AI Uses Zero-Shot Learning to Find Existing Drugs for Treating Rare Diseases

A groundbreaking drug-repurposing AI model could bring new hope to doctors and patients trying to treat diseases with limited or no existing treatment options. Called TxGNN, this zero-shot tool helps doctors find new uses for existing drugs for conditions that might otherwise go untreated.

The study, recently published in Nature Medicine and led by scientists from Harvard University, could reduce the time and cost for drug development—delivering effective treatment to patients much more quickly.

“With this tool, we aim to identify new therapies across the disease spectrum but when it comes to rare, ultrarare, and neglected conditions, we foresee this model could help close, or at least narrow, a gap that creates serious health disparities,” study lead Marinka Zitnik, an assistant professor of biomedical informatics in the Blavatnik Institute at Harvard Medical School, said in a recent news post. 

Globally, more than 300M people are affected by over 7,000 rare or undiagnosed diseases. Of these rare diseases, around 7% have an FDA-approved drug treatment. This means that many patients are waiting and hoping for new therapies. 

This new AI tool addresses a limitation of most current models, which rely on known drugs for similar diseases. These models struggle with rare and poorly understood conditions due to a lack of data.

TxGNN overcomes this issue by using graph neural networks (GNNs) to analyze and identify complex relationships and patterns in large medical data sets, which include information on diseases, drugs, and proteins. 

The researchers trained and fine-tuned the model using NVIDIA V100 Tensor Core GPUs and NVIDIA H100 Tensor Core GPUs. According to Zitnik, the GPUs were critical in processing the large medical knowledge graph spanning 17,080 diseases and almost 8,000 drugs. 

TxGNN relies on graph neural networks (GNN) to reason over complex biological data and generate transparent explanations for experts to review insights into its predictions. By analyzing these underlying connections, the AI model can understand and predict how a drug could influence a specific condition.

In testing, the AI model improved treatment predictions by up to 19% without being trained on the specific disease. It also performed better than existing models at predicting contraindications, which are situations when a drug shouldn’t be used.

Its treatment suggestions also matched medications that doctors often prescribe even when they aren’t approved for a specific condition. 

Visit the TxGNN Explorer to learn more and experience the visual interface. 

Read the full story from Harvard Medical School.