What Role Can AI Play in the Diagnosis and Research of Rare Diseases?
This question was at the heart of the roundtable discussion organized by ALCIMED on January 21, 2026.
AI and Rare Diseases: From Technological Promise to Real-World Impact on Diagnosis and Care
Each rare disease affects fewer than one in 2,000 people, yet collectively they impact nearly 3 million people in France and 300 million worldwide. More than 6,500 rare diseases are currently described in Orphanet, the European reference database.
Despite this, the diagnostic odyssey in France still averages five years, a delay with significant medical, social, and psychological consequences.
In this context, artificial intelligence raises great expectations. But how can it be deployed in practice when data are scarce, heterogeneous, and highly sensitive?
A Foundational Paradox: Few Patients, Massive Complexity
Contrary to common belief, the challenge in rare diseases is not only the lack of data, but also their extreme diversity: clinical reports, imaging, biological data, genomic, phenotypic data, and more.
This heterogeneity far exceeds human analytical capacities and makes AI tools particularly relevant—provided they are properly framed and governed.
The first response lies in the collective structuring of knowledge. Public reference databases such as Orphanet play a central role, as do international collaborations and research consortia, such as the European ERDERA project led by INSERM.
A striking example illustrates this dynamic: an international study published in Nature in 2024 (Chen et al., 2024) identified a de novo variant responsible for a neurodevelopmental syndrome through retrospective analysis of patient records from multiple countries. This collaborative approach, involving laboratories from the French Genomic Medicine Plan (PFMG), led to genetic diagnoses that had previously been inaccessible.
Rare Data: How AI Addresses Small Sample Sizes
Faced with the limited number of patients per condition, several approaches are emerging:
-
Transfer learning, leveraging models trained on more common diseases
-
Synthetic data and digital twins, enabling the generation of artificial data or profiles while limiting re-identification risks
-
Hybrid methods combining clinical expertise with algorithmic power
These tools do not replace real-world data, but they help to test, guide, and prioritize research hypotheses.
Diagnosis: Real Potential, but Still Difficult to Measure
AI is often presented as a major lever to reduce diagnostic delays. In practice, the reality is more nuanced.
Rare disease diagnosis is intrinsically complex: high phenotypic variability, overlapping clinical signs, and multiple gene–disease relationships (a single gene may be involved in several distinct syndromes, and conversely, multiple genes may be linked to the same rare disease).
AI’s value lies in its ability to rapidly cross-reference thousands of heterogeneous data points.
Several applications have been developed in close collaboration with hospitals and research institutes to combine human medical expertise with AI’s analytical capacity. This is notably the case for three tools developed based on clinicians’ expertise from Necker Hospital: AYDI (Institut Imagine) for facial phenotyping, Soniopedia (Sonio) for prenatal diagnosis, and RDK, developed using Orphanet’s reference data to support diagnostic orientation and referral to expert centers.
However, no large-scale prospective study has yet demonstrated a measurable reduction in diagnostic delay attributable to AI alone. The available evidence remains limited, particularly for recent tools. The potential is clear, but proof of impact on the patient journey still needs to be established.
Therapeutics and Discovery: A Long-Term Strategic Lever
It is probably upstream, in therapeutic research and development, that AI is currently delivering its most tangible results.
Recent work suggests that AI could:
-
Accelerate molecule discovery
-
Optimize patient selection and recruitment
-
Improve clinical trial design
Some studies suggest that generative AI and automation could reduce clinical trial costs by up to 50% and shorten timelines by 12 months or more in optimistic scenarios (Niazi & Mariam, 2025), by improving molecule discovery, patient selection and recruitment, and protocol design. In addition, AI-discovered molecules show higher success rates in Phase I trials (80–90% versus 40–65% historically) (Serrano et al., 2024).
These significant reductions in time and cost must, however, be weighed against the intrinsic costs of AI itself (infrastructure, energy consumption, governance).
Going further, it is possible to envision AI-based tools capable of accelerating the entire R&D cycle, from exploratory phases to validation and optimization stages.
Developing tools that support researchers in navigating global scientific knowledge in a fast and structured way is a major challenge. This is precisely the perspective behind the design of OIP Discovery applications. These platforms aim to enable researchers working on rare diseases, cancer, or infectious diseases to access, in a unified and efficient manner, the state of the art in their field—including scientific publications, clinical signs, associated genes and variants, molecules of interest, clinical trials, and patents. The integration of AI agents provides complementary methodological support for exploration, structuring, and analysis of this information, while avoiding source fragmentation and the need to consult multiple heterogeneous databases.
AI also enables the exploration of drug repositioning, by identifying shared biological mechanisms between common and rare diseases. Genes such as BRCA1, involved both in oncology and certain rare diseases, illustrate this scientific continuity. Likewise, the cancer treatment repurposed by Professor Guillaume Canaud for CLOVES syndrome (Venot et al., 2018) stands as an inspiring example.
Areas for Progress and Blind Spots
In the short term, several challenges remain critical:
-
Model evaluation and quality control
-
Data security, governance, and sharing
-
Integration of AI tools into existing clinical environments
-
Prioritization of use cases in light of regulatory and economic constraints
AI in rare diseases does not raise fundamentally different questions from other medical fields—but it amplifies their demands.
Conclusion: Useful AI, or No AI at All
The future of AI in rare diseases will not be determined by algorithmic sophistication, but by its ability to transform fragmented data into actionable clinical decisions, with transparency, restraint, and demonstrated impact.
Hybrid models—combining human scientific expertise with algorithmic power—currently appear the most promising. Provided they remain guided by a clear objective: to meaningfully improve the patient journey, rather than merely showcase technological prowess.
Sources :
Targeted therapy in patients with PIK3CA-related overgrowth syndrome. Venot Q, Blanc T, Rabia SH, Berteloot L, Ladraa S, Duong JP, Blanc E, Johnson SC, Hoguin C, Boccara O, Sarnacki S, Boddaert N, Pannier S, Martinez F, Magassa S, Yamaguchi J, Knebelmann B, Merville P, Grenier N, Joly D, Cormier-Daire V, Michot C, Bole-Feysot C, Picard A, Soupre V, Lyonnet S, Sadoine J, Slimani L, Chaussain C, Laroche-Raynaud C, Guibaud L, Broissand C, Amiel J, Legendre C, Terzi F, Canaud G. Nature. 2018 Jun;558(7711):540-546. doi: 10.1038/s41586-018-0217-9. Epub 2018 Jun 13. PMID: 29899452
De novo variants in the RNU4-2 snRNA cause a frequent neurodevelopmental syndrome. Chen Y, Dawes R, Kim HC, Ljungdahl A, Stenton SL, Walker S, Lord J, Lemire G, Martin-Geary AC, Ganesh VS, Ma J, Ellingford JM, Delage E, D'Souza EN, Dong S, Adams DR, Allan K, Bakshi M, Baldwin EE, Berger SI, Bernstein JA, Bhatnagar I, Blair E, Brown NJ, Burrage LC, Chapman K, Coman DJ, Compton AG, Cunningham CA, D'Souza P, Danecek P, Délot EC, Dias KR, Elias ER, Elmslie F, Evans CA, Ewans L, Ezell K, Fraser JL, Gallacher L, Genetti CA, Goriely A, Grant CL, Haack T, Higgs JE, Hinch AG, Hurles ME, Kuechler A, Lachlan KL, Lalani SR, Lecoquierre F, Leitão E, Fevre AL, Leventer RJ, Liebelt JE, Lindsay S, Lockhart PJ, Ma AS, Macnamara EF, Mansour S, Maurer TM, Mendez HR, Metcalfe K, Montgomery SB, Moosajee M, Nassogne MC, Neumann S, O'Donoghue M, O'Leary M, Palmer EE, Pattani N, Phillips J, Pitsava G, Pysar R, Rehm HL, Reuter CM, Revencu N, Riess A, Rius R, Rodan L, Roscioli T, Rosenfeld JA, Sachdev R, Shaw-Smith CJ, Simons C, Sisodiya SM, Snell P, St Clair L, Stark Z, Stewart HS, Tan TY, Tan NB, Temple SEL, Thorburn DR, Tifft CJ, Uebergang E, VanNoy GE, Vasudevan P, Vilain E, Viskochil DH, Wedd L, Wheeler MT, White SM, Wojcik M, Wolfe LA, Wolfenson Z, Wright CF, Xiao C, Zocche D, Rube...Nature. 2024 Aug;632(8026):832-840. doi: 10.1038/s41586-024-07773-7. Epub 2024 Jul 11. PMID: 38991538
Artificial intelligence in drug development: reshaping the therapeutic landscape. Niazi SK, Mariam Z. Ther Adv Drug Saf. 2025 Feb 24;16:20420986251321704. doi: 10.1177/20420986251321704. eCollection 2025. PMID: 40008227
Artificial Intelligence (AI) Applications in Drug Discovery and Drug Delivery: Revolutionizing Personalized Medicine.Serrano DR, Luciano FC, Anaya BJ, Ongoren B, Kara A, Molina G, Ramirez BI, Sánchez-Guirales SA, Simon JA, Tomietto G, Rapti C, Ruiz HK, Rawat S, Kumar D, Lalatsa A. Pharmaceutics. 2024 Oct 14;16(10):1328. doi: 10.3390/pharmaceutics16101328.PMID: 39458657
