Dupuytren's disease & drug repurposing
How an existing drug gets approved for a new use in the UK, EU and US, and how many patients a trial would need.
Why it’s usefulCost is the biggest barrier. This page shows where the money goes, lets you size a trial yourself, and sets out where AI and projects like this one could bring the cost down.
Usually because its target turns up in the disease's biology. In Dupuytren's the strongest example is TNF: patient tissue is full of immune cells and TNF, and adding TNF to palm fibroblasts from patients (but not other cytokines, and not cells from people without the disease) turns them into contractile myofibroblasts PMID 23431165.
Existing safety data, pharmacology and any accidental human experience. A drug already taken by millions has a known safety profile, and that is the central economic advantage of repurposing.
Dupuytren's is unusually well suited to this: diseased tissue is removed in routine surgery, so patient-derived fibroblasts and whole-tissue explants are easy to obtain. That is why it is described as a tractable human model of fibrosis PMID 36566101. But lab success is not enough: 5-fluorouracil passed this step and failed its trial.
A dose-finding study, then a randomised controlled trial against placebo or standard care. The endpoint chosen here determines the cost of everything downstream: see the calculator below. RIDD shows an academic sponsor can do this: the University of Oxford ran a placebo-controlled trial of a licensed biologic in this disease under MHRA authorisation PMID 35949922.
A licence is sought to add the new indication, and a health-technology body decides whether the health service will pay. The route differs by region:
MHRA national procedure. Since Brexit the MHRA runs its own national route. RIDD was authorised by the MHRA (EudraCT 2015-001780-40) with the University of Oxford as sponsor PMID 35949922.
Hybrid application, relying partly on the original product's dossier. Revised EU legislation has two articles on repurposing, but a review concludes academic and not-for-profit sponsors still face barriers the new incentives don't remove: limited regulatory know-how, pharmacovigilance obligations and limits on data protection PMID 39756647.
505(b)(2) new drug application, which lets the applicant rely partly on data it did not generate. Of 1,001 NDAs approved 2010–2020, 570 used this route, mostly for new formulations PMID 37040834. It was designed for manufacturers; a university or charity has no straightforward way to use it. A proposed "labeling-only" 505(b)(2) would let non-manufacturers add a new indication to a well-established generic drug PMID 39444616.
Public money is generally available for early research, but not for the confirmatory trials that a licence needs. With mostly academic and non-profit sponsors who don't aim for a licence, the chance of successful authorisation falls below 30% PMID 41208860. A review of 192 studies identified 33 distinct barriers across nine themes, and stakeholder groups disagree about which matter most PMID 41199589.
There is also a dispensing problem: where prescribing is by active ingredient, a pharmacy may substitute a cheaper generic not licensed for the new use, removing any return on the investment that produced the evidence PMID 41208860.
Same drug, same trial, opposite conclusions
per QALY for adalimumab, judged over the 12-month trial
per QALY over a lifetime, assuming repeat courses every three years
95% CI £7,534–£42,698; 77% probability of best value at £20,000/QALY PMID 36378072. The lifetime figure rests on an untested assumption about how often the disease reactivates.
Slow
Contracture worsened only on the small-finger side, by about 1.75–6.25° a year PMID 35404338. A trial that waits for the finger to bend needs years.
Uneven
11–16% of hands were stable or improved over five years untreated, so an unselected trial enrols people who would never have progressed.
Unpredictable
None of the traditional risk factors predicted progression after multivariable selection, so you can't easily enrich a trial for people who will progress.
The field's answer has been to measure the nodule instead of the finger. Nodule hardness on a durometer is biologically anchored: it tracks myofibroblast load on ultrasound PMID 34617826. But it has no validated minimal clinically important difference and has never been shown to predict contracture.
Standard two-sample formulas, checked against both published trials before use: RIDD's inputs give 68 per arm (published 69) and DISC's give 283 (published 284). Planning-grade, not protocol-grade.
People to randomise
DEPART randomised 404 hands to observation or low-dose radiotherapy. At 36 months, hand-related disability (QuickDASH) favoured radiotherapy: 10.9 vs 17.8, a statistically significant difference of 6.9 points PMID 41027544.
But the pooled minimal clinically important difference for QuickDASH is 12–15 points PMID 38438144. A treatment can improve a functional score detectably without improving it meaningfully. DEPART's efficacy endpoint, progression, has not yet been reported.
Limits: these are planning estimates. The spread in nodule hardness comes from a 25-person pilot, and the sample size changes with its square. A real protocol would use a mixed-effects model with baseline adjustment and the documented 14% rate of unobtainable durometer readings PMID 35949922. No FDA or EMA guidance specific to Dupuytren's was found.
The expensive part is the large confirmatory trial, and AI can't pay for that. What it can do is make everything before the trial cheaper, and make the trial itself smaller. These are the places I think the savings are.
1 · Gathering the evidence
Claude Science searched PubMed, the trials registry, genetics and drug databases and built this whole evidence base: 112 studies, 80 trials and 12 scored candidates, all referenced. Because it's scripted, it can be rerun when new studies come out.
Limit: it makes mistakes. Its own read-back found 1 error in 44 records, and more turned up later (see corrections). People still need to check it.
2 · Dropping weak candidates early
5-fluorouracil looked good in the lab and failed in patients. Comparing each new candidate with 5-FU in a pre-registered lab study (OP_04) is a medium-sized project that could stop a weak drug before it costs a trial.
3 · A smaller trial
The choice of what to measure moves a trial from about 136 to about 708 patients (try the calculator above). Recruiting people whose disease is visibly getting worse also cuts the numbers. The calculator runs these scenarios in seconds, using inputs taken from published trials.
4 · Data that already exists
Three of the most useful datasets in this disease already exist: RIDD's longer follow-up, a 400-patient steroid trial and DEPART's progression results (OP_07). Only 14 of 80 registered trials have posted results. Tracking what's missing across registries is easy for AI and costs far less than a new trial.
5 · Small checks that unblock funding
Three open problems are rated small effort: checking UK and EU patent status for the leading drugs (OP_05), testing whether Hedgehog signalling is active in Dupuytren's tissue (OP_09) and auditing old claims back to their sources (OP_12). AI can do much of the searching for all three.
6 · Cheaper follow-up
Trials here need two to three years of follow-up, plus time off the drug to see if the benefit lasts. A phone-based hand measure could replace many clinic visits. More on that idea.
Where AI doesn't help
It can't fund the confirmatory trial, recruit and treat patients, or take responsibility for a regulatory submission. For off-patent drugs that funding gap is the real bottleneck: fewer than 30% of these projects reach authorisation PMID 41208860. What AI can do is make a stronger, cheaper case for that funding.