revise the Abstract to carry the same "single paired cohort, hypothesis not established result" hedge on the refractory-core claim that appears in Results. 3. Add sensitivity/specificity or at least a differential-diagnosis discussion for the diagnostic claims, ideally by testing signature/marker overlap against a public reflux-esophagitis or non-EoE eosinophilic-GI transcriptomic dataset if one exists. 4. State explicitly whether celiac disease was clinically excluded in the index patient, given the DQ2.2 finding, or flag this as an open question the wheat-discordance interpretation depends on. 5. Report concordance metrics (e.g., rank correlation, AUC) between mhcnuggets and netMHCIIpan-4.1 specifically for the allergen/allele set used, not just cite that cross-validation was performed. 6. Temper "repurposing an established cardiac-biomarker platform" language for sST2 with an explicit note on background population variability and lack of esophagus-specificity. **Recommendation:** Major revision — the pipeline is a legitimate and often appropriately hedged in-silico discovery exercise, but several specific clinical-plausibility gaps (SIGLEC6 tissue selectivity, sST2 specificity, the DQ2.2/celiac question, single-cohort refractory-core claim inflating into the Abstract) need to be addressed before the paper's clinical claims are appropriately calibrated to its bioRxiv/research-article venue. --- ## Cross-cutting comment The pair is unusual in that Paper D's stated thesis — that the campaign's value lies in its calibration and self-correction — is itself only as credible as Paper E's underlying clinical reasoning, and my read is that the correction ledger over-indexes on the three showcased catches (MHC-I/II, missed TCR paper, HLA-equity) while several equally consequential clinical judgment calls in Paper E (SIGLEC6 tissue selectivity, sST2 specificity, the DQ2.2/celiac question) were not flagged as needing the same scrutiny, suggesting the "human domain catch" process was not applied uniformly. I'd also flag that the single index-case calibration is doing more argumentative work across both papers than an n=1 observation can bear, and both manuscripts should independently and explicitly cap its evidentiary weight rather than each treating it as the campaign's headline validation. ## Reviewer 2 ### Reviewer 2: Computational biologist / biostatistician — multi-cohort transcriptomics, batch correction, protein language models, MHC-II epitope prediction **Manuscripts assessed:** Paper D and Paper E --- ## PAPER D (Nature Perspective) **Brief summary** This is a process narrative arguing that a human-supervised AI agent can be trusted to run a multi-phase EoE discovery-to-design campaign because its errors were caught, logged, and corrected. The evidence offered is a "correction ledger" of eight episodes (one self-caught, three human domain catches, others calibration/containment) plus a claim that the agent's findings were mostly literature-anchored rather than novel. It explicitly defers the underlying biology to Paper E. **Strengths** - The three-way taxonomy of error types (machine self-correction, human-only catches, contained contradictions) is a genuinely useful framework, and the HLA-burden-recruitment episode is a strong, concrete illustration of a failure mode no statistical check would flag. - Honest framing of the human/agent division of labor (Table 3, Fig. 2) rather than claiming full autonomy. - The blinded index-case calibration is a legitimate, if narrow, piece of prospective validation and is appropriately flagged as single-patient. **Major concerns** 1. **The evidentiary base for "calibration" is entirely self-reported and unaudited.** The 8-episode ledger, the 12-finding novelty classification (8 grounded/3 novel/1 contradictory), and the claim that literature verification was done "against CrossRef" are all asserted by the same agent whose trustworthiness is under evaluation. There is no external or blinded audit of the ledger itself — no second reviewer (human or independent model) checked whether the agent's self-classification of "novel" vs. "grounded" is accurate, or whether other errors went uncaught. A trust argument built entirely on the trusted party's own bookkeeping is circular; at minimum the paper needs to disclose whether the ledger was constructed prospectively (during the campaign) or reconstructed post hoc, and whether any independent check was applied to it. 2. **Denominator problem for "self-correction."** Reporting 8 corrected episodes says nothing about calibration unless we know how many total claims/decisions were at risk of error. Twelve "principal findings" is a curated, presumably small, hand-picked list — the paper needs to state how "principal" was defined and by whom (agent or human), otherwise the 8-episode ledger could reflect selection of the interesting errors rather than a true error rate. 3. **The single strongest calibration claim (blinded index-case) is statistically thin and needs its numerator/denominator stated.** "Both true triggers ranked top-two" out of how many candidate foods, and what is the null-model expectation (i.e., probability of two true positives landing in the top two by chance alone, given N foods and ties)? Without this, "top-two out of ten" reads as impressive but is not quantified — a single patient is anecdote, not calibration, unless accompanied by a permutation/null calculation showing the observed ranking is unlikely under random presentation-based ranking. 4. **The refractory-core caveat is repeated as a virtue but not resolved.** Both papers rely on exactly one paired pre/post-treatment cohort for the "treatment-refractory core" claim, foregrounded here as a triumph of transparent uncertainty-carrying. Transparency about a single-cohort finding is good; but the Perspective repeatedly uses this same single-cohort result as one of the pillars of "what the campaign found," which is in tension with treating it as merely a flagged uncertainty. 5. **"Recovered every canonical marker with correct direction"** is repeated verbatim from Paper E without independent methodological scrutiny in this piece — for a paper about self-correction and calibration, no mention is made of whether the multiple-testing correction, cohort weighting, or heterogeneity statistics that underlie this claim were themselves checked or reversed at any point. If none of the ledger's 8 episodes touch the core statistical pipeline, that itself needs to be stated as a limitation (the human catches were conceptual/ethical, not statistical). **Minor concerns** 1. "23 hours of active compute... within a one-week campaign" is a strange metric to headline (Table 1) — active-compute-hours is not really a measure of scientific rigor and risks reading as a marketing figure; contextualize against actual analyst-hours-equivalent or drop the framing. 2. Figure/table cross-referencing (Fig. 3a "Blinded clinical calibration") should be double-checked against the actual figure legend in the linked artifact — the text says this appears in Fig. 3, but Fig. 3 is described as the correction ledger; ensure the panel is genuinely present in the artifact and not conflated from Paper E's Fig. 7c. 3. The claim that "two-thirds of its findings were anchored to existing work" (8/12) is presented as inherently reassuring, but 3 "novel" claims and 1 "contradictory" is also 4/12 (33%) unanchored/discordant — frame both directions. 4. Reference 15 (Jacobse et al., meta-analysis synthesis) is listed but never cited in text — check citation completeness. 5. The provenance graph (Fig. 4, 36 artifacts / 83 edges) is a nice touch, but the paper should state whether this graph was auto-generated and machine-verified for cycle-freedom/completeness, or hand-curated for the figure. **Feasible revisions** 1. Report, for the "principal findings" set, the total number of discrete claims/decisions the agent made (denominator) so 8/? or 12/? is interpretable as a rate, not a curated list. 2. Add a permutation-based null calculation for the blinded index-case ranking (e.g., probability that 2 truly-triggering foods land in top-2 of 10 by chance, and sensitivity to how "casein family completion" changed the ranking pre/post). 3. State explicitly whether the correction ledger was compiled prospectively (logged at time of error) or reconstructed retrospectively for the manuscript, and whether any party other than the agent/its supervisor reviewed the ledger for completeness. 4. Add one sentence clarifying that none of the 8 ledger episodes involved reversing a core statistical/methodological claim (meta-analysis, batch correction, cell-type annotation) — the human catches were domain/ethical, not quantitative — so readers don't over-generalize "self-correcting" to the statistical backbone. 5. Replace or supplement the "23 hours compute" metric with a more informative effort metric, or remove it as unnecessary color. **Recommendation:** Major revision — the framework and the ethical-catch episode are valuable, but the central "calibration" claim needs a non-circular evidentiary basis (denominators, null models, or an independent audit trail) before it can support the paper's trust argument, which is unusually important for a Nature Perspective making a methodological claim about agentic science. --- ## PAPER E (research article) **Brief summary** This is the science paper: a 9-cohort (235-sample) bulk transcriptomic meta-analysis integrated with a 166k-cell single-cell atlas nominates a 567-gene EoE signature, prioritizes three targets (SIGLEC6, IL1RL1/ST2, CCL26) into structure-grounded design specs with companion diagnostics, and layers on an MHC-II antigen-mapping arm calibrated against one blinded index patient, closing with a four-phase preclinical roadmap. **Strengths** - Appropriate general use of random-effects (DerSimonian–Laird) meta-analysis and BH correction for the bulk DE work — a defensible design in principle. - The mast-cell compositional result (14-fold expansion, Mann-Whitney P=7.2×10⁻⁵) is a clean, plausible, literature-concordant finding. - Genuinely commendable candor about limitations in the Discussion: single-cohort refractory core, single-patient calibration explicitly labeled non-clinical, IP/freedom-to-operate flagged, everything labeled in silico. - The HLA-agnostic trial-design correction is a substantively good methodological and ethical decision, well justified statistically (avoiding ancestry-confounded eligibility). **Major concerns** 1. **Multiple testing across the whole discovery funnel is not addressed holistically.** BH correction is stated for per-study DE, but there is no accounting for the multiplicity introduced across the full pipeline: 567 genes selected, then subset to druggable (120), then to 10 Tier-A, then 3 leads — each stage is itself a form of selective inference, and the headline p-values quoted (e.g., CCL26 adjusted P=3.4×10⁻⁴, mast-cell P=7.2×10⁻⁵) are being read by the reader as confirmatory even though they emerge from a highly multi-stage, partly overlapping selection process (single-cell composition test on the same axis used for signature interpretation). Please state whether single-cell composition/DE tests were pre-registered/hypothesis-confirmatory versus exploratory relative to the bulk signature, and correct across the union of tests actually performed (not just within one panel). 2. **The "9 case-control cohorts, 235 samples" harmonization needs batch-integration diagnostics.** Nine cohorts on presumably different platforms (microarray + RNA-seq mixed almost certainly, given EoE literature) pooled by random-effects meta-analysis is reasonable, but the paper gives zero heterogeneity statistics (I², τ², forest plots) for the headline genes, no description of how genes measured on some but not all platforms were handled, and no sensitivity analysis (leave-one-cohort-out) to show the 567-gene signature isn't dominated by 1-2 large cohorts. This is a fixable, in silico addition (forest plots + I² per gene, leave-one-out stability) and is standard for any meta-analysis claiming a "high-confidence" signature. 3. **The refractory-core claim (9/10 targets persisting in non-responders) is a single paired cohort with an unspecified sample size and unspecified statistical test.** "Nine of the ten prioritized targets... every one with paired-cohort coverage... normalized in responders but persisted in non-responders" (Fig. 4d) reads as a clean categorical result but no n, no effect sizes, no correction for multiple targets tested, and no discussion of whether responder/non-responder status was itself derived from the same dataset used to define the signature (circularity risk). This is the single most clinically consequential finding in the paper (it's used to argue these are the targets that "most need a novel agent") and it currently has the least statistical scaffolding. 4. **Single-cell annotation and the mast/basal-epithelial composition shift lack methodological detail needed to trust cell-type calls.** No marker gene lists, no clustering resolution/parameter sensitivity, no doublet-filtering description, no cross-cohort batch-integration method (Harmony? scVI? BBKNN?) specified beyond "standard QC... batch integration and Leiden clustering." Given eosinophils are explicitly noted as "under-captured by droplet methods" (per Paper D) yet eosinophil biology is central to EoE, please state how mast cells (also small in number, ~0.2–2.6% of cells) were distinguished from basophils/other granulocytes, and provide a cluster-stability metric (e.g., bootstrapped ARI) rather than a single Leiden run. 5. **ESM per-residue entropy/conservation as an epitope-conservation proxy is asserted without validation.** "Per-residue ESM conservation... confirmed that the proposed epitopes sit on conserved, structurally-defined surfaces" (Fig. 5) treats ESM pseudo-likelihood/entropy as a ground-truth conservation signal, but ESM per-residue scores reflect a mixture of structural/functional constraint learned from training-set homology, not literal cross-species sequence conservation, and are known to be confounded by secondary structure and burial. No comparison against an actual multiple-sequence-alignment-based conservation score (e.g., ConSurf, ortholog MSA entropy) is shown, so "conserved" is potentially a mislabeling of "ESM-confident." This needs either (a) a validating correlation against MSA-based conservation for the three targets, or (b) rewording throughout to "ESM-predicted structural confidence" rather than "conservation." 6. **The mhcnuggets vs. netMHCIIpan-4.1 "cross-validation" is not actually described as a validation.** The Methods say predictions were made "with mhcnuggets, cross-validated against IEDB netMHCIIpan-4.1" but no concordance statistic (rank correlation, AUC against IEDB-annotated true epitopes, or agreement rate on the 31,512 predictions) is reported anywhere in Results or Methods. This is the single most important methodological check for the entire antigen-directed arm (Fig. 7, the preclinical roadmap gating decision) and is currently a claim without a number. Please report: (a) the overlap/correlation between the two predictors on the same peptide-allele set, (b) whether burden rankings (Fig. 7a) are robust to which predictor is used, and (c) which predictor's calls fed the index-case ranking in Fig. 7c. 7. **Per-allele vs. summed epitope burden is conflated.** "31,512 predictions" across 10 allergens × 13 HLA-DRB1 alleles are apparently summed into an aggregate "epitope burden" per allergen (Fig. 7a) and separately into "presentation capacity" per allele (Fig. 7b), but the index-case ranking (Fig. 7c) requires a patient-specific two-allele (or four-allele, given both DRB1 and DQ) burden — the paper does not specify whether burden was summed across all 13 alleles (irrelevant to a given patient), restricted to the patient's own alleles, or some normalized score, and whether DQ2.2/DQ6.2 presentation (explicitly invoked for the wheat false-positive) was modeled at all despite the Methods only mentioning "HLA-DRB1" prediction. If DQ-restricted presentation for wheat gliadin was estimated by a different tool or heuristic, that needs to be stated — right now the wheat "discordance" mechanism (DQ2.2 presentation) is asserted without describing how DQ-allele binding was actually predicted, given the stated pipeline only models DRB1. 8. **The "≥92% / 100% addressable severe-endotype" diagnostic coverage figures need denominators and endotype-assignment method disclosed.** How many patients constitute the "severe endotype," across how many cohorts, and how was endotype assignment cross-validated against the 3-endotype clustering in Fig. 4b? Coverage statistics computed on the same patients used to derive the endotypes are circular unless out-of-sample checked. 9. **GWAS overlap (7 genes, e.g., CAPN14 OR 1.77, DSG1 OR 2.22) needs a background/enrichment test, not just a list.** Is 7/567 more than expected by chance given the GWAS trait's total credible gene set size and the signature's size? An enrichment test (hypergeometric/permutation against matched gene-set sizes) is standard and currently absent. **Minor concerns** 1. Fig. 2c's 76% secreted/surface figure among "up-regulated druggable genes" — clarify the druggability annotation source/version (Open Targets tractability bucket definitions change between releases) and whether "druggable" here means antibody-tractable specifically or includes small-molecule tractability too. 2. CCL26 log2FC of 4.56 "up in 9/9 cohorts" — report the effect-size heterogeneity (τ²/I²) for this gene specifically, since it's the headline number. 3. SIGLEC6 IHC companion diagnostic P=3×10⁻³¹ (84.5% vs 34.2%) — with what sample size? A p-value this extreme for a proportion comparison implies either a very large n or a potential test-choice issue (e.g., chi-square on pooled cells rather than patients as the unit of analysis, which would inflate significance via pseudoreplication). Please confirm the unit of analysis is patient-level, not cell-level. 4. Table 1's "prioritization score" — the formula/weights combining fold-change, accessibility, refractory persistence, and addressable fraction into a single score should be given explicitly (and ideally a sensitivity analysis showing lead selection is robust to reasonable reweighting). 5. Reference 1 (Rothenberg 2015 Gastroenterology) is cited generically as "Foundational EoE pathogenesis (see also refs 2-4)" — this is not a proper citation practice for a research article; give the actual claims it supports. 6. It would help to report the actual sample size for the "one paired pre/post-treatment cohort" underlying Fig. 4d directly in the Results text, not only implicitly via caveat in the Discussion. **Feasible revisions** 1. Add forest plots and I²/τ² heterogeneity statistics for the top signature genes (at minimum CCL26, POSTN, DSG1) plus a leave-one-cohort-out sensitivity analysis for the 567-gene list. 2. Report concordance (rank correlation or AUC vs. IEDB reference epitopes) between mhcnuggets and netMHCIIpan-4.1 on the shared prediction set, and state which tool's output is actually used downstream. 3. Clarify and report the exact per-patient burden computation (which alleles, DRB1 vs. DQ, summed vs. max) used for Fig. 7c, and explicitly state how DQ2.2/DQ6.2 binding (wheat gliadin) was predicted given Methods only describe DRB1 modeling. 4. Add a validation of ESM per-residue scores against an orthology/MSA-based conservation metric for the three lead targets, or rename the metric throughout to avoid implying evolutionary conservation. 5. Report sample size, unit of analysis (patient vs. cell), and exact test for the SIGLEC6 IHC diagnostic statistic and the refractory-core Fig. 4d result; run a permutation or exact test appropriate to the true n. 6. Add a hypergeometric/permutation enrichment test for the GWAS-gene overlap against a matched background gene-set size. 7. Provide clustering-stability diagnostics (resolution sensitivity, bootstrapped ARI) for the single-cell cell-type and mast-substate calls, and specify the batch-integration method used across single-cell samples/cohorts. 8. State explicitly whether endotype/severity assignment and diagnostic-addressability calculations used independent held-out folds or the same data, and if not, add a cross-validated addressability estimate. **Recommendation:** Major revision — the biology is plausible and largely literature-concordant, and the caveats that are disclosed are handled honestly, but several load-bearing statistical claims (refractory core, ESM-as-conservation, mhcnuggets/netMHCIIpan concordance, GWAS enrichment, diagnostic p-values) currently lack the methodological detail or validation needed to support the confidence with which they're presented, which matters more for a research article than for the companion Perspective. --- ## Cross-cutting comment The pair is well-constructed as a division of labor — Paper D's honesty about *process* (especially the HLA-equity catch) is a genuine contribution regardless of the underlying biology's ultimate correctness — but Paper D's trust argument leans on Paper E's numbers being solid, and several of those numbers (the single-paired-cohort refractory core, the unquantified mhcnuggets/netMHCIIpan agreement, ESM-as-conservation) are exactly the ones I'd flag as needing more scaffolding. I'd want the "calibration" story in Paper D to explicitly acknowledge that none of its 8 ledger episodes touch these core statistical/computational choices — the self-correction narrative is strongest on conceptual and ethical judgment calls and essentially untested on the quantitative pipeline that produces the headline targets. Neither paper should publish independently of the other addressing these gaps, since Paper D's central claim (the campaign is trustworthy) is only as strong as Paper E's unaudited statistical core. ## Reviewer 3 ### Reviewer 3: AI-in-science methodologist and research-ethics reviewer (agentic-science evaluation, reproducibility, AI-disclosure norms, trial equity) **Manuscripts assessed:** Paper D and Paper E --- ## PAPER D — "From disease name to design specification: a self-correcting AI-agent campaign for EoE" **Brief summary:** This Perspective narrates the operating model of a human-supervised AI agent that ran an EoE discovery-to-design campaign, using an 8-episode "correction ledger" and a 12-claim novelty classification as evidence that the campaign was calibrated and self-correcting. It argues for five generalizable practices for agentic science (orthogonal confirmation of novelty, auditable corrections, defined human judgment roles, calibration against withheld ground truth, provenance tracking). **Strengths** - The explicit taxonomy of correction types (machine self-correction vs. human domain catch vs. contained contradiction vs. carried-forward uncertainty vs. validated novelty) is a genuinely useful contribution to the "how do we evaluate agentic science" literature — more granular than most disclosure statements in the genre. - The S100A8/9 → S100A4 reversal (an orthogonal cross-cohort check catching the agent's own error) is a concrete, verifiable instance of self-correction rather than an assertion of it. - Foregrounding the rejected HLA-burden recruitment design as the paper's central ethical exhibit is the right instinct — it is a real, specific, non-trivial equity catch, not a generic "we care about fairness" gesture. - Reasonably direct acknowledgment that "none has been validated at the bench" and that GPU-heavy design steps were deferred as specifications rather than run. **Major concerns** 1. **The ledger is a curated sample of unknown denominator, and the paper never says so.** Eight episodes and twelve "principal findings" are presented as if they characterize the campaign's error rate, but the paper gives no accounting of how many claims, sub-analyses, or intermediate decisions were made in total, nor how the twelve "principal findings" or eight "episodes" were selected from that larger pool. Without a denominator, "8/12 grounded, 3 novel, 1 contradictory" is a curated highlight reel, not a calibration statistic. A genuinely evidentiary ledger would need either (a) an exhaustive accounting of all claims made during the campaign, or (b) an explicit, pre-specified sampling rule for which claims were entered — otherwise this is closer to a highlight reel selected by the same agent (and supervisor) whose trustworthiness it is meant to establish. 2. **Asymmetric epistemic weight between "agent caught it" and "human caught it" is asserted, not measured.** The paper's central rhetorical move — "the agent's contribution was high-volume and mechanical; the human's was sparse and decisive" — is plausible but is supported by exactly three human-catch anecdotes. Three data points do not establish a general claim about where the agent's failure modes systematically lie. The paper should either soften "the agent's failure modes are specific and predictable" to "in this campaign, three failure modes recurred," or provide a broader audit (e.g., how many decisions were reviewed and passed without correction, so the catch rate can be contextualized). 3. **The "self-correcting" framing in the title oversells what one clean example supports.** Self-correction is demonstrated once, cleanly (S100A8/9). Everything else in the ledger is either human-caught, contained-not-reversed, or carried forward as unresolved uncertainty. Calling the whole campaign "self-correcting" in the title, when the paper's own analysis shows the decisive catches were human, is a mismatch between title-level claim and body-level evidence. Consider retitling or explicitly scoping "self-correcting" to the one instance it fits, with the human oversight loop named as the actual source of reliability. 4. **The recruitment-ethics episode, while substantive, is described only from the outcome side.** We're told the agent proposed HLA-burden-stratified enrollment and the supervisor rejected it — but not what prompted the agent to generate that design in the first place, whether the agent was ever asked to consider equity as a constraint before proposing it, or whether the agent, once corrected, could recognize the class of error (e.g., did it flag similar equity risk unprompted anywhere else in the campaign?). Without that, this reads as one good catch by the human, not evidence about the agent's capacity (or incapacity) to reason about equity, which is the stronger and more useful claim the paper gestures at. 5. **"Verified against CrossRef" is not the same as "supports the claim as cited."** The closing note that all 17 references were retrieved and CrossRef-verified only confirms the references exist and resolve — it says nothing about whether the agent (or the humans) checked that each citation supports the specific quantitative or mechanistic claim attached to it. Given the paper's own thesis is about catching agent overclaiming, this is a place the authors should apply their own standard and show it, not just assert it. 6. **Table 1's "23 hours of active compute" and "214 artifacts" are presented as quality signals without a baseline.** Speed and volume are not evidence of correctness; the paper should be explicit that these are efficiency metrics, not validity metrics, and should avoid the implicit framing that fast + voluminous + occasionally self-corrected = trustworthy. A slower or smaller campaign with the same error-catch rate would be equally (in)valid. **Minor concerns** 1. Figure/table cross-referencing to Paper E is heavy (many identical numbers restated); the Perspective would benefit from being clearer about exactly which claims are original to this paper versus imported from the companion research article, since a reader of D alone cannot distinguish "reported here" from "reported elsewhere and repeated here." 2. The phrase "the causal role of dietary antigen is established by the efficacy of elimination diets" (also in Paper E) is a stronger causal claim than the cited review (ref. 10) generally supports for all patients/subtypes; efficacy of elimination diets is strong circumstantial evidence, not proof of a single causal mechanism, and the wording should be softened (e.g., "strongly implicated by"). 3. "Novelty" operational definition is mentioned as a "negotiated" decision between agent and human (a nice detail) but the actual operational definition used is never stated in the text — readers cannot audit whether the 3 "novel" claims meet a reasonable bar without knowing the criterion. 4. The competing-interests and author-contribution sections are placeholders; for a paper whose subject is trust and auditability, incomplete AI-use/competing-interest disclosure is a notable irony that should be flagged as a condition of publication, not an afterthought. **Feasible revisions** 1. Report the full denominator of claims/decisions made during the campaign (even approximately, e.g., "N discrete analytical claims were generated, of which these 12 were designated 'principal'") and state the selection rule for what entered the correction ledger and the 12-claim novelty set. 2. Reframe the title/abstract claim of "self-correcting" to accurately scope it — distinguish the one true machine self-correction from the human-supervised correction process as a whole, and rename the operating model accordingly (e.g., "human-in-the-loop corrigible" rather than "self-correcting"). 3. Add an explicit accounting of how many agent proposals or design decisions were reviewed and passed by the human supervisor without modification, to give the three human-catch anecdotes a denominator and let readers assess catch rate rather than catch existence. 4. State the operational definition of "novel" used to classify the twelve findings, and show the classification rubric (even a supplementary table) so the 8/3/1 split is auditable rather than asserted. 5. For each of the 17 citations, add a supplementary note (can be done editorially, no new experiments) confirming that the specific claim attached to each citation is supported by that source's actual findings, not merely that the citation resolves via CrossRef. 6. Complete the AI-authorship and competing-interest statements before submission; given the paper's thesis, an incomplete disclosure here undercuts the argument being made. **Recommendation:** **Major revision.** The correction-ledger concept is a genuine and useful contribution to trust-in-agentic-science reporting, but the paper's central claims of "self-correction" and "calibration" are evidenced by a small, unquantified, seemingly hand-picked set of episodes rather than a systematic audit, and the title/abstract's framing outruns what the body actually demonstrates — both fixable through better accounting and more careful scoping rather than new data collection. --- ## PAPER E — "An integrated computational pipeline for eosinophilic esophagitis..." **Brief summary:** This research article reports the underlying science: a 567-gene meta-signature from 9 harmonized bulk cohorts validated against 166,420 single cells, three prioritized therapeutic targets (SIGLEC6, IL1RL1/ST2, CCL26) advanced to structure-grounded design specs with companion diagnostics, an MHC-II antigen-mapping arm across ten allergens, a blinded single-patient calibration test, and a four-phase preclinical roadmap gated on ex-vivo T-cell validation. **Strengths** - Consistently and correctly labels all outputs as in-silico hypotheses; the Discussion's limitations paragraph explicitly flags the single-cohort refractory finding, the single-patient calibration, and the unassessed freedom-to-operate issue — this is honest work for the venue. - The HLA-agnostic trial design rationale is stated crisply and appears in the Results (not buried), which is good practice for a paper making an equity claim. - The blinded index-case test (wheat as an instructive false positive due to presentation-without-pathology) is a legitimately clever and well-reported piece of self-critical validation — it strengthens rather than just decorates the antigen-directed arm. - Explicit "not clinical guidance" caveat on the index-case section is present and appropriately blunt. **Major concerns** 1. **N=1 clinical calibration is used to bear far more argumentative weight than a single case can support.** The blinded index-patient test is repeatedly invoked (abstract, Results, Discussion) as validating the "presentation is a prior, not proof" principle and as supporting the antigen-directed arm generally. A single patient, with only 7 candidate foods ranked, is a compelling illustration but statistically underpowered to "establish" anything (the paper's own word choice in the Discussion — "together establish" — is too strong). This should be reframed uniformly as a single illustrative case, and the paper should report how many foods/negative controls were actually ranked, what the ranking method's expected false-positive rate would be by chance with n=7-10 items, and whether any pre-registration of the ranking (before unblinding) exists as an artifact. 2. **The treatment-refractory "core" finding — arguably the most clinically important result — rests on one paired cohort and is stated with more confidence than the caveat allows.** "Nine of the ten prioritized targets... normalized in responders but persisted in non-responders" is a striking number, but Results states it plainly before the caveat appears; the caveat ("we carried it forward as a hypothesis rather than an established result") should be attached at first mention, not follow the strong claim by a paragraph, and ideally the abstract itself should not present this as a settled finding equivalent to the meta-signature. 3. **Multiple-comparisons handling is unclear for several headline statistics.** The mast-cell expansion P=7.2×10⁻⁵ and the GWAS overlap odds ratios (CAPN14 1.77, DSG1 2.22) are reported without stating whether these were corrected for the number of cell types/genes tested, or whether the GWAS enrichment itself was tested against a background/null (e.g., permutation of random gene sets of the same size) rather than reported as raw per-gene odds ratios. Seven of 567 genes overlapping GWAS loci needs a background enrichment test (hypergeometric or permutation) against the genome-wide GWAS catalog, not just individual odds ratios, to support "concentrated in susceptibility... genes" as a signature-level claim. 4. **91/120 (76%) "druggable and accessible" framing risks conflating annotation categories with actual tractability.** Open Targets druggability annotation flags known families/domains; "secreted or surface-accessible" doesn't by itself establish an actionable epitope, developability, or lack of on-target liabilities (e.g., expression in other tissues). This is a coarse funnel step, and the manuscript should state the false-positive rate one would expect from this filter alone (i.e., how many antibody-drug programs against "secreted/surface accessible, druggable-annotated" targets fail later for off-target expression) to calibrate reader expectations of what "76% tractable" actually promises. 5. **Companion diagnostic addressability (≥92%, up to 100% for two of three) is a strong, headline-grade claim that needs its denominator and validation cohort specified.** Is "addressability" computed on the same cohorts used to derive the signature (circularity risk), or an independent set? If it's circular (derived and evaluated on overlapping single-cell/bulk data), the 92-100% figures are optimistic upper bounds, not external estimates, and this should be stated plainly. 6. **Competing interests / freedom-to-operate note is a one-line disclosure with no actual analysis.** "Freedom-to-operate has not been formally assessed" is honest, but given the paper explicitly builds an antigen-directed arm whose "specificity-defining measurement" is explicitly modeled on a very recently disclosed, functionally-validated prior-art finding (ref. 8, 2025), the paper should say more concretely what aspects (assay design? epitope target? diagnostic concept?) are potentially covered, even in a qualitative, non-legal sense — "has not been formally assessed" currently reads as a placeholder rather than engagement with a live IP question the authors themselves raise. **Minor concerns** 1. Figure 7's "31,512 predictions" across 10 allergens × 13 alleles is a large number that could hide substantial redundancy (overlapping peptides, shared HLA supertypes); a brief note on how many unique 15-mers/binding cores this represents would help readers gauge real epitope diversity versus prediction-count inflation. 2. mhcnuggets vs. netMHCIIpan-4.1 cross-validation is mentioned but no concordance statistic (e.g., rank correlation, top-k overlap) is given — this is an easy in-silico addition that would substantiate "cross-validated" beyond naming two tools. 3. The abstract states dupilumab "does not target the esophageal mast-cell compartment that expands most dramatically in disease" — worth citing evidence (or softening to "is not designed to directly target") since IL-4Rα blockade's downstream effects on mast-cell biology aren't fully characterized either way. 4. "Repurposing an established cardiac-biomarker platform" for soluble-ST2 ELISA is a nice practical note but should mention the very different expected serum ranges/dynamic range considerations between cardiac and esophageal-inflammatory contexts, since assay repurposing across indications is not automatically valid. 5. Table 1's Tier-A scoring formula (how fold-change, accessibility, refractory class, and addressable fraction are combined into "prioritization score") is not given in the Methods; this should be specified precisely enough to be reproducible. **Feasible revisions** 1. Reframe every mention of the index-case calibration (abstract, Results, Discussion) as "a single illustrative case" rather than language implying general validation ("together establish"), and report the chance-level expectation for top-2 ranking of 2 true triggers among ~7-10 candidates. 2. Move the "hypothesis, not established result" caveat on the PPI-refractory core to the same sentence/paragraph as the headline 9/10 statistic, and reflect the caveat's weight in the abstract's phrasing, not just the Discussion. 3. Add a formal enrichment test (permutation or hypergeometric, against a matched background gene set/GWAS catalog) for the GWAS risk-loci overlap, rather than reporting isolated per-gene odds ratios as if they demonstrate signature-level enrichment. 4. State explicitly whether companion-diagnostic addressability percentages were computed on data independent of signature derivation; if not, relabel them as "upper-bound estimates on the derivation cohort" and note the need for external validation. 5. Report concordance between mhcnuggets and netMHCIIpan-4.1 predictions (e.g., Spearman correlation of predicted affinities, or top-decile overlap) as a supplementary figure/table. 6. Expand the freedom-to-operate paragraph to identify, qualitatively, which specific components (functional assay design, target epitope, diagnostic readout) potentially overlap with the cited 2025 disclosure, even absent a formal legal opinion. 7. Provide the exact scoring formula/weights used for Tier-A prioritization in Methods, sufficient for independent recomputation from the released artifacts. **Recommendation:** **Major revision.** The pipeline is broad, mostly appropriately caveated, and the biology is largely concordant with known EoE literature, but several headline numbers (refractory core, diagnostic addressability, GWAS enrichment) are presented with more confidence in the primary text than the underlying single-cohort or potentially circular evidence supports; these are statistical/presentational fixes achievable without new data. --- **Cross-cutting comment:** The two-paper structure is itself a reasonable innovation — separating "what we found" from "how we ran the process" lets each be evaluated on its own terms — but it also creates a risk of mutual reinforcement: Paper D cites the ledger as evidence its findings are trustworthy, while Paper E cites the same episodes as validating its biology, and a reader of either alone may over-credit claims that are only as strong as the shared, unquantified evidentiary base underneath both. The single strongest piece of genuine external validation across the pair — the blinded index-case wheat/milk/soy result — is real and well-designed, but it is n=1 and is asked to carry rhetorical weight in both manuscripts that a single patient cannot bear; tightening that claim uniformly across D and E would materially improve the pair's credibility.