# Revision Plan — EoE de novo binder-design manuscript

**In response to:** Mock peer-review (Major Revision; both main + supplement).
**Scope discipline:** every action below is achievable in silico or by editing — no wet-lab, no new cohorts. Each item is tagged with the artifact/section it touches, the effort tier, and whether the data is already in hand or needs a new (in-silico) computation.

**Effort tiers:** **E** = edit/language only · **A** = re-tabulate or plot data already computed · **C** = new in-silico computation from existing inputs.

**Priority order** (editor's adjudication): do the quantitative-audit fixes (R2 controls, R3 statistics) first — they decide whether the two leads can be trusted at all — then the narrative-calibration fixes (R1), then polish.

---

## Tier 1 — Highest value: does the evidence support the two leads?

### RP-1. Decoy / negative-control calibration of the Boltz-2 thresholds  — tier C — addresses Essential #2 (R2 Major #1–2)
**Problem:** "59/60 passed ipTM>0.5 & pLDDT>0.7" has no null distribution, so the pass rate is uninterpretable.
**Action:** generate 20–30 scrambled/shuffled binder sequences (same length/composition, sequence identity destroyed) and, per target, ~10 length-matched random-decoy binders; co-fold each against the target with the *identical* Boltz-2 protocol (3 recycles, 3 diffusion samples, `--use_msa_server`). Report the decoy ipTM/pLDDT distribution beside the 60 designs.
**Deliverable:** new panel in Fig 4 (or Fig S9) — design vs decoy confidence; new decoy rows in Table S4. Restate pass criteria as "exceeds the decoy-calibrated confidence band."
**Inputs in hand:** target structures (1G2S ch A, 5WT7 ch A) and the design sequences (`eoe_binder_validation.csv`, `binder_seq` column). Needs GPU co-folding run.

### RP-2. Orthogonal-predictor agreement on the leads  — tier C — addresses Essential #8 (R2 Major #1, Feasible #2)
**Problem:** single co-folding model = self-consistency, not validation (circularity).
**Action:** re-fold the 2 leads (and ideally the top ~10 by on-target score) with an architecturally distinct predictor (Chai-1 or AlphaFold3-class). Report per-complex ipTM/pLDDT concordance vs Boltz-2.
**Deliverable:** small concordance table (design × predictor × ipTM/pLDDT) + one scatter panel; text sentence on agreement. Move this from "future work" into Results.
**Inputs in hand:** lead complex sequences/structures. Needs a second predictor run.

### RP-3. On-target composite score — justify or demote  — tier A — addresses Essential #2 (R2 Major #3)
**Problem:** the score `ipTM × (0.5 + 0.5·coverage) × min(pLDDT/0.7, 1)` has undefended constants and may just track ipTM.
**Action:** compute Spearman ρ between on-target score and raw ipTM across all 60 (from `eoe_binder_validation.csv`); show how many lead-selection decisions flip if you rank by ipTM alone vs the composite. State plainly whether the composite changes conclusions; if not, present it as a tie-breaker heuristic, not a validation metric, and give the rationale for the 0.5/0.5 split and the 0.7 pLDDT floor.
**Deliverable:** one correlation number + sentence in §2.4; footnote deriving the constants. **Data fully in hand.**

### RP-4. Reframe "59/60 passed" as a threshold rate on a pre-filtered subset  — tier E — addresses Essential #2 (R2 Major #4 + R3 Major #6, convergent)
**Action:** in Abstract and §2.4, state the funnel explicitly: 60 = 30/target hand-selected (by MPNN score + entropy) from 1,074 complexity-passing of 1,920 designed on 80 backbones. "59/60 pass confidence thresholds" is a threshold-crossing rate on that curated subset, **not** a design success rate.
**Deliverable:** two edited sentences; cross-reference the Fig 3A funnel. **Edit only.**

---

## Tier 2 — Statistical audit trail (R3)

### RP-5. Show the meta-analysis statistics, don't assert them  — tier A — addresses Essential #3 (R3 Major #1–3)
**Action:** add per-cohort log2FC/SE/direction and I² for CCL26 and POSTN to the text/Table S1. Report the total count of FDR-significant genes so the adjusted p-values have a denominator.
**Grounded values already in artifacts** (`eoe_target_shortlist.csv`, `eoe_crossref_lit_vs_omics.csv`):
- CCL26: pooled log2FC +4.56, adj p 3.4×10⁻⁴, **k=9 cohorts, up in 9/9, I²≈98%**
- POSTN: pooled log2FC +5.55, adj p 2.5×10⁻¹⁶, **k=8 cohorts, up in 8/8, I²≈72%**
**Note the high heterogeneity** (CCL26 I²≈98) explicitly — with random-effects it widens the CI; report the CI, and consider a leave-one-cohort-out sensitivity check (LOCO artifact `eoe_meta_signature_loco_robust.csv` already exists — cite it).
**Deliverable:** expanded Table S1 rows + one Results sentence per lead. **Data in hand.**

### RP-6. Resolve the "9 cohorts" vs "8/8 POSTN" discrepancy  — tier E — addresses Essential #3 (R3)
**Resolution (confirmed from data):** the meta-analysis spans 9 cohorts, but **POSTN was quantified in only 8** of them (k=8, up 8/8); CCL26 was in all 9 (k=9, up 9/9). This is differential per-gene cohort coverage, not an inconsistency.
**Action:** add a half-sentence — "POSTN passed expression filtering in 8 of the 9 cohorts (up in 8/8); CCL26 in all 9 (9/9)." **Edit only.**

### RP-7. Disclose the druggability-shortlist scoring formula  — tier A/E — addresses Essential #5 of supplement + #2 main (R1, R3 convergent)
**Problem:** the Boltz on-target formula is disclosed but Table S2's `priority_score` is not — asymmetric transparency. Score exists (`priority_score` column) but weights over effect size / concordance / single-cell specificity / AB-tractability are not stated.
**Action:** write out the formula and weights in Supplementary Methods (or state it was a qualitative composite and relabel as "priority tier"). **Edit + one methods paragraph.**

### RP-8. Disclose the literature/trial corpus search methodology  — tier E — addresses Essential #4 (R3 Major #5, Minor #1)
**Action:** add to Supplementary Methods the PubMed query string(s), inclusion criteria, and the date range — **fix the "2015–2026" range** (flag/correct the apparent 2026 typo, or justify if a forward-dated preprint search was intended). Describe how the 45-target Table S3 subset was drawn from the 591-abstract / 157-trial corpus. Makes "zero direct trials" falsifiable. **Edit only.**

---

## Tier 3 — Claim calibration & narrative (R1)

### RP-9. Calibrate "validated" throughout  — tier E — addresses Essential #1 (R1, R2, R3 convergent)
**Action:** replace unqualified "validated" with "computationally validated / in-silico assessed" at all three Abstract+Results occurrences (Abstract "we validated 60…"; §2.4 "We validated the 60…"; §2.5 header "Two validated lead binders") and in figure legends. Apply from the Abstract onward, not only the Discussion. **Edit only.**

### RP-10. Reconcile CCL26 "single most discriminating transcript" with its 11th/20 druggability rank  — tier E — addresses Essential #5 (R1 #6, R3 #4 convergent)
**Resolution (from data):** the two statements measure different things — discriminating power (DE effect/specificity vs GERD) vs a composite druggability priority in which POSTN ranks 3rd and CCL26 11th (CCL26's lower rank is driven by its very high heterogeneity I²≈98 and adj-p, not low effect). 
**Action:** add a sentence acknowledging the tension and explaining it; cite the shortlist formula (RP-7). **Edit only.**

### RP-11. Engage with *why* these targets are undrugged  — tier E — addresses Essential #4 (R1 Major #1–2)
**Action:** add a Discussion paragraph on (a) eotaxin redundancy — CCL11/CCL24 also signal through CCR3, so trapping CCL26 alone may be partially bypassed; (b) periostin's history as a biomarker rather than a validated target and its physiological tissue-repair role (agonist-vs-antagonist and selectivity risk — already partly noted, strengthen). Frames "undrugged" as opportunity *with* a mechanistic reason, not just a pipeline gap. **Edit only** (literature already in project review artifacts).

### RP-12. Temper POSTN clinical/barrier framing  — tier E — addresses Essential #6 (R1 Major #4–5)
**Action:** label the incremental benefit of direct POSTN neutralization over existing IL-13-pathway blockade (dupilumab/cendakimab) as an untested hypothesis; clarify that "histologic remission" is an eosinophil-count endpoint, not proof of ongoing symptomatic disease. **Edit only.**

### RP-13. State the isolated-domain caveat in Results §2.5, not only Discussion  — tier E — addresses Essential #7 (R2 Major #5)
**Action:** one sentence in §2.5 — validation used isolated FAS1-IV (PDB 5WT7), not full-length multi-domain POSTN; binder behavior in the intact protein/integrin context is untested. **Edit only.**

---

## Supplement-specific

### RP-14. Deposit full gene-level meta-analysis table  — tier A — (R1, R3 convergent)
Table S1 shows top-60; deposit the full 25,654-gene table (artifact `eoe_meta_signature.csv` already exists) or at minimum per-cohort stats for the two leads. **Data in hand — link the existing artifact.**

### RP-15. Extend Table S4 with buried SASA + interface counts for all 60; add decoy rows  — tier C — (R2)
**Status:** `eoe_binder_validation.csv` has `n_iface` for all 60 but **no buried-SASA column** (only the two leads were SASA-scored). 
**Action:** compute buried SASA for all 60 complexes from their Boltz-2 output structures (Shrake–Rupley, same method as leads); merge with decoy rows from RP-1. **Needs recompute from the 60 complex PDBs.**

### RP-16. Version-pin the pipeline  — tier E — (R2)
Record commit hashes/weights for RFdiffusion (`Complex_base_ckpt.pt`), ProteinMPNN (`v_48_020`), Boltz-2, and random seeds (or documented run-to-run variance). **Edit / metadata capture.**

---

## Summary: effort distribution

| Tier | Items | New in-silico compute | Data-in-hand re-tabulation | Edit-only |
|---|---|---|---|---|
| Compute (C) | RP-1, RP-2, RP-15 | ✔ (GPU co-folding + SASA) | — | — |
| Re-tabulate (A) | RP-3, RP-5, RP-7, RP-14 | — | ✔ | — |
| Edit (E) | RP-4, RP-6, RP-8, RP-9, RP-10, RP-11, RP-12, RP-13, RP-16 | — | — | ✔ |

**Critical path:** the three compute items (RP-1 decoys, RP-2 orthogonal predictor, RP-15 SASA-all-60) are the only ones needing GPU time and are the highest-value — they directly answer the circularity/controls objection that drove the Major decision. Everything else is disclosure, re-tabulation of data already computed, or language calibration.

*Every fix maps to a convergent (≥2-reviewer) or single-reviewer essential revision from the mock panel; no action requires wet-lab, clinical, or new-cohort data.*