Advances in biological AI models have shown promise when it comes formulating new drugs. But can the technology also find ways to treat rare diseases with existing medications?
That’s what researchers at Harvard Medical School set out to learn with a new foundation model designed to do just that. The TxGNN model, described in a recent paper in Nature, is purportedly the first AI system that seeks to pinpoint already approved drugs that might specifically treat rare diseases for which there are currently no treatments on the market.
These diseases might be rare on their own, but taken together, the more than 7,000 conditions classified as such affect around 300 million people worldwide, according to The Lancet. And only 5–7% of these ailments currently have drugs approved by the Food and Drug Administration (FDA). Meanwhile, nearly one-third of drugs approved by the FDA end up with more than one approved use later on, often as many as 10 or more, the authors write.
The problem, according to the team, is that discovering these new uses tends to be “a serendipitous and opportunistic endeavor”—doctors more or less stumble upon them in the course of working with patients.
“Predicting the efficacy of all drugs against all diseases would enable us to select drugs with fewer side effects, design more effective treatments targeting multiple points in a disease pathway, and systematically repurpose existing drugs for new therapeutic use,” the authors wrote.
Breakthroughs like Google’s AlphaFold protein prediction model, which won the Nobel Prize for chemistry this month, have led to a boom in the application of generative AI to the drug discovery pipeline. Plenty of companies are also building models meant to tackle drug repurposing, according to GlobalData.
Keep up with the innovative tech transforming business
Tech Brew keeps business leaders up-to-date on the latest innovations, automation advances, policy shifts, and more, so they can make informed decisions about tech.
The team behind TxGNN claims that it’s nearly 50% better on average than leading repurposing models when it comes to identifying drug candidates and 35% better at predicting contraindications, or reasons why a patient should not take a given drug.
The model is also broader in scope than many AI drug models on the market, according to a Harvard post announcing the research. While many of these models tend to focus on one specific disease or small groupings, TxGNN is designed to more comprehensively draw parallels between “mechanisms based on genomic underpinnings” shared by rare diseases and more well-documented ones, per the announcement.
The tool was trained on public repositories of data like DNA information and clinical notes and validated on almost 1.3 million deidentified patient records from Mount Sinai Hospital in New York. The model also includes an explainer component that seeks to elucidate the step-by-step reasoning behind the predictions.
The team has made the tool available for free for other scientists to use in an effort to spur discoveries.
“This is precisely where we see the promise of AI in reducing the global disease burden, in finding new uses for existing drugs, which is also a faster and more cost-effective way to develop therapies than designing new drugs from scratch,” said Marinka Zitnik, an author on the paper and an assistant professor of biomedical informatics at Harvard Medical School’s Blavatnik Institute.