A case study from drug repurposing for rare diseases

In 2010, Dr David Fajgenbaum became critically ill with Castleman Disease (CD) while at medical school. Told he had no treatment options, he couldn’t wait for a new treatment that would cost billions of dollars and at least 10 years to develop. Repurposing an existing drug was his only hope.

Dr Fajgenbaum and his team set about researching options and found a drug (sirolimus) that had previously been approved as an immunosuppressant but never been used for CD. As a result of this repurposed drug, David has been in remission for more than 11 years.

It’s an inspiring story, and Dr Fajgenbaum later founded EveryCure, using AI and ML methods to identify potential repurposing candidates for other rare diseases.

Open Access and Open Source for AI

But what about access to data? The most obvious source of drug-target and drug-pathway relationships is in biomedical literature. Early-stage companies building AI applications frequently use open access (OA) articles since these are easier to acquire. However, the challenge is addressing concerns that any predictions are less reliable.

This preprint by the EveryCure team (https://www.medrxiv.org/content/10.1101/2024.12.31.24319817v1.full.pdf) with contributions from Nascent Studio, explores this question.

It shows that OA articles generated more than 50% of potential repurposing candidates for rare diseases and subscription articles 45%.

The analysis also shows that although a larger proportion of subscription articles contained relationships and potential repurposing candidates, this is due to the number relationships (edges) supported by single references in the knowledge graph used. It does not indicate an intrinsic difference in quality between OA and subscription articles and thus predictions made from OA content are just as effective.

Are you exploring data sources for your model training? Contact us and we can guide you to the most effective sources.