Using AI to repurpose existing drugs to treat rare diseases — Harvard Gazette

There are over 7,000 rare and undiagnosed diseases worldwide.

Although each of these diseases affects only a small number of people, they collectively cause staggering human and economic damage, affecting approximately 300 million people worldwide. Yet they remain largely untreated or undertreated, with only 5 to 7 percent of these diseases having an FDA-approved drug.

Developing new drugs is a daunting challenge, but a new tool based on artificial intelligence can accelerate the discovery of new therapies based on existing drugs and bring hope to patients with rare and neglected diseases and to the physicians who treat them.

The AI ​​model, called TxGNN, is the first specifically designed to identify drug candidates for rare diseases and conditions for which there is no treatment.

Drug candidates have been identified from existing drugs for more than 17,000 diseases, many of which have no treatment yet. This is the largest number of diseases that a single AI model can handle to date. The researchers point out that the model could be applied to even more diseases than the 17,000 it handled in the first experiments.

The work, described in Nature Medicine on Wednesday, was led by scientists at Harvard Medical School.. The researchers are making the tool available free of charge and would like to encourage clinical scientists to use it in the search for new therapies, especially for diseases for which there are no or limited treatment options.

“We hope to use this tool to identify new therapies across the entire spectrum of diseases. But when it comes to rare, ultra-rare and neglected diseases, we believe this model could help close or at least narrow a gap that leads to serious inequalities in health care,” said lead researcher Marinka Zitnik, assistant professor of biomedical informatics at the Blavatnik Institute at HMS.

“This is where we see the potential of AI in reducing the global burden of disease and finding new uses for existing drugs. In addition, developing therapies is faster and more cost-effective than developing new drugs from scratch,” added Zitnik, who is an associate professor at Harvard University's Kempner Institute for the Study of Natural and Artificial Intelligence.

The new tool has two key features: one identifies treatment candidates along with possible side effects, and another explains the reasons for the decision.

In total, the tool identified drug candidates for 17,080 diseases, including conditions for which there are no treatments, from nearly 8,000 drugs (both FDA-approved and experimental ones currently in clinical trials). It also predicted which drugs would have side effects and contraindications for certain conditions – something the current approach to drug discovery mostly determines through trial and error during early clinical trials that focus on safety.

Compared to leading AI models for drug repurposing, the new tool was, on average, nearly 50 percent better at identifying drug candidates and 35 percent more accurate at predicting which drugs would have contraindications.

Advantages of using already approved medicines

Repurposing existing drugs is a promising way to develop new treatments because it is based on drugs that have been studied, have well-established safety profiles, and have passed regulatory approval processes.

Most drugs have multiple effects beyond the specific targets for which they were originally developed and approved. However, many of these effects go undetected during initial testing, clinical trials, and reviews and are poorly studied. They only become apparent after years of use by millions of people. In fact, nearly 30 percent of FDA-approved drugs have received at least one additional treatment indication after initial approval, and many have received dozens of additional treatment indications over the years.

This approach to drug repurposing is haphazard at best. It relies on patient reports of unexpected beneficial side effects or on doctors' intuition about whether to use a drug for a condition for which it is not intended. This practice is called off-label use.

“We rely on luck and chance rather than strategy. This limits drug discovery to diseases for which there are already drugs,” says Zitnik.

The benefits of drug repurposing extend beyond incurable diseases, Zitnik noted.

“Even for more common diseases for which there are already approved treatments, new drugs could offer alternatives with fewer side effects or replace drugs that are ineffective in certain patients,” she said.

What makes the new AI tool better than existing models

Most current AI models used for drug discovery are trained on a single disease or a handful of conditions. Rather than focusing on specific diseases, the new tool has been trained to make new predictions using existing data. It does this by identifying common features across multiple diseases, such as shared genomic aberrations. For example, the AI ​​model identifies common disease mechanisms based on shared genomic underpinnings, allowing it to infer from a well-understood disease with known treatments to a poorly understood disease with no treatments.

This capability, the research team said, brings the AI ​​tool closer to the kind of reasoning a human clinician could use to generate new ideas if they had access to all the pre-existing knowledge and raw data that the AI ​​model has but that the human brain cannot possibly access or store.

The tool was trained on massive amounts of data, including DNA information, cell signals, gene activity levels, clinical notes, and more. The researchers tested and refined the model by having it perform various tasks. Finally, the tool's performance was validated using 1.2 million patient records and asked to identify drug candidates for various diseases.

The researchers also asked the tool to predict which patient characteristics would make the identified drug candidates contraindicated for certain patient groups.

Another task was to ask the tool to identify existing small molecules that could effectively block the activity of specific proteins involved in disease-causing pathways and processes.

In a test designed to measure the model's ability to reason like a human clinician, researchers asked the model to find drugs for three rare conditions that it had not been exposed to as part of its training: a neurodevelopmental disorder, a connective tissue disease and a rare genetic disorder that causes water imbalance.

The researchers then compared the model's drug therapy recommendations with current medical knowledge about how the suggested drugs work. In each example, the tool's recommendations were consistent with current medical knowledge.

In addition, the model not only identified medications for all three diseases, but also provided the reasons for its decision. This explanatory function ensures transparency and can increase doctors' trust.

The researchers point out that any therapies identified by the model would require additional research regarding dosing and timing of administration. However, they add that with this unprecedented capacity, the new AI model would accelerate drug repurposing in ways that were not previously possible. The team is already working with several rare disease foundations to identify potential treatments.

Co-authors included Kexin Huang, Payal Chandak, Qianwen Wang, Shreyas Havaldar, Akhil Vaid, Jure Leskovec, Girish N. Nadkarni, Benjamin S. Glicksberg, and Nils Gehlenborg.

This work was supported by the National Science Foundation CAREER Award (grant 2339524), the National Institutes of Health (grant R01-HD108794), the U.S. Department of Defense (grant FA8702-15-D-0001), Amazon Faculty Research, the Google Research Scholar Program, AstraZeneca Research, the Roche Alliance with Distinguished Scientists, the Sanofi iDEA-TECH Award, Pfizer Research, the Chan Zuckerberg Initiative, the John and Virginia Kaneb Fellowship at HMS, the Biswas Family Foundation Transformative Computational Biology Grant in partnership with the Milken Institute, the HMS Dean's Innovation Awards for the Use of Artificial Intelligence, the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University, and the Dr. Susanne E. Churchill Summer Institute in Biomedical Informatics at HMS.