Dupuytren's disease & drug repurposing
The tools, data sources and decisions behind the site, the mistakes we found, and the checks I'd run on a project like this.
Why it’s usefulAI research tools are fast, but they make mistakes. This page lists ours so you can judge how far to trust the rest.
I ran the research in Claude Science, Anthropic's AI research workbench, in five phases. After each phase I read the output, asked questions and set the direction for the next one.
I then built this website with Claude, which also cross-checked the files against each other and found several of the mistakes listed below.
Decisions I made along the way
Claude Science reaches databases through "connectors". These are the eight it used. The databases are what to cite; the connector is just the route in.
| Connector | Database | What it did here |
|---|---|---|
| pubmed | PubMed / PMC | Every literature search and citation: the 566-paper corpus and the 112-study evidence map |
| clinical-trials | ClinicalTrials.gov | All 80 registered studies, with their endpoints |
| human-genetics | GWAS Catalog | 11 studies and 255 genetic associations |
| clinical-genomics | Open Targets Platform (API v26.9.0) | 458 associated targets and drug records |
| genes-ontologies | UniProt, Reactome (release 98), MyGene | Protein identity and pathway membership |
| chembl | ChEMBL | How each candidate works and how far it has got clinically |
| drug-regulatory | Drugs@FDA (2026-10-29), openFDA labels | Approval dates, generic counts, boxed warnings and dosing |
| structures-interactions | RCSB PDB, AlphaFold DB (v6) | Finding and checking the protein structures |
Skills
figure-style for publication-quality figures (it made the genetics chart) and parts of indication-dossier: its rules against made-up facts and its regulatory and trials module. Each connector also came with its own documentation, loaded before first use.
Available, but not used on purpose
Boltz, AlphaFold2, Chai-1 and DiffDock could have predicted how the drugs bind. Claude Science chose not to run them, because pirfenidone and metformin have no established target to dock against. That was deliberate: a predicted complex would have looked like evidence without being evidence.
My hand-check
Loading…
Claude Science's own checks
Ten so far. Most were in the written summaries rather than the data. The full error log shows which research phase each one came from.
Phase 2 · fixed before the site
The first report said only one drug had reached a randomised trial. It had counted trials by treatment type, which filed drug-plus-surgery trials under surgery. Tamoxifen's randomised trial PMID 24740662 had been missed.
Phase 2 · found by Claude Science
One record said recurrence was higher after fasciectomy. The source says fasciotomy PMID 26648251.
Phase 2 · found while building the site
The report said company-funded collagenase studies looked better. The paper shows the opposite: independent studies reported higher success PMID 38527344.
Phase 3 · found while building the site
Pirfenidone's "first FDA approval, 2011" is its EU and Japan date. The US approval was October 2014.
Phase 3 · found while building the site
The notes say the network has 7 weak links. The data file has 9. The site now counts directly from the data.
Phases 3–4 · found from your question
The metformin case cited a study where Dupuytren's was rarer in metformin users. A 2004 study in the same search results found the opposite PMID 15336742. Neither is a trial, and both are confounded by diabetes.
Hand-check
The check file arrived with all 20 rows marked correct, but I had only checked two. It now shows 18 pending.
If you use an AI research tool on your own topic, or rerun this one, these are the checks I'd build in. Most are quick, and each would have caught something here.
| Check | Why | What it caught or would have caught here |
|---|---|---|
| Recount every headline number from the data | Summaries drift from the data they describe. | Weak-link and trial counts that were out of date (E06, E07) |
| Check every "higher" or "lower" against the figures | Direction errors read fluently and are easy to miss. | The sponsorship finding written backwards (E03) |
| Count the same thing two ways | A category can hide items. If two counts disagree, find out why. | Drug trials hidden under "surgery" (E01) |
| Say what each count is counting | Reports, trials and drugs get mixed up. | "Ten trials" that were really ten reports (E04) |
| Check that a fact agrees in every file | Later phases fix things earlier files still get wrong. | Pirfenidone's approval date (E05) |
| Hand-check a random sample against the sources | The only way to estimate the real error rate. | Swapped words like fasciectomy/fasciotomy (E02) |
| Check inputs sit inside their stated ranges | Odd values signal a mix-up in what a field means. | A questionnaire threshold outside its own range (E08) |
| Record who checked what | A blank, a model check and a human check look the same in a spreadsheet. | Rows marked as checked that weren't (E09) |
| Search for evidence against each main conclusion | AI summaries tend to cite the study that fits the story. | A 2004 study suggesting metformin users have more Dupuytren's, not less (E10) |
| List what it refused to guess | Good behaviour is worth confirming too. | Passed: the untreated progression rate and the durometer threshold were left as "not established" |
| Recompute key results independently | Confirms the maths, not just the wording. | Passed: all 20 trial-size scenarios matched exactly |