# Project Tolera — Reconciled Accomplishments Ledger
_Single source of truth for the website, deck, talking points, and prize submission. Every number traces to a named source; computed values were recomputed from the source artifact, not recalled._

**Built with Claude: Life Sciences Hackathon (2026) · Ruth-Anne Pai, PhD — immunologist and person living with EoE**

---

## Headline (use these consistently everywhere)
- **1 week**, **1 person** (a PhD immunologist and person living with EoE who does not write code), working in Claude Science.
- **3 integrated work streams → 3 preprint-style manuscripts**, spanning **2 diseases** (eosinophilic esophagitis + esophageal adenocarcinoma).
- **10 open-source GitHub repositories** + **1 reusable Specialist Agent** (the Therapeutic Program Architect, a 10-stage pipeline).
- **A public website** (6 pages, 6 audiences) at **https://ruthannepai.netlify.app**.

> ⚠️ **Two numbers changed since the current site/deck/submission were written and must be updated everywhere:**
> **manuscripts: 2 → 3** (EAC/Barrett's is the new third), and **repositories: 8 → 10** (verified live against the GitHub org on the day of this ledger).

---

## Stream 1 — Personalized pMHC-II tolerance (the flagship EoE pivot)
*A patient-specific antigen-selection engine for personalized pMHC-II tolerance therapy in EoE.*
- **3 dominant EoE allergens** (milk β-casein, wheat α/β-gliadin, soy β-conglycinin) scanned across **5 common HLA-DRB1 alleles**.
- **4,640 peptide × allele evaluations → 832 strong binders** (IC₅₀ < 500 nM).
- Presented-epitope load is strongly HLA-dependent — **13 to 352 strong binders per allele (27-fold range)**: the evidence that antigen priority is *per-patient and must be computed*.
- Worked example: a **milk β-casein epitope on HLA-DRB1\*07:01** clears an **ipTM ≥ 0.87** structural gate with full in-groove engagement.
- Delivery: **20 nm iron-oxide nanoparticle, 5 pMHC copies, ~24× avidity**; construct concept is **PACT** — one shared MHC-II backbone with a swappable peptide cassette.

## Stream 2 — CCL26 / POSTN effector-target binders (omics-mined)
*De novo design of neutralizing protein binders against CCL26 and POSTN.*
- **9-cohort EoE transcriptomic meta-analysis** nominates **2 targets** with no direct therapeutic: **CCL26/eotaxin-3** (most discriminating EoE transcript) and **POSTN/periostin** (persists in histologic remission).
- Design campaign: **80 RFdiffusion backbones → 1,920 SolubleMPNN sequences → 60 binder–target complexes** assessed by Boltz-2.
- Lead binders: **CCL26 ipTM 0.942** (4/5 epitope hotspots, 1,220 Ų buried); **POSTN ipTM 0.901** (pLDDT 0.911, 3/4 hotspots, 842 Ų).
- **Honesty result (a feature, not a footnote):** a **36-complex scrambled/random decoy calibration** showed **92–100% of decoys cross the same ipTM/pLDDT gate** that 59/60 designs cross — so a high pass rate is close to the null. Only focal on-epitope engagement separates designs from decoys. This is the self-correction story worth telling.

## Stream 3 — EAC / Barrett's dual-arm program (NEW; proves the platform generalizes)
*Steering into competitive whitespace: GUCY2C and DKK1 across the esophageal-adenocarcinoma trajectory.*
- Genomics from **TCGA/cBioPortal (esca, n=182)**: **TP53 87%, CDKN2A del 39%, CCND1 amp 35%** — a copy-number-driven cancer.
- Competitive map: **351 interventional EAC trials (106 active)**; **16 candidate targets triaged** → **2 nominated**: **GUCY2C** (T-cell-engager treatment arm) and **DKK1** (neutralizing-trap interception arm in high-risk Barrett's).
- De novo binders: **16 designs (8 backbones/target), 15/16 clear ipTM > 0.5**. Top **GUCY2C gucy2c_bb2 ipTM 0.915** (flagged: high-alanine developability liability); top **DKK1 dkk1_bb1 ipTM 0.858, pLDDT 0.908** — the cleanest design in the set.
- Program verdict: **TRACTABLE (0.87)**; two make-or-break questions (GUCY2C normal-gut window; DKK1 dysplasia-regression surrogate endpoint) are named as open, not claimed.

## Cross-stream design totals (state carefully)
- **76 binder–target complexes computationally assessed** across streams 2 & 3 (60 + 16) via two distinct pipelines. Lead-binder ipTM spans **0.858–0.942**.
- Reporting rule: these are **pilot-scale, in-silico** leads. ipTM is a single-model docking-confidence proxy, **not a measured affinity**. No experimental validation has been performed on any design.

## Software & platform
- **10 public GitHub repositories** (github.com/ruthannepai-tech): antigen-epitope-pipeline, claude_science_videos, eoe-protein-design-skills, inflammatory-target-discovery-skills, manuscript-architect-agent, omics-target-mining, patient-centered-market-and-survey, podcast-debrief-skill, synthetic-peer-review, therapeutic-program-architect.
- **1 Specialist Agent** — the **Therapeutic Program Architect**: any disease → a **10-stage**, human-in-the-loop, gate-stopping drug-program dossier. (A Manuscript Architect agent also exists.)

## Media & outreach
- **6-page website**, 6 audiences, walkthrough video embedded.
- **4 videos** (site walkthrough, T-cell three-fate animation, CCL26 & POSTN 360° binder renders — ChimeraX/Blender on Modal GPUs).
- **1 podcast/audio debrief**; **3 executive summaries** and **3 plain-language summaries** (one per stream).

## Process & narrative (the story the numbers support)
- **Author:** PhD immunologist, **person living with EoE**, who had never written Python or used a protein model.
- **Method, not one drug:** every step is an open, reusable skill; the whole workflow is packaged as an installable specialist that runs for any antigen-specific disease.
- **Honesty-gated science:** Claude Science repeatedly flagged and corrected over-claims (e.g., calling structure prediction "validation"; the decoy-calibration result). **2 synthetic peer-review rounds** (Major → Minor) hardened the manuscripts.
- **Democratization:** the through-line — patient-driven, patient-centered drug development now within reach of the community it serves.

---
### Provenance notes / watch-items
- "9 constructs / three formats" is **older pMHC-nanoparticle framing**; the *current* flagship pMHC manuscript is the personalized antigen-selection engine (832 binders, PACT cassette). Prefer the engine framing on all live deliverables; keep "9 constructs" only if explicitly referencing the earlier nanoparticle panel.
- Artifact `S1_binder_designs.csv` in the store actually contains **EAC (DKK1/GUCY2C)** data (naming collision) — CCL26/POSTN design counts come from the manuscript, which is authoritative.
- Live-URL note: submission links **ruthannepai.netlify.app**; site footer also shows **ruthannepai.com** — treat netlify as canonical live site; confirm primary domain.