New artificial intelligence (AI) model finds potential treatments from existing medications for countless diseases, including rare ones that currently lack therapies. This AI tool autonomously generates valuable insights, applies them to conditions outside its initial training, and provides explanations for its conclusions. The use of AI could speed up the creation of more targeted therapies with fewer side effects, at a significantly reduced cost compared to traditional drug development approaches.
Over 7,000 rare and often undiagnosed diseases are recognized worldwide.
While each individual disease may affect only a limited number of people, the overall impact is substantial, impacting around 300 million individuals worldwide and imposing a significant healthcare and economic burden.
Despite this, merely 5 to 7 percent of these diseases have an FDA-approved treatment, leaving a majority untreated or inadequately treated.
The challenge of creating new medications is formidable, but a groundbreaking AI tool could facilitate the discovery of new therapies from already available drugs, bringing renewed hope to patients with rare or neglected conditions as well as the healthcare providers treating them.
This AI model, named TxGNN, is pioneering as it has been specifically designed to identify drug candidates for rare diseases and conditions that currently lack treatments.
It has successfully identified potential drug candidates from existing medications for over 17,000 diseases, many of which do not have known treatments. This achievement marks the largest scope of diseases addressed by any AI model to date, with researchers suggesting the potential to apply the model to even more conditions.
This research, published on September 25 in Nature Medicine, was led by a team at Harvard Medical School. The researchers have made the tool accessible at no cost and hope to encourage clinician-scientists to leverage it in their quest for novel therapies, particularly where treatment options are minimal or nonexistent.
“Our goal with this tool is to discover new therapies across a broad range of diseases, but specifically for rare, ultra-rare, and neglected conditions, we believe this model could significantly contribute to reducing health discrepancies,” said lead researcher Marinka Zitnik, assistant professor of biomedical informatics at the Blavatnik Institute at HMS.
“We perceive the potential of AI in lessening the global burden of diseases by identifying new uses for existing drugs, which is a more rapid and cost-effective approach compared to dreamt-up new drug development,” added Zitnik, who is also associated with the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University.
The innovative tool boasts two main functions—one for identifying treatment candidates and their potential side effects, and another that explains the reasoning behind its choices.
In total, the tool identified drug candidates from nearly 8,000 medications (including FDA-approved and those currently in clinical trials) for 17,080 diseases, many of which lack treatment options. It also enabled predictions about which medications could cause side effects or contraindications for specific conditions—an aspect that current drug discovery methodologies generally assess through trial and error in initial clinical trials focused on safety.
Compared to leading AI models for drug repurposing, the new tool is nearly 50 percent more effective, on average, in identifying drug candidates. Additionally, it is 35 percent more precise in forecasting which drugs may pose contraindications.
Benefits of Repurposing Existing Drugs
Repurposing existing medications for new treatments is appealing as it capitalizes on drugs that have undergone extensive study, possess well-established safety profiles, and have received regulatory approval.
Many medicines exhibit multiple effects beyond their specific initial purposes. However, numerous effects remain undiscovered or under-researched during original testing and clinical trials, typically becoming recognized only after widespread use. Nearly 30 percent of FDA-approved medications have obtained at least one additional treatment indication following their initial authorization, with many accruing multiple additional indications over the years.
This current method of drug repurposing is largely random, depending on patient reports of unforeseen beneficial side effects or doctors’ instincts on whether to prescribe medications for unintended conditions, referred to as off-label use.
“Historically, we have relied on chance rather than strategy, restricting drug discovery to existing medications,” Zitnik stated.
The advantages of repurposing drugs extend beyond those without existing treatments, Zitnik observed.
“Even for more prevalent diseases with available treatments, novel drugs can provide alternatives with fewer side effects or replace ineffective medications for specific patients,” she added.
What Sets the New AI Tool Apart from Existing Models
Most existing AI models in drug discovery are trained only on specific diseases or a limited array of conditions. In contrast, this new tool employs a training method that enables it to leverage current data for new predictions by identifying common features among multiple diseases, such as shared genomic variations.
For instance, the AI model detects shared mechanisms of diseases based on common genomic foundations, allowing it to draw parallels from well-researched diseases with known treatments to less understood ones without any therapies.
This capability aligns the AI tool’s reasoning more closely with that of a human clinician, providing creative insights based on access to an extensive pool of pre-existing knowledge and data that the human mind cannot entirely possess or retain.
The model received training on vast datasets encompassing DNA information, cell signaling, gene activity levels, clinical documentation, and more. Researchers validated the model’s performance against 1.2 million patient records while tasking it with identifying drug candidates for diverse diseases.
Furthermore, the researchers queried the AI to predict which patient characteristics could lead to contraindications for the identified drug candidates in certain patient groups.
Another evaluation examined the AI’s capability to pinpoint existing small molecules that could effectively hinder the activity of disease-causing proteins.
In one test, designed to assess the model’s reasoning ability akin to a human clinician, researchers prompted it to find medications for three rare conditions that were not part of its training regimen—a neurodevelopmental disorder, a connective tissue disease, and a rare genetic disorder causing water imbalance.
The team then compared the model’s drug recommendations against existing medical knowledge about the functionality of the suggested medications. In each instance, the model’s suggestions matched current medical understanding.
Additionally, the model not only identified medications for all three conditions but also provided a rationale for its choices, promoting transparency and fostering physician confidence.
Researchers caution that any therapies suggested by the model would need further assessment regarding appropriate dosing and administration timing. However, they emphasize that this unprecedented capability of the new AI model will fast-track drug repurposing in ways previously unachievable. The team is already collaborating with various rare disease foundations to explore potential treatments.
