OpenEndo · Research agenda
The invisible disease meets
visible compute.
190 million patients. Decades of underfunding. And a force multiplier — AI, agents and open algorithms — that costs nothing to share. Here is what compute can actually do for endometriosis, and what we're already running.
Why compute matters here
Endometriosis research is starved of money but not of data: every trial registry, every PubMed abstract, every national health register is a dataset waiting to be mined. Compute is the cheapest way to multiply scarce research capacity — and it is already working:
- A computational drug-repurposing pipeline recently flagged simvastatin and primaquine as candidate therapies (iScience, Aug 2026 — PMID 42668641).
- The first AI-based diagnostics are already in clinics: the Ziwig Endotest saliva test (miRNA signature + ML, CE-IVD) shortens a 7–10 year diagnostic odyssey to one swab.
- Denmark's national registries — every diagnosis, prescription, surgery and absence from work, linkable per citizen — are among the world's best resources for causal, population-scale endometriosis research. They are barely being used for this disease.
1 · Understand — building broader understanding
Living evidence synthesis running
LLM-assisted reviews that never go stale: the weekly PubMed feed is ingested, summarized and pushed to subscribers — instead of a static review written once every five years. We run this every Monday.
Knowledge graphs & literature mining
Map the disease: genes, mechanisms, drug targets, trials, guidelines — as one linked graph. Ask questions no single paper answers: "which mechanisms are drugged, which are ignored?"
Registry & EHR analytics high impact
Danish registries can quantify the diagnostic delay per region, the cost of untreated disease, surgical recurrence rates, and the fertility trajectories of 190,000 women. Causal inference, not just correlation.
-Omics integration
GWAS, single-cell and miRNA data are piling up. ML can cluster the disease into biological subtypes (endotypes) that explain why one woman responds to hormones and another doesn't.
Patient-generated data
Symptom trackers, cycle apps and pain diaries hold the lived experience no trial captures. Aggregated and anonymized, they are evidence — the patient's voice as a dataset.
Imaging AI
Ultrasound and MRI segmentation models trained on verified lesions can make non-invasive diagnosis routine — the difference between years of dismissal and months to answers.
2 · Research — doing actual useful research
Drug repurposing proof exists
Transcriptomics, connectivity maps and LLM screening over approved drugs can surface candidates for ~10× less cost and time than de novo development. The simvastatin/primaquine finding is the template — it needs replication and clinical follow-up.
Biomarker development
ML over -omics to build and validate diagnostic and prognostic signatures — the Endotest path, made open. Every dataset we publish is a training set someone else can use.
Trial intelligence running
Gap analysis over trial registries (what's being studied, where, by whom), patient-trial matching, and synthetic control arms that make trials smaller and faster. The dashboard on this site is the visible part.
Economic modeling
Cost-of-illness models (€69B/year globally) and ROI of earlier diagnosis — the ammunition that wins funding arguments, one health minister at a time.
In-silico & mechanistic models
Simulations of hormonal dynamics and lesion progression to test hypotheses cheaply before clinical work — a sandbox for the disease.
Federated learning
Train models across hospitals without moving a single patient record. Privacy-preserving ML is the only ethical way to scale to the data that already exists.
3 · Create — building solutions
Open infrastructure running
This project: trials, papers, funding and policy as open, weekly-refreshed data. The public-good layer every other tool can build on.
Decision support
Triage tools for primary care that shorten the diagnostic delay: symptom profiles + imaging hints + red flags, built from registry-grade evidence — a GP's second opinion in 30 seconds.
Patient-facing tools
Plain-language trial finders, treatment explainers, symptom trackers that speak the patient's language — including translations, because half the world's patients aren't served in English.
Autonomous research agents running
Agents that monitor registries and regulators, synthesize the week's findings, flag funding deadlines, and draft grant sections — with verification as a hard rule. One of them wrote parts of this page.
Policy simulation
What-if models for screening programs, capacity planning and diagnostic-delay targets — the numbers a politician can act on before the next election.
What we run today
- Weekly research monitor — ClinicalTrials.gov + PubMed + news + funding deadlines, synthesized and delivered every Monday.
- Live open dashboard — this site, regenerated weekly by
scripts/update_data.py. - Funding watch — deadlines with countdowns (next: 13 Sep 2026, the Danish 160M DKK women's health centre; 22 Sep 2026, DoD PRMRP).
- Danish one-pager — evidence-pack for politicians, ready to print.
Starter projects for contributors
EUCTR mirror new
Pull EU Clinical Trials Register data into the same schema — Europe's trials are invisible in ClinicalTrials.gov.
Weekly paper digest new
An LLM pipeline that turns the weekly PubMed feed into a readable, source-linked digest in multiple languages.
Repurposing tracker new
A living list of computational repurposing candidates (like simvastatin/primaquine) with evidence links and trial status.
Diagnostic-delay map new
Quantify time-to-diagnosis from registries where possible, patient surveys elsewhere — region by region. The 7–10 years becomes a number politicians can't ignore.
Translation agents new
Use LLMs to translate the site and digest into ES/FR/DE/PT — with native-speaker review as the gate.
Ground rules — non-negotiable
- Evidence-first: no hallucinated citations, ever. Every claim links to a verifiable source; every link is checked.
- Patient-respect: the people behind the data come first — their privacy, their language, their dignity. No "sufferers", no miracle cures.
- No hype: "promising" is a compliment we earn; "breakthrough" is a word we ban.
- Open by default: MIT, data as public good, models and prompts published when they can be.
- Human-in-the-loop: agents monitor, synthesize and draft; clinicians and patients decide. Nothing clinical ships without a human.