New uses for existing medicines in rare childhood epilepsies
AI can already suggest which approved medicines might help a rare disease. Proving it needs real patient data, and that is where most projects stop. Ithaca combines a working AI prototype with the legal know-how to reach that data lawfully, with consent.
Dravet syndrome. A severe genetic epilepsy that starts in infancy. A handful of medicines are approved; many children still have frequent seizures.
SCN1A gene. Most children with Dravet carry a change in this sodium-channel gene. The model uses links like this to reason about which medicines could act on the disease.
Known treatments. Approved medicines such as stiripentol and cannabidiol. We hide links like these and check whether the model can find them again.
Safety filter. A naive model could suggest sodium-channel blockers because they act on SCN1A, yet they can make Dravet worse. Our filter removes medicines known to harm a condition before any ranking.
Algorithms are plentiful. Lawful access to patient data is not.
Few treatments
There are thousands of rare diseases and most have no approved treatment. Repurposing an approved medicine is faster and cheaper than developing a new one, because its safety is already known.
Algorithms are becoming open
Well-funded players such as the non-profit Every Cure publish their repurposing models openly. Ranking candidates is no longer the scarce part.
Lawful data access
Testing a candidate needs real patient records, often of children and across borders. New rules such as the EU's European Health Data Space open a legal route for research use, but using it takes legal expertise. That is our starting point.
Four studies behind the plan
Our qualitative view, not measured data. Tap a company.
Ithaca. We start from the scarce part: consent, data-protection compliance and partnerships that let candidates be tested on real-world data. Our own models are good enough to produce candidates; the moat is lawful access and trust.
Every Cure. US non-profit with publicly funded, open-source repurposing models. Shows that ranking candidates is becoming a free commodity.
Healx. UK company using AI to repurpose drugs for rare diseases, with programmes in clinical trials. Proves the model works and that rare diseases attract investment.
Recursion. Large AI drug-discovery platform with very high spending. Shows how capital-intensive the "build everything" route is.
BenevolentAI. Once a leading AI drug-discovery company; left the public markets in 2025. A lesson that algorithms alone are not a business.
We chose rare genetic developmental and epileptic encephalopathies: the gene is known, the outcome (seizure frequency) is measurable, off-label use is common and patient organisations are active. Select a condition:
Dravet syndrome. Starts in the first year of life, mostly caused by SCN1A. Some approved treatments exist, which lets us test whether the model rediscovers them. Sodium-channel blockers can worsen it, a key test for our safety filter.
CDKL5 deficiency disorder. Early-onset seizures and severe developmental delay. Very few targeted treatments. In the open graph it is merged into a broader group, so better data matters.
SCN2A-related epilepsy. The same gene change can either over- or under-activate the channel, so the right medicine depends on the variant. A clear case for patient-level data.
KCNQ2 encephalopathy. Seizures in the first days of life. Potassium-channel biology gives clear drug targets for repurposing hypotheses.
GRIN-related disorders. Changes in NMDA-receptor genes. Some approved medicines act on these receptors, which makes them natural repurposing candidates.
KCNT1-related epilepsy. Severe infant epilepsy where off-label medicines have been tried with mixed results; real-world data could show who benefits.
Risk control: if fewer than 3 of the first 10 to 15 organisations we interview confirm interest, we move to our fallback group, autoinflammatory diseases.
-
Now · open data only
No patient data
Public research sources only: Orphanet, Mondo, HPO, Open Targets, ChEMBL, ClinicalTrials.gov and adverse-event reports. Sources limited to non-commercial use stay out of the product.
-
Next · academic partner
Consent-based pilot
A study with a university and a patient organisation: tiered consent designed for parents of children, ethics approval, a data-sharing agreement, analysis only inside a secure research environment.
-
2029+ · European Health Data Space
Scale across countries
EU rules on secondary use of health data start to apply from 2029, creating a regulated route to data across member states. We aim to be ready to use it on day one.
Principle: data stays in the country where it was collected and under that country's law. We never pool data across jurisdictions without a lawful basis.
-
Week 1
Data review
Explored PrimeKG, an open graph of 129,375 genes, diseases, drugs and phenotypes with 4 million links. Found that it has no contraindications for Dravet, so a safety filter was needed first.
-
Week 2
Safety filter and first model
Added a clinical contraindication filter based on the international Dravet consensus. A first model found known treatments well but mostly repeated drugs already in use.
-
Week 3
Graph neural network
Switched to a graph neural network and a harder test that mimics a rare disease with no known drugs. It beat both baselines.
Can the model find a treatment it was never shown?
We hide known treatments and ask the model to rank all 7,957 medicines in the graph. The harder test hides every treatment for one in ten diseases, which is what a rare disease with no approved drug looks like.
Safety filter in action: Dravet syndrome
Part of the model's raw list for Dravet. Switch the filter off to see what it removes. These are research hypotheses, not recommendations.
| Rank | Medicine | Note |
|---|---|---|
| 6 | Lorazepam | benzodiazepine, used for seizure emergencies |
| 8 | Diazepam | benzodiazepine, used for seizure emergencies |
| 10 | Clonazepam | benzodiazepine, used in some epilepsies |
| 30 | Vigabatrin | removed may worsen myoclonic seizures |
| 77 | Carbamazepine | removed sodium-channel blocker, can worsen Dravet |
| 83 | Lamotrigine | removed sodium-channel blocker, can worsen Dravet |
Lists are still noisy and some top candidates are implausible; every list must be reviewed by clinicians. Source of the filter: Wirrell et al., Epilepsia 2022, international consensus on Dravet syndrome.
Roadmap
Done
The groundwork an investor would ask about is in place.
- Market review: the gap is lawful data access, not algorithms
- Focus disease group chosen, with a fallback
- Data-and-law plan for consent, privacy and cross-border rules
- Working prototype with a safety filter, tested on unseen diseases
- Bilingual site and a list of first partners to contact
Prove clinical value
Goal: clinicians and families confirm the problem and the approach.
- Candidate lists reviewed by paediatric neurologists
- 10 to 15 interviews with patient organisations and clinics
- A bioinformatician in rare-disease genomics joins as adviser
- Model upgrade: clinical features, larger training runs
- First letters of intent from partners
First real-world evidence
Goal: test our best hypotheses on real data, lawfully.
- Company incorporated in a jurisdiction that suits investors and data partners
- Ethics-approved study with an academic data partner
- Tiered-consent pilot with a patient organisation
- First paid data-and-compliance projects
Build the data layer
Goal: recurring revenue from lawful research access.
- Consent-based access to real-world data in several countries
- Researchers and companies pay for access to approved studies
- Strongest candidates move towards clinical studies with partners
Scale
Goal: become the trusted route to rare-disease data in Europe and beyond.
- EU rules on secondary use of health data apply from 2029
- Extend to other rare-disease groups
- Own repurposing programmes, licensed to companies that can bring them to patients
From a service to a data layer
-
Stage 1 · Early revenue
Data and compliance service
Help repurposing teams, biotech companies and patient organisations design consent, data protection and data-access applications for rare-disease studies, together with candidate lists from our models.
-
Stage 2 · Core business
Lawful patient-data layer
Consent-based access to real-world data for rare childhood epilepsies, built with clinics, registries and patient organisations in several countries. Researchers and companies pay for access to ethics-approved studies.
-
Stage 3 · Long term
Own repurposing programmes
The strongest candidates, confirmed on real-world data, go into studies with clinical partners and are licensed to companies that can bring them to patients.
Legal expertise first, science around it
- Founder
Law and data governance
Lawyer specialising in medical data, consent and regulation. Leads partnerships, data agreements and compliance.
- In house
AI and engineering
Models, data pipeline and safety filter built in house. Code and technical reports are available to investors on request.
- Joining
Science and clinical advisers
We are recruiting a bioinformatician in rare-disease genomics, paediatric neurologists as clinical advisers, and university AI researchers as academic partners.
What we are looking for
We are talking to angel investors, pre-seed funds, accelerators and grant programmes about a first round to fund the next 12 months.
Science
A bioinformatician and a clinical advisory board to validate candidates.
Legal and data setup
Company formation, ethics and data-protection setup, the first data-sharing agreement.
First pilot
Computing, data-access fees and the first real-world study with an academic partner.
If algorithms are open, what is the moat?
Consents, ethics approvals, data-sharing agreements and the trust of patient organisations take years to build and are hard to copy. Models are replaceable; lawful access is not.
Why rare childhood epilepsies?
The causal gene is usually known, seizure frequency is a measurable outcome, off-label use is already common, and families are organised and motivated. It is a narrow market to win first, with a clear path to other rare diseases.
Do you use patient data today?
No. We use open research data only. We will work with patient data only under consent, ethics approval and a data-sharing agreement, inside a secure research environment.
Is there a company?
Not yet. We will incorporate in a jurisdiction that suits investors and data partners before taking investment. Nothing on this site is an offer of securities.
How will you make money?
First from data and compliance projects for research teams and companies, then from paid access to the lawful data layer, and in the long term from licensing our own repurposing programmes.
Who we would like to hear from
Investors and accelerators
Angels, pre-seed funds, accelerators and grant programmes interested in rare diseases, health data or AI for drug discovery.
Clinicians and researchers
Paediatric neurologists, epilepsy researchers and bioinformaticians who could review candidate lists or join as advisers.
Patient organisations and data partners
Charities, family groups, universities and research data programmes in any country interested in an ethics-approved study.
What this project is and is not
- Every candidate we produce is a research hypothesis. None is medical advice, and none should change a child's treatment without their own clinician.
- Conclusions are checked by qualified clinicians and scientists before they are shared.
- We use only open research data today. We will work with patient data only under consent, ethics approval and a data-sharing agreement, inside a secure research environment.
- We are not a healthcare provider and do not sell or offer treatments.
Privacy, sources and terms
Privacy
This site sets no cookies, runs no analytics or tracking and loads nothing from other websites. It remembers your language choice in your own browser only. Our hosting provider, Cloudflare, processes technical data such as IP addresses to deliver the site and protect it from abuse (Cloudflare privacy policy). If you email us, we use your address and message only to reply and to discuss collaboration, and we delete the correspondence on request. Privacy questions: ithaca-rare@outlook.com.
Sources
Figures are our own results on the public PrimeKG dataset: Chandak P, Huang K, Zitnik M. Building a knowledge graph to enable precision medicine. Scientific Data 10, 67 (2023). Model: Schlichtkrull M et al. Modeling relational data with graph convolutional networks. ESWC (2018). Safety filter: Wirrell EC et al. Epilepsia (2022). Company descriptions are our own summary of public information. Medicines are named for scientific discussion only; trademarks belong to their owners.
Terms
Content is provided for general information about a research project. It may contain errors and may change without notice. It is not medical, legal or investment advice and is not an offer of any product, service or securities. Do not send us medical records or other health data about yourself or others.