Scientists from Harvard Medical School are using artificial intelligence (AI) to solve the problem of finding treatments for rare and neglected diseases. They have developed a new method, called TxGNN, that could help identify new drug candidates for more than 7,000 rare and undiagnosed diseases. Details of the method are discussed in a Nature Medicine paper titled, “A foundation model for clinician-centered drug repurposing.”
To date, just 5–7% of rare and neglected diseases have an FDA-approved drug. The rest remain either untreated and undertreated. AI-based applications like TxGNN can help address this challenge by propelling the discovery of new therapies from existing medicines, offering hope for both patients and clinicians.
According to its developers, TxGNN is the first method developed specifically for identifying drug candidates for rare diseases and conditions with no treatments. It identified drugs from existing medicines for more than 17,000 diseases, many of which did not already have treatments, representing the largest number of diseases handled by a single AI model to date. And that’s just a start, the researchers say, as the model could be applied to even more disease types than those used for this study.
So how does it work? The tool is a graph foundation model with two central features. The first identifies treatment candidates along with possible side effects and another feature that explains the rationale for the decision. It was trained on large quantities of data including genomic data, cell signaling information, gene activity data, clinical notes, and more. The researchers tested and refined the model by asking it to perform various tasks and validated its performance on 1.2 million patients records.
The team then asked the model to select drug candidates for various diseases. They also requested that the model make some additional predictions. For example, they asked the tool to predict patient characteristics that would prevent the selected drug candidates from being used to treat particular populations. In another task, the scientists tasked the tool with identifying small molecules that effectively blocked the activity of some proteins implicated in disease-causing pathways and processes.
In another test, the researchers asked the model to identify drugs for three rare conditions that were not included in its training dataset. The conditions were a neurodevelopmental disorder, a connective-tissue disease, and a condition that causes water imbalance. They then compared the model’s therapy suggestions to clinical knowledge about how the suggested drugs work. Results showed that not only did TxGNN’s recommendations align with current medical knowledge, the tool also provided the rationale behind its decision.
The Harvard team has made its tool available for free and they hope clinician-scientists use it. Of course any therapies identified using the model would require additional evaluation for dosing and timing of delivery before being applied to patients. The team is already collaborating with several rare disease foundations to help identify possible treatments for their diseases of interest.
“With this tool we aim to identify new therapies across the disease spectrum but when it comes to rare, ultra rare, and neglected conditions, we foresee this model could help close, or at least narrow, a gap that creates serious health disparities,” said Marinka Zitnik, PhD, lead researcher and an assistant professor of biomedical informatics in the Blavatnik Institute at Harvard Medical School.
