I'm Ruth-Anne Pai, PhD — an immunologist, protein designer, and person living with eosinophilic esophagitis (EoE). In one week I built an end-to-end body of work: mining public omics to nominate targets, designing antigen-specific protein therapeutics with ESM-family and folding models, and packaging it all into reusable software, three manuscripts, and a development plan — then proved the same engine generalizes by pointing it at a second disease, esophageal cancer. Everything here is open.
How to read this work. This is a one-week, one-person citizen-science project, produced with substantial AI assistance (Anthropic Claude Science) under my direction. Every result here is computational (in-silico) and has not been experimentally or clinically validated; the manuscripts are preprints that have not been peer reviewed. It is shared openly as a method and a set of hypotheses — not medical advice, and not established results.
The work spans data mining, protein design, software, and a full biotech plan. Pick the entry point that fits you — each card points to the pages built for you.
The Therapeutic Program Architect turns your community's disease into a full, human-in-the-loop drug program. And the EoE plan was written patient-voice-first.
Read the methods and structures behind the pMHC-II tolerance work, the CCL26/POSTN effector binders, and the GUCY2C/DKK1 cancer-interception binders — three de novo campaigns, one pipeline.
The full package: platform thesis, market, competitive whitespace, regulatory path, budget, and milestones.
Ten open repositories — skills for target mining, epitope mapping, protein design, manuscript drafting, market analysis, peer review, and media.
A ten-stage pipeline from public GEO/PRIDE/CELLxGENE data to a ranked, druggability-annotated target shortlist.
How a citizen-scientist took EoE from "here is a disease" to IND-ready design and a plan — all in public.
The throughline isn't a single molecule — it's a reusable pipeline and a pair of AI specialists that take any disease from public data to a full program and a written manuscript. Two agents do the work end to end: the Therapeutic Program Architect (disease → program) and the Manuscript Architect (program → preprint). See the engine that works for any disease →
To prove it, I ran it end to end on three programs across two diseases:
An antigen-selection engine that turns a person's HLA type into a ranked, manufacturable epitope set — 4,640 peptide×allele evaluations → 832 strong binders — with a folded, groove-validated milk lead on a tolerogenic nanoparticle.
A 9-cohort transcriptomic meta-analysis surfaced two high-evidence, no-trial targets; 80→1,920→60 de novo binder designs followed — reported with an honest negative-control calibration.
The same engine pointed at esophageal cancer: TCGA genomics → a whitespace map → two nominated targets (GUCY2C, DKK1) → 16 de novo binders, 15/16 clearing the interface line — proof the method generalizes.
{hl}
' if hl else "" return f"""""" repo_sections = "" for group, repos in REPOS.items(): cards = "".join(repo_card(r) for r in repos) repo_sections += f"""Ten open-source repositories from the hackathon — a disease-agnostic drug-program pipeline, a manuscript-architect agent, target-discovery, epitope-mapping and protein-design tool stacks, market and peer-review skills, and a scientific-media engine. Most are packaged as Claude Science skills you can drop in and run.
From a disease name to a ranked, druggability-annotated target shortlist, using only public data. Each stage
ends by saving artifacts, so a run is fully auditable. Two core statistics (moderated_de,
meta_analysis) are dependency-light and usable stand-alone.
from kernel import moderated_de, meta_analysis
# per-study DE on log2 expression (genes x samples)
de = moderated_de(expr, labels, case="disease", control="control")
# pool per-study effects (log2FC and SE across studies)
meta = meta_analysis(fc_df, se_df, min_studies=3)Give it any disease — common or genetic/rare — and it runs a full, human-in-the-loop drug-development program: from unmet need and patient priorities, through data- and literature-mined target and modality selection, AI/ML design and in-silico pressure-testing, to scientific, regulatory, and commercial strategy. It stops and asks you at every consequential decision.
The disease, target, and modality are all parameters resolved from data and your decisions, not hard-coded. The same modality router that produced a tolerizing vaccine for eosinophilic esophagitis produces a ligand-trap, antibody, or small molecule for another disease depending on the target biology. If your community has a disease and a question, this is built for you.
A single JSON program dossier is the source of truth for a whole run — disease, scope, every decision and who made it, per-stage artifacts, gate status, and open questions. A program can be paused, resumed, and audited.
The pipeline never silently commits a target, modality, endpoint, price, or go/no-go. It stops at named gates and asks a structured question with its recommendation and options.
Idiopathic pulmonary fibrosis. The modality router lands on a ligand-trap / antibody / small molecule path driven by the target biology — not the tolerizing vaccine EoE called for.
Spinal muscular atrophy. Exercises the genetic/rare archetype branch — a monogenic, mechanism-driven program with a different modality decision tree.
Eosinophilic esophagitis. The complete run behind Project Tolera — a personalized pMHC-II tolerance vaccine, end to end.
It comes in two pieces: drug-program-orchestration (the conductor skill) and the Therapeutic Program Architect (a specialist agent that preloads the skill and has full access to the skill catalog and all data/regulatory connectors).
git clone {GH_ARCHITECT}.gitRun the drug program for <your disease>.Load the drug-program-orchestration skill
and run the pipeline for <your disease>.The design stage (Stage 6) benefits from a GPU compute target — the orchestrator checks for one and runs the CPU-feasible parts or defers design if none is configured. Connector calls (ClinicalTrials.gov, Open Targets, FDA, GEO/PRIDE, ChEMBL) run automatically inside the specialist. All forward-looking scientific, regulatory, and commercial statements are development hypotheses — not established results, medical advice, or guaranteed outcomes.
A personalized antigen-selection engine for pMHC-II tolerance in EoE; a target-and-binder brief on the CCL26 / POSTN effector axis; and a dual-arm de novo design campaign for esophageal cancer that shows the same engine generalizes to a second disease. All three are computational preprints (not peer reviewed), openly documented, with structures and data available to download.
Ruth-Anne Pai, PhD · immunologist and person living with EoE · produced with Claude Science (AI) support
Rather than broadly suppressing the immune system, this work re-educates the specific food-reactive T-cell clones that drive EoE — and it computes which epitopes matter for a given person. Because presented-epitope load is strongly HLA-dependent, antigen priority is per-patient. The engine scans the dominant dairy, wheat, and soy allergens across common HLA-DRB1 alleles, then hands a folded, groove-validated lead to a tolerogenic nanoparticle (PACT: one shared backbone, a swappable peptide cassette).
Three pMHC-II:peptide complexes were co-folded (ESMFold2-Fast, GPU) and inspected for canonical groove
geometry. Download the atomic coordinates below — .pdb files open in PyMOL, ChimeraX, or any
Mol* viewer.
| Complex | Peptide (15-mer) | ipTM | Structure |
|---|---|---|---|
| Dairy pMHC-II | KIHPFAQTQSLVYPF | 0.872 | dairy_pmhc.pdb |
| Wheat pMHC-II | IHNVVHAIILHQQQQ | 0.896 | wheat_pmhc.pdb |
| Soy pMHC-II | AYPFVVNATSNLNFL | 0.891 | soy_pmhc.pdb |
Ruth-Anne Pai, PhD · immunologist and person living with EoE · produced with Claude Science (AI) support
CCL26 and POSTN emerged from a nine-cohort transcriptomic meta-analysis as high-evidence, no-trial biology — the eosinophil-recruitment axis and the barrier-remodeling axis. From 80 RFdiffusion backbones → 1,920 SolubleMPNN sequences → 60 Boltz-2-assessed complexes, lead binders engage the intended functional epitopes: CCL26 ipTM 0.942 (4/5 hotspots, 1,220 Ų buried) and POSTN ipTM 0.901 (pLDDT 0.911, 3/4 hotspots).
A neutralizing anti-CCL26 IgG1 must occlude the N-loop / 40s-loop docking surface to block CCR3 site-1 engagement — a high-affinity (KD < 1 nM) requirement for the small, 71-aa antigen. Alternatively, a small-molecule CCR3 antagonist targets the invariant receptor and captures the other eotaxins (CCL11/CCL24) on the same axis for a broader anti-eosinophil effect.
The sole ESM escape-risk position in the CCL26 epitope (H39) sits inside the N-loop docking surface — so the design co-engages the disulfide-rigid, conserved 40s loop rather than relying on H39. For the IL1RL1 interface, CDRs are steered toward conserved aromatic hotspots (Y119 / F245) and away from the escape-prone rim.
Effector-specific eosinophil-recruitment blockade — narrower and more targeted than dupilumab's broad Th2 blockade, and universal across EoE endotypes.
We ran 36 scrambled and random decoy complexes through the same pipeline. 92–100% of decoys cleared the same ipTM/pLDDT gate that 59 of 60 designs cleared — so a high structural pass rate is close to the null and, on its own, does not demonstrate a real binder. What separates the leads is focal, on-epitope engagement, not the confidence score. Claude Science flagged this over-claim during the work; reporting it is part of the point.
Ruth-Anne Pai, PhD · immunologist and person living with EoE · produced with Claude Science (AI) support
The strongest test of a platform is whether it works on a disease it wasn't built for. I pointed the same engine at esophageal adenocarcinoma (EAC) — a copy-number-driven cancer at the end of the Barrett's trajectory — and asked it to find tractable targets in competitive whitespace, then design binders. It nominated a dual-arm program: GUCY2C for a T-cell-engager treatment arm, and DKK1 for a neutralizing-trap interception arm in high-risk Barrett's.
gucy2c_bb2 (ipTM 0.915; flagged for a high-alanine developability liability)
and DKK1 dkk1_bb1 (ipTM 0.858, pLDDT 0.908 — the cleanest design in the set). Reported as
pilot-scale in-silico leads: ipTM is a docking-confidence proxy, not a measured affinity.A gut-restricted surface antigen (GTEx: ~85% GI-restricted expression) nominated #1 in 80.7% of 20,000 weight-perturbation draws, with CDH17 as a backup. Designed for a T-cell-engager format. Open question: the normal-gut therapeutic window.
A secreted Wnt modulator targeted with a neutralizing trap for high-risk Barrett's — interception before invasive cancer. Open question: a dysplasia-regression surrogate endpoint that a trial could actually read out.
Program verdict: TRACTABLE (0.87). The two make-or-break questions above are named as open, not claimed. This is a design-stage dossier, not a clinical result.
A citizen-scientist's roadmap for a personalized pMHC-II tolerance vaccine in eosinophilic esophagitis — from AI/ML design to first-in-patient. EoE is, at its root, antigen-specific: a food peptide presented on an MHC-II molecule to a pathogenic T-cell clone. Yet every approved and pipeline therapy is a broad, chronically-dosed symptom suppressor. Project Tolera maps the path to a one-course therapy that retires the clone.
Personalized therapy is only investable if manufacturing is tractable. PACT uses a shared, GPU-validated MHC-II backbone with a swappable peptide cassette — only the peptide changes per patient, so the expensive, regulated components are standardized and validated once. Vein-to-vaccine target: 6–10 weeks.
| Layer | Definition | Patients (US) |
|---|---|---|
| Prevalence | US EoE (~1 in 2,000) | ~167,500 |
| Diagnosed | Clinically diagnosed (~60%) | ~100,500 |
| SAM | Biologic-eligible (moderate–severe / failing conventional care) | ~55,300 |
| SOM (mature) | Realistic annual serviceable share (~12% of SAM) | ~6,600 / yr |
Antigen-specific tolerance is a hard, worthwhile problem, and the field has built the groundwork Tolera stands on. EoE has no dedicated patient-focused drug-development (PFDD) meeting and no pMHC program yet — an opening to bring patient priorities in early, capturing them before the trial is designed rather than after.
An EoE pMHC vaccine follows the trail blazed by Tzield (T1D), KAN-101 (celiac), and dupilumab (EoE). The plan pursues an expedited-designation sequence and engages FDA early on whether histologic remission can serve as a registration-enabling surrogate.
~$8M carries the concept from validated in-silico design to a cleared IND and first-in-human dosing — the single largest de-risking step available today. A further ~$35M funds through the Phase 2 durable-remission proof-of-concept. These figures scope the work, not a raise.
The complete Project Tolera package — business plan, scientific plan, financial model, market analysis, and investor deck. Everything below is downloadable.
To test whether the pipeline behind Project Tolera was specific to EoE or genuinely disease-agnostic, I pointed the same engine at a very different problem — esophageal adenocarcinoma (EAC) and its precursor, Barrett's esophagus. Starting from public genomics and the competitive trial landscape, it nominated two targets and delivered de novo binder leads for each: a GUCY2C T-cell-engager for established cancer, and a DKK1 neutralizing trap to intercept the Barrett's→cancer transition. Same method, second disease.
Why this page exists. This is not a second company — it is a generalization test. Project Tolera is the deep, patient-voice-first EoE program; this page is the proof that the underlying engine re-runs on a different disease, biology, and modality without hand-tuning. The two disease pages reference each other by design.
Public genomics from TCGA/cBioPortal (esca cohort, n=182) show EAC is dominated by copy-number change rather than druggable point mutations — TP53 altered in 87%, CDKN2A deleted in 39%, CCND1 amplified in 35%. Against 351 interventional EAC trials (106 active), most cluster around the same checkpoint and HER2 mechanisms. The engine triaged 16 candidate targets and nominated two that sit in open space.
GUCY2C is a cell-surface receptor with GI-restricted normal expression, retained in a large fraction of GI adenocarcinomas — a handle for a T-cell-engaging biologic in established EAC. The engine ranked it first among the triaged targets and designed de novo binders against it.
Make-or-break question: the normal-gut therapeutic window. Named as an open question in the manuscript.
DKK1 is a secreted Wnt modulator implicated in the Barrett's→dysplasia→EAC progression. A neutralizing trap offers a route to intercept the transition in high-risk Barrett's patients rather than only treating cancer once it forms.
Make-or-break question: a dysplasia-regression surrogate endpoint. Named as an open question in the manuscript.
Eight backbones per target were carried through the same design-and-gate pipeline used across the other programs. 15 of 16 cleared the ipTM > 0.5 interface threshold; the leads are a GUCY2C binder at ipTM 0.915 and a DKK1 binder at ipTM 0.858. As everywhere on this site, these are computational structure predictions — leads to test, not validated binders.
gucy2c_bb2 at ipTM 0.915
(flagged for a high-alanine developability liability); DKK1 lead dkk1_bb1 at ipTM 0.858, the
cleanest of the set. 15/16 designs clear the interface line.Program verdict: TRACTABLE (0.87). The dual-arm program is judged tractable, with two named make-or-break questions — the GUCY2C normal-gut window and a DKK1 dysplasia-regression surrogate endpoint — that would gate any real-world development. Everything here is in-silico and not experimentally validated.
The full dual-arm manuscript and the two lead complexes are downloadable. The manuscript sits alongside the EoE work on the manuscripts page.
The molecules on this site are demonstrations. The real deliverable is the engine that produced them: a pair of human-in-the-loop AI specialists that a non-coding scientist can point at any disease. Part 1 turns a disease into a full, decision-gated drug-development program. Part 2 turns that program into a written, peer-review-style manuscript. Together they take one person from “here is a disease” to a program and the paper that documents it — the same path this site walked three times, across two diseases.
Why it matters. I am one person, living with the disease I started on, and I do not write production code. In one week these two agents let me run a credible, honest, end-to-end drug-discovery program — and then generalize it to a second disease I have no personal stake in. That is the case: not a single result, but a repeatable, inspectable method that puts a full program within reach of any motivated researcher or patient community.
Give it any disease — common, rare, or genetic — and it runs a ten-stage, human-in-the-loop program: from unmet need and patient priorities, through data- and literature-mined target and modality selection, AI/ML design and in-silico pressure-testing, to scientific, regulatory, and commercial strategy. It stops and asks you at every consequential decision, and it flags its own over-claims.
A program is only useful if others can read, check, and build on it. The Manuscript Architect turns a completed program dossier into a structured, bioRxiv-style preprint — then hardens it through two rounds of synthetic peer review, generating reviewer critiques and point-by-point responses so the weak claims surface before a human reviewer ever sees them.
Assembles the program's targets, designs, methods, and figures into a complete manuscript with the sections and rigor a preprint server expects.
Generates independent reviewer critiques and the author responses to them — a self-correcting loop that catches over-claims and gaps early.
Every computational result is labeled in-silico, never “validated.” The honesty line is enforced in the writing, not bolted on afterward.
The three manuscripts on this site were produced this way — read them →
The same two agents produced every program on this site. The clearest evidence that the method generalizes is that it re-ran cleanly on a disease with entirely different biology and a different modality.
Personalized antigen selection for eosinophilic esophagitis — 4,640 peptide×allele evaluations → 832 strong binders.
A 9-cohort meta-analysis to two no-trial targets, then 80→1,920→60 de novo designs with an honest negative-control calibration.
The generalization test: public genomics → two nominated targets → 16 binders, 15/16 clearing the interface line.
Both specialists and the skills behind them are public, installable, and reproducible — part of ten open repositories spanning target mining, epitope mapping, protein design, market analysis, peer review, and media.
Immunologist, protein designer, and a person living with eosinophilic esophagitis. I build open, patient-first therapeutic science — and this project is what one week of it looks like when the patient and the scientist are the same person.
For many patients the path to an EoE diagnosis is long and indirect, and the treatment that follows is effective for many but fundamentally incomplete: PPIs, swallowed steroids, elimination diets, and/or dupilumab may reduce inflammation, but none is curative, all require indefinite adherence, and inflammation typically returns when treatment stops. Burden doesn't show up in a biopsy — chewing every bite an unusual number of times, drinking with every mouthful, a quiet, rational anxiety around eating out, and social anxiety and isolation. This is a disease that asks for constant, low-grade attention, meal after meal, for life.
That lived reality is not incidental to this work; it is its origin. When the immunologist and the patient are the same person, patient experience becomes a design input rather than a marketing note. It shapes which priorities lead, which endpoints matter, and how a community would be engaged.
With access to computational tools (folding and ESM-family models on GPU, public literature and omics, epitope prediction), clear immunological principles, and one week of focus, this project went from “here is a disease” to: novel epitopes identified from the EoE literature, structurally validated at atomic resolution; a multivalent tolerance platform grounded in published precedent; a complete preclinical development path with decision gates and budgets; and reusable software so anyone can do the same for their disease. To test that last claim, I pointed the same engine at a second, very different disease — esophageal cancer — and it produced a dual-arm target-and-binder program of its own. The next steps — laboratory validation and clinical translation — require a team. But the design is ready to be picked up, tested, and, if it holds, carried to patients.
A disease-agnostic drug-development specialist — for any community, any disease.
Three preprints: pMHC-II tolerance, the CCL26 / POSTN effector axis, and a GUCY2C / DKK1 program for esophageal cancer.
If this resonates with you — as a researcher, clinician, patient-advocate, funder, or biotech partner — I'd like to hear from you.
Work developed during the Built with Claude: Life Sciences Hackathon, 2026.