In October 26th, Tekkare’s CEO Bruno Sarfati was in Amsterdam for the World Orphan Drug Congress Europe 2026, the largest meeting dedicated to orphan drugs and rare diseases. He spoke in the workshop “AI for Rare Diseases: From Promise to Practice,” organised by ERDERA, the European Rare Diseases Research Alliance, of which Tekkare is a partner.
The session asked a fair question: can AI really reduce the risk of developing a drug for a rare disease? Bruno’s answer was deliberately narrow. Not the biological risk, which stays biology. But a large part of the risk sits elsewhere, and that part AI can reduce.
“AI doesn’t make the decision. It makes sure you decide with everything already known, and that you hear the day something changes.”
Bruno Sarfati, CEO and Founder, Tekkare
The rare disease knowledge paradox
We often say there is no data in rare diseases. For most of them, the opposite is true: a great deal is already known. The trouble is that it sits in different places, trial registries, publications and case reports, patent offices, regulators and payers, in several languages, each updating on its own calendar.
The practical cost is real. For a single rare disease, a team can spend weeks just finding out what is already known, and still miss part of it.
Two kinds of risk
Bruno drew a line between two risks. The first is scientific risk: does the mechanism work in patients? That is biology, small samples and endpoints, and AI helps only at the margin. The second is knowledge risk: did we miss something that changes the plan? A competitor entering the clinic, a new patent, a payer decision. This risk is not static, it moves every week, and this is where AI genuinely changes the game.
What it takes to stay current on a single disease
To keep up with one rare disease, a team has to watch science (PubMed, published patient cases, Orphanet, genes and variants), clinical trials (ClinicalTrials.gov and the EU registry, including first-in-human starts), patents, and the health agencies, FDA and EMA alongside Europe’s main five, HAS, G-BA, AIFA, AEMPS and MHRA.
Around 18 sources. 5 languages. Updated daily, each on its own calendar. Multiplied by every disease in a portfolio.
No human team can follow all of this, every day, across a portfolio. Something will be missed. That is precisely the job AI is good at, and it is what Tekkare built OIP Discovery® to do.
The live case: an oral FGFR3 inhibitor for achondroplasia
To make it concrete, Bruno took the position of a biotech developing an oral FGFR3 inhibitor for achondroplasia. On day one, OIP Discovery® gives the baseline across every source: 106 patients described one by one in the literature, 65 clinical trials including 12 in the EU registry, 100 patent families of which 16 are new in 2026, and 2 approved drugs with 10 orphan designations.
A snapshot, though, is not enough, because strategy breaks between snapshots. So the real value is turning the watch on. Here is what it would have sent over 22 months, from six sources and five regions:
- Dec 2024, Trials, Japan: Kyowa Kirin starts a phase 3 of KK8398.
- Feb 2025, Agency, HAS: Vosoritide downgraded from ASMR III to ASMR V in children under 5.
- Mar 2025, Trials, EU and US: TYRA-300, an oral FGFR3 inhibitor, enters phase 2.
- Nov 2025, Trials, EU: Infigratinib moves into infants, with a dose-finding part in the EU.
- Dec 2025, Trials, China: ABSK061, an FGFR2/3 inhibitor, registers its first-in-patient trial.
- Feb 2026, Agency, FDA: Navepegritide approved, the second drug in the disease.
- Jul 2026, Trials, EU: BMN 333 head-to-head against vosoritide authorised.
- 2026, Patents: 16 new patent families mentioning achondroplasia.
No single database, and no single person, would have caught all of them. Three in particular would change a development plan. When ABSK061 enters the clinic in children, a molecule close to yours from a company you may not be watching, the response is to check selectivity, dosing and patents now, not at the next quarterly review. When the French HAS moves vosoritide from ASMR III to ASMR V in under-5s, because the primary endpoint was not met and no quality-of-life gain was shown, the lesson is to plan quality-of-life endpoints from phase 2. When the FDA approves navepegritide, a second therapy, placebo-controlled trials become harder to justify.
From alerts to a decision
Read together, the alerts become a few sentences a team can act on: a second FGFR inhibitor has entered the clinic, a second CNP analogue has been approved, and French payers no longer see added benefit in under-5s without quality-of-life data, so consider quality-of-life endpoints, an infant cohort and an active comparator. Every sentence links back to an official source the team can open, the trial registry, the FDA decision, the HAS opinion. No web search, no guessing. The agent suggests options, it does not decide. The decision stays with the team.
Where AI is good enough, and where it is not
Bruno was clear about the limits, which is part of why the approach is trusted.
Good enough: watching every source every day, structuring and linking the evidence, flagging what changed, summarising with sources.
Not good enough: predicting the biology, seeing what was never published, making the decision.
One live example of the guardrails: in the data, 54 papers tagged “achondroplasia” actually describe another condition. The system caught them and kept them out of the counts. The rule Bruno set is simple: AI is good enough when every claim can be traced to a source a human can check.
The take-home
AI does not make the decision. It makes sure the team decides with everything already known, and hears the day something changes. Less knowledge risk means fewer surprises at phase 3, at approval and at reimbursement. It is also what ERDERA is building at scale with its Data Hub and accelerator: the right knowledge, at the right time, for the people who decide.
Figures shown during the event are from OIP Discovery®, data as of 19 October 2026. Patient counts measure literature coverage, not prevalence. Tekkare is ISO 27001 certified and HDS certified for the hosting of health data.
#RareDiseases #OrphanDrugs #WODC2026 #DrugDevelopment #ERDERA #KnowledgeRisk #OIPDiscovery
