# Project Tolera — 3-Minute Talking Points
*Built with Claude: Life Sciences Hackathon · Researcher Track · Ruth-Anne Pai, PhD*

**Format:** 12 slides in ~3:00. Target ~430 spoken words. Timings are cumulative; the **[SAY]** lines are the spoken script, **(note)** lines are delivery cues.

---

**Slide 1 — Title (0:00–0:15)**
[SAY] "I have EoE — eosinophilic esophagitis. I'm a PhD scientist and a rare-disease advocate, and I don't code. This week, with Claude Science, I designed a therapy for my own disease."
(note) Land the personal hook. Pause after "for my own disease."

**Slide 2 — Why we're here (0:15–0:25)**
[SAY] "EoE has effective treatments, but none is curative — they all reduce inflammation and none removes the cause. I wanted to go after the cause."

**Slide 3 — We began with the omics (0:25–0:45)**
[SAY] "I started where the brief pointed: public data. I mined bulk and single-cell EoE transcriptomes and ran a nine-cohort meta-analysis to surface the genes and cells that are actually dysregulated."
(note) Gesture to the volcano plot.

**Slide 4 — Differential expression → target (0:45–1:00)**
[SAY] "That pointed me to two convergent effector targets — the eosinophil chemokine CCL26 and the remodeling protein POSTN — and I designed de novo protein binders docked to both."

**Slide 5 — The pivot: target the cause (1:00–1:15)**
[SAY] "But effectors are downstream. The cause in EoE is a food-driven immune response — so I pivoted to antigen-specific tolerance: retrain the immune system to the trigger foods instead of blocking inflammation forever."

**Slide 6 — From target to designed protein (1:15–1:30)**
[SAY] "I'd never touched a protein model. With Claude I built pMHC-II tolerance constructs — screened 4,640 peptide–allele pairs, found 832 strong binders, and folded leads for dairy, wheat, and soy."

**Slide 7 — Learning by building (1:30–1:45)**
[SAY] "Everything was new. I created my first GitHub and Modal accounts, published my first reusable skills, and learned ChimeraX and Blender to render my own molecular videos — on Claude Science and Modal GPUs."

**Slide 8 — From week to roadmap (1:45–2:00)**
[SAY] "In one week this became three peer-review–style manuscripts with synthetic reviewer responses, a preclinical-to-IND roadmap, and a regulatory and market strategy — a credible line from a public dataset to the clinic."
(note) Emphasize **three** manuscripts.

**Slide 9 — One platform, many diseases (2:00–2:15)**
[SAY] "And the method isn't specific to EoE. Every step — mine the data, nominate targets, design and honestly gate binders — is a repeatable pipeline."

**Slide 10 — The reusable engine (2:15–2:40)**
[SAY] "So I packaged the whole thing as an installable Specialist Agent — point it at any disease and it runs a ten-stage program. To prove it, I re-ran the entire engine on a completely different disease — esophageal cancer — and it produced a dual-arm program of its own: a GUCY2C T-cell engager and a DKK1 trap, with de novo binder leads."
(note) This is the "it generalizes" beat — the strongest proof. Slow down here.

**Slide 11 — The front door (2:40–2:52)**
[SAY] "It's all open at ruthannepai.netlify.app — three manuscripts, ten repositories, two diseases, binder interfaces from 0.86 to 0.94 ipTM. Every result, skill, and structure is downloadable and reproducible."

**Slide 12 — Why this matters (2:52–3:00)**
[SAY] "Science, democratized. As a patient, I was empowered to design a therapy for my own disease — and if I can, a growing wave of AI-empowered researchers can too. Thank you — Anthropic, Gladstone Institutes, and Cerebral Valley."

---

## Backup Q&A one-liners
- **"Is any of this validated?"** — "No. Everything is computational — in-silico structure prediction, not experimental validation. I'm careful to call it that, and Claude flagged me every time I drifted toward overclaiming."
- **"Why should I trust the binders?"** — "Because I ran negative controls. For CCL26/POSTN, scrambled-sequence decoys cleared the same confidence gate 92–100% of the time — so the fold-back score alone isn't proof, and I say so in the manuscript."
- **"What's the honest headline?"** — "One person, one week: two diseases, three manuscripts, a reusable engine — and an honest account of what's a lead versus what's validated."
- **"What would you do with more time?"** — "Wet-lab the Tier-A dairy construct and the GUCY2C lead first — those are the two closest to a real experiment."

## Numbers cheat-sheet (matches deck + site + ledger)
- 3 manuscripts · 2 diseases (EoE + esophageal cancer) · 10 GitHub repos · 1 Specialist Agent (10-stage)
- pMHC-II: 4,640 peptide×allele evaluations → 832 strong binders (IC50<500 nM), 13–352/allele
- CCL26/POSTN: 9-cohort meta-analysis; 80 backbones → 1,920 sequences → 60 complexes; CCL26 lead ipTM 0.942, POSTN 0.901; 36 decoys clear the gate (the honesty result)
- EAC: TCGA n=182; 351 EAC trials (106 active); 16 designs, 15/16 pass ipTM>0.5; GUCY2C 0.915, DKK1 0.858; verdict TRACTABLE 0.87
- Overall lead ipTM range: 0.86–0.94