# Paper D — Outline & Draft Plan
## An AI agent runs an end-to-end therapeutic target-discovery and protein-design campaign on public data: a week-long eosinophilic esophagitis case study

**Working title options**
1. *From disease name to design spec in a week: an autonomous-agent campaign for eosinophilic esophagitis therapeutic discovery*
2. *Human-supervised AI science: an agentic pipeline from public omics to protein-therapeutic design*
3. *A self-correcting AI research agent for drug-target discovery: the eosinophilic esophagitis case*

**Article type:** Methods / Perspective hybrid with a worked case study (a "systems demonstration" paper).
**Primary audience:** computational biologists, AI-for-science community, translational drug-discovery groups, and the *Built with Claude: Life Sciences* hackathon audience.
**Central thesis:** A general-purpose AI agent, supervised by a domain scientist, can execute a *coherent, reproducible, and self-correcting* research campaign — spanning omics mining, mechanistic validation, target prioritization, and computational protein design — end-to-end on public data in a compressed timeframe. The scientific value lies not only in the EoE outputs but in the **workflow, the human–AI division of labor, and the calibration/self-correction behavior** that make the outputs trustworthy.

---

## Why this paper (and how it differs from Papers A/B/C)

- **Paper A** = the reusable pipeline (method as product).
- **Paper B** = the EoE biology (disease-discovery science).
- **Paper C** = the antigen-directed pMHC framework.
- **Paper D (this)** = the *process*: how an agent + scientist actually produced A/B/C's content, what the collaboration looked like, where the agent self-corrected, where the human caught errors, and what this implies for AI-assisted drug discovery. EoE is the running example, not the point.

The novel, publishable claim is about **agentic scientific practice**, evidenced by a real campaign — including its failures and corrections, which most AI-for-science papers omit.

---

## Manuscript structure

### 1. Introduction
- Drug-target discovery is bottlenecked by the breadth of public data and the number of manual analytical steps between "disease" and "actionable target/design." (Motivate with EoE: a chronic, food-antigen-driven type-2 disease with only one approved biologic, dupilumab, and a large treatment-refractory population.)
- AI agents promise end-to-end execution, but the field lacks candid, reproducible accounts of an agent doing *the whole thing* — with honest treatment of error and correction.
- Contribution: (i) a documented end-to-end agentic campaign; (ii) an explicit human–AI supervision model; (iii) empirical evidence of self-correction and calibrated uncertainty; (iv) a fully artifact-tracked, reproducible output set.

### 2. The campaign at a glance (Figure 1 — hero schematic)
- Four phases: **Discovery → Validation → Design → Prospective framework.**
- One master schematic: disease name → 57 public datasets → 567-gene meta-signature → single-cell + endotype validation → Tier-A targets → structure-guided design specs + pMHC framework. Overlay where the human intervened and where the agent self-corrected.
- Headline metrics box: 57 datasets surveyed; 9 harmonized bulk cohorts (235 samples); 166,420 single cells; 3 protein-design leads; ~1 week wall-clock; N artifacts under lineage tracking.

### 3. The human–AI operating model (Figure 2 — swimlane / RACI diagram)
- **Agent responsibilities:** dataset retrieval + QC, statistics (moderated-t, DerSimonian–Laird meta-analysis), single-cell processing, enrichment, druggability annotation, structure retrieval, ESM-2 analysis, drafting.
- **Human (scientist) responsibilities:** framing goals, domain priors, sanity-checking biology, catching conceptual errors, deciding scope, approving plans.
- **Shared/negotiated:** target prioritization criteria, modality selection, what counts as "novel."
- Present as a swimlane figure across the four phases; annotate the specific hand-off points.

### 4. Case study results (compressed — this is NOT Paper B; report headlines + point to companion papers)
- **4.1 Discovery:** 567-gene reproducible meta-signature; IFN + IL-6/JAK-STAT + NF-κB program over barrier loss; 13-fold mast expansion; Tier-A targets (CCL26, IL1RL1/ST2, SIGLEC6, POSTN, ALOX15, …).
- **4.2 Validation:** epithelial differentiation arrest; type-II/IFN-γ resolution; reproducible severity endotypes (6/6 cohorts); PPI-refractory persistence; GWAS risk-gene-vs-effector split.
- **4.3 Design:** three protein-tractable leads with interface maps, ESM-2 epitope conservation, developability, companion-Dx pairing.
- **4.4 Prospective framework:** food-antigen → MHC-II/CD4 tolerance rationale, multi-allergen epitope mapping, HLA-based stratification.
- Keep this section tight: figures are reused/cited from A/B/C; here they are *evidence the agent produced valid science*, not the primary contribution.

### 5. Self-correction and calibration — the core evidence (Figure 3 — the "correction ledger")
This is the section that distinguishes the paper. Concrete, documented episodes (sourced from the project phase-report artifacts: S100A8/9→S100A4 and FLG from `eoe_phase2_report.md`; MHC-I→II from `food_antigen_pmhc_review.md` §2. These are recorded in the report prose, not recoverable via a session-transcript grep — verify each against the cited artifact when drafting):
- **The S100A8/9 self-correction.** Agent initially nominated an S100A8/9 alarmin axis from cell-cell communication; independent bulk-cohort check showed the ligands were *down* (tracking differentiation loss) and the true mast-remodeling signal was **S100A4**. Agent reversed its own claim when the data contradicted it. *(Machine self-correction forced by orthogonal evidence.)*
- **The MHC-I vs MHC-II catch.** Agent's first pMHC analysis scored MHC-I/CD8 peptides; the human pointed out that EoE is a CD4/Th2 food-antigen disease, so the disease-central arm is **MHC-II**. Agent integrated the correction and re-scoped the antigen-directed strategy to class-II-led. *(Human domain-prior catch.)*
- **The FLG contradiction, kept not hidden.** FLG was down in 8/9 cohorts but *up* in one recurrence cohort; the agent recorded it as a cohort-specific contradiction rather than suppressing it.
- **Honest limitation flagging:** eosinophil dropout in droplet scRNA-seq; PPI-refractory finding resting on a single paired cohort; "surface" druggability labels needing manual triage; GPU-dependent design steps deferred and delivered as executable specs.
- Present as a **correction ledger table/figure**: each episode → trigger → who caught it → resolution → effect on conclusions. Tie to the existing `novelty_ledger.csv`.

### 6. Reproducibility and provenance (Figure 4 — artifact lineage graph)
- Every result is an artifact with tracked lineage (code + inputs + environment).
- The pipeline itself was packaged as a reusable skill (`omics-target-mining`: `SKILL.md` + `kernel.py`) — the campaign produced a *tool*, not just answers.
- Show a lineage DAG for one representative target (e.g., SIGLEC6: meta-signature → single-cell mast state → independent confirmation → design brief) to make provenance concrete.

### 7. What worked, what didn't, and limits of agentic science (Discussion)
- **Worked:** breadth of data integration at speed; statistical rigor when methods were specified; honest uncertainty; end-to-end continuity across four phases.
- **Didn't / needed human:** conceptual framing (MHC-I vs II), biological plausibility judgments, deciding when a signal is disease-specific vs artifact.
- **Hard limits:** no wet-bench validation; GPU-gated design steps; all findings in silico; agent can be confidently wrong without an orthogonal check → argues for *designed-in* verification steps.
- **Generalizable practices:** require orthogonal confirmation for any novel claim; keep a correction ledger; make the human a supervisor not a spectator; track provenance for everything.

### 8. Conclusion
- A supervised agent compressed a multi-month discovery-to-design arc into a week and produced calibrated, reproducible, self-corrected outputs. The methodological contribution is a *template for trustworthy agentic science*, with EoE as proof of concept.

---

## Figure plan (mapped to existing artifacts)

| Fig | Content | Built from existing artifacts | New work needed |
|-----|---------|-------------------------------|-----------------|
| **1** | Campaign schematic (4 phases, data→design) | phase summary PNGs, dataset inventory | New composite schematic (compose) |
| **2** | Human–AI swimlane / operating model | — | New diagram |
| **3** | Self-correction & calibration ledger | `novelty_ledger.csv`, `novelty_ledger.png`, `independent_confirmation.png` | New ledger figure + table |
| **4** | Artifact lineage DAG (SIGLEC6 exemplar) | lineage graph of design brief | New lineage render |
| **S1–Sx** | Case-study evidence (reused/cited from A/B/C) | meta-analysis, endotypes, ESM epitope, interface, pMHC panels | Mostly reuse |

## Tables
- **T1** Campaign metrics (datasets, samples, cells, targets, artifacts, wall-clock).
- **T2** Correction ledger (episode → trigger → catcher → resolution → impact).
- **T3** Human vs agent task allocation across phases.

---

## Gaps to fill before submission
1. **Quantify the campaign:** exact artifact count, total datasets/samples/cells, wall-clock per phase, number of human interventions vs autonomous steps. (Computable from artifact store + session record.)
2. **Formalize the correction ledger** from `novelty_ledger.csv` + the documented episodes into T2/Fig 3.
3. **Build the two new schematics** (Fig 1 campaign, Fig 2 operating model) and one lineage render (Fig 4).
4. **Decide venue & framing** (see below) — perspective vs methods changes tone and length.
5. **Ethics/limitations statement:** in silico only; not clinical advice; human oversight essential.
6. **Author-contribution / AI-disclosure statement** consistent with target-journal policy on AI-generated content — this paper should be exemplary here given its subject.

## Candidate venues
- *Nature Methods* / *Nature Biotechnology* (Perspective or Correspondence) — high bar, high fit for "AI runs a campaign."
- *Cell Reports Methods*, *Patterns* (Cell Press, data-science-for-biology) — strong fit for the reproducibility + workflow angle.
- *Bioinformatics* / *Bioinformatics Advances* — if framed methods-forward.
- *Drug Discovery Today* / *Nature Reviews Drug Discovery* (Perspective) — for the translational-process framing.
- The hackathon's own write-up / blog + a preprint (bioRxiv) — fastest path, pairs well with a peer-reviewed submission.

## Relationship to A/B/C (avoid self-overlap)
- D **cites** A (pipeline), B (EoE biology), C (pMHC framework) rather than re-deriving them.
- D's *owned* figures: campaign schematic, operating model, correction ledger, lineage graph — none of which belong to A/B/C.
- If A/B/C aren't yet submitted, D can cite them as "companion manuscripts (in preparation)" or fold minimal results inline.