# Project Tolera — Gladstone Prize Text Submission (revised) *All numbers reconciled with the accomplishments ledger, website, and deck. First-person voice preserved; edits are surgical.* --- ## Project Name Project Tolera: From One Disease to a Therapeutic Program Architect for Any Disease ## Project description Project Tolera is an open, AI-native pipeline for antigen-specific therapy design, built end-to-end in one week for eosinophilic esophagitis (EoE), a chronic food-driven disease I live with. Starting from public bulk and single-cell omics, I mined EoE datasets into a ranked, druggability-filtered target landscape (surfacing CCL26 and POSTN) and designed de novo protein binders against them. But I saw an opportunity to pivot: every approved and pipeline EoE drug blocks a downstream cytokine and is taken indefinitely, yet EoE is antigen-specific — a food peptide presented on MHC-II to a pathogenic Th2 T cell. Targeting that triad could induce durable tolerance instead of lifelong suppression. As proof of concept, I built a patient-specific antigen-selection engine: from 4,640 peptide–allele evaluations it nominated 832 strong-binding tolerizing epitopes, and I modeled lead pMHC-II complexes for dairy, wheat, and soy with high structural confidence. Every step is an open, reusable skill, and I packaged the whole workflow into a Specialist Agent that runs the same ten-stage program for any disease. To prove that claim, I re-ran the engine end-to-end on a second, unrelated disease — esophageal adenocarcinoma — and it produced a dual-arm program of its own: a GUCY2C T-cell engager and a DKK1 trap, with de novo binder leads across the esophageal-cancer trajectory. It matters because it puts patient-driven drug development within reach of the community it serves. I built a website () to share and describe these tools. ## Link to your work https://ruthannepai.netlify.app/ — project hub with links to all 10 open GitHub repositories (github.com/ruthannepai-tech), 3 manuscripts, and every construct and skill. ## How did you use Claude? I used Claude Science as my whole research environment. I don't write Python and had never used a protein model, and Claude Science let me work at the level of scientific reasoning while it handled the code. It mattered most in four places. First, omics mining: it pulled and analyzed public EoE datasets into a defensible target landscape. Second, protein design: it ran ESM-family and structure-prediction models to design de novo binders and pMHC-II constructs with confidence scoring I could interrogate — and it ran the entire pipeline twice, once for EoE and once for esophageal cancer, from the same reusable engine. Third, honesty: when I overclaimed — calling structure prediction "validation," or a target "proven" before trials — it caught and corrected me; and when my negative controls showed scrambled-sequence decoys clearing the same confidence gate as my real designs, it made me report that limitation plainly rather than bury it. That is what made the science defensible. Fourth, building: it taught me to publish skills to GitHub, run jobs on Modal, and make molecular videos in ChimeraX and Blender. I also packaged everything into a reusable Specialist Agent and built a website. Claude Science turned a one-person week into a full, reproducible AI-to-clinic program that anyone can pick up and run. ## Thoughts/feedback on building with Claude Science The most striking part was how quickly it collapsed the distance between a scientific question and a working result. What I expected to accomplish in a week was often accomplished in one day. As a bench-trained scientist who doesn't code, I expected tooling to be the wall, but instead it disappeared. What I valued most was that Claude Science and I worked together. I shared disease context, led in every step of the project strategy, and flagged ethical and accessibility concerns. And Claude proposed and executed elegant workflows and pushed me to state things I could stand behind, which is exactly what you want from a collaborator. There were several times when I thought of an idea (e.g. creating videos in Blender, building a website) and figured I'd see if Claude could help. We were always able to figure it out together, and I became more confident running code through this experience. Publishing my first skills for Claude Science and seeing them become reusable by anyone was the moment the "democratization" framing stopped being a slogan and felt real. It's been transformative and I can't wait to keep building.