New Uses for Old Drugs

This is the second in a three-part series from Medscape Medical News on the impact of artificial intelligence (AI) on drug discovery and development. Part 1 is about AI’s role in designing speedier, more effective clinical trials. Part 3 reports on AI’s ability to create new proteins from scratch, streamlining the creation of protein-based therapeutics.

Scientists the world over are racing to end Alzheimer’s disease. Over two decades, they’ve conducted hundreds of clinical trials and spent billions in funding. Yet only a handful of Alzheimer’s medications have been approved.

But what if there were drugs already on the market that could help treat or even prevent this devastating disease?

If such drugs exist, geneticist Gyungah R. Jun, PhD, is determined to find them — using AI.

“Using big genetic and molecular data from patients and AI, I can predict everything in silico,” she said, including who is at a risk for Alzheimer’s and how these individuals will respond to existing drugs.

The idea is not far-fetched. About a third of US Food and Drug Administration (FDA)–approved drugs are later found to offer at least one new use. Some have racked up more than 10 post-approval indications.

Historically, these discoveries have happened by chance (think glucagon-like peptide 1 receptor agonists for weight loss), but AI is giving researchers like Jun an edge in the process.

The goal is simple: To quickly uncover new uses for approved drugs or previously studied chemicals, thus skirting the slow pace of new drug development and carving a fresh path to treatments for diseases that desperately need them.

Finding a New Way to Alzheimer’s Treatments

Candidate drugs for Alzheimer’s often target sticky amyloid plaques, and the development process typically ends with trying to find a target population of patients for a clinical trial.

Jun believes that’s backwards.

“My suggestion is we should start with precision medicine,” said Jun, a PhD and associate professor of biomedical genetics at Boston University Chobanian & Avedisian School of Medicine, Boston.

Her strategy focuses on genetics, which accounts for about 80% of Alzheimer’s cases. Jun began developing her approach several years ago, while working on genome-guided Alzheimer’s drug discovery projects for a large pharmaceutical company. She wanted to start at the cellular level inside the brain, identifying specific genes within networks of genes that would be “prime targets” for drugs already in use. But she could only explore drugs produced by her company, so it didn’t work.

When she returned to academia in late 2018, Jun expanded her efforts to include PubChem, a massive public database of drug compounds. Her method relies on machine learning, a type of AI that identifies patterns in large datasets.

Scientists already know many genetic markers associated with Alzheimer’s disease. Jun groups those markers into network-based “subtypes” according to cell type and function. Next, she identifies which gene within a subtype to target. To do that, she uses a graph neural network, a type of machine learning that makes predictions based on relationships between data points (called nodes) linked by lines (edges).

Once a target is in her sights, Jun sets out to find existing drugs that match up. She uses “unsupervised learning” — giving AI unlabeled data (in this case, PubChem) to interpret without any instruction.

Jun’s lab has tested this approach, landing on the APOE gene in the astrocyte subtype as their top contender. They used autopsied brains and astrocytes derived from human pluripotent stem cells to validate their findings.

Results suggest that estradiol, an FDA-approved estrogen replacement therapy, could be effective for treating Alzheimer’s in the APOE genotype. As it happens, at least three clinical trials repurposing estradiol for Alzheimer’s disease have already been completed or are underway. But Jun believes that they need to be reconsidered based on individuals’ genetic risk profiles.

“Clinical trials with patient selection markers will dramatically improve success by up to 90%,” Jun said.

Her hope is that pharmaceutical companies could borrow her technique and design clinical trials around specific genetic subtypes.

“Hopefully it can be quickly applied,” she said, noting that women have a higher risk of developing the disease.

Taking Aim at Rare Diseases

The odds of finding an effective drug are especially slim for patients with rare diseases, those affecting fewer than 200,000 patients. There are more than 7000 rare diseases, and just 5% have any FDA-approved treatment.

“It’s actually shocking,” said Marinka Zitnik, PhD, assistant professor of biomedical informatics at Harvard Medical School, Boston. “As someone who works in AI, that was such an important statistic in terms of how making even small improvements in saving time or decreasing the failure rate in drug discovery could have a huge effect on patients down the line.”

Her lab is investigating the use of existing drugs for diseases with no effective treatments, an effort named zero-shot drug repurposing. “Zero-shot learning” happens when a machine learns to recognize things it’s never seen before. It’s useful for drawing conclusions from large quantities of unlabeled data — in this case, rare, complex, and neglected diseases.

“We developed a model that, for the first time, can nominate a drug for a disease, even though that disease has zero known treatments,” Zitnik said.

Her lab specializes in geometric deep learning, a field of deep learning that can consider the geometry of molecules (as mapped in graph neural networks) to make predictions or generate new designs. They used their model to analyze more than 17,000 diseases, most of them lacking treatments and not well understood.

When the researchers asked the model to find potential drugs for a rare disease, it was able to identify previously hidden relationships to other diseases, including shared pathways, phenotypes, and pathologies. The findings enabled the model to identify existing drugs that could be effective. Its predictions were consistent with off-label prescriptions made by clinicians in a large health system, and because the model explains its reasoning, it may give clinicians greater confidence in its predictions.

In a case study, the model identified several approved drugs that may have positive therapeutic effects on Wilson’s disease, a rare disorder associated with excessive copper accumulating in the body that can trigger cirrhosis. Patients often develop intolerance to a common existing treatment, making long-term care challenging. But the model surfaced a promising potential therapeutic candidate, which previous studies suggest may remove iron from the liver. The lab is doing follow-up biological tests to confirm its effects on copper.

Predicting How a Disease Will Respond to a Treatment

The prospect of repurposing existing drugs for rare diseases “has a lot of low-hanging fruit,” said Shantanu Singh, PhD, a principal investigator at the Carpenter –Singh Lab at the Broad Institute of MIT and Harvard. Rare diseases are generally less studied, despite FDA support. “It’s hard to incentivize pharma companies to develop a drug that benefits just 1000 or 10,000 people,” Singh said. That leaves ample opportunity for researchers in academia and biotech to step in.

The Carpenter–Singh Lab uses AI to search for insights in cellular images, yet another approach that could help realize the potential of existing drugs.

When biologists look under a microscope, they rely heavily on their own knowledge to understand the image and detect abnormalities or changes, said Singh. This took him by surprise. In his previous role computing for an auto manufacturer, “my work was completely focused on this idea that you can make computers see, just like humans can see,” he said.

The auto company had been developing an early version of a self-driving car. Singh used data from satellite images to build predictive models that could recognize objects and people. It’s a type of AI known as computer vision, which lets computers recognize and glean information from images and video.

At the Carpenter–Singh Lab, automated robot microscopes take pictures of cells treated with different chemicals and compounds. AI and machine learning help “describe” what’s happening inside the cells.

“Just by looking at similarities between the patterns of different chemicals’ effects on cells, you can tell a lot,” Singh said.

Rare diseases are often genetic, and many feature mutations in more than one gene. But some — roughly 3000 or so — are thought to involve only one gene.

By introducing a single mutation into a cell line, researchers can create simplified models of these single-gene diseases. From there, they can analyze how the cell responds to different drugs, looking for any that revert the disease phenotype back to a healthy state.

“You’re able to now, at scale, across the 3000 diseases and across the [20,000-plus] drugs that are on the market, find potential drug-disease combinations,” Singh said.

It’s the technique that launched biotech company Recursion. (Anne Carpenter, PhD, senior director of Broad’s Imaging Platform, is a scientific advisor.) The company’s founders used cell imaging to identify a repurposed candidate for cerebral cavernous malformation (CCM), a rare neurovascular disease. Recursion’s repurposing efforts draw on the company’s database of about 17,000 known molecules, including approved drugs, and chemicals that other organizations attempted to develop but later shelved.

Available drugs only treat symptoms of CCM, such as seizures and bleeding in the brain. Cell imaging pointed to the effectiveness of a small molecule known as a superoxide scavenger. The researchers were surprised, but when they found previous research linking a CCM gene to oxidative stress, they moved forward with animal studies. Now, that molecule is in phase 2 trials, with data expected later this year.

For the approximately 350,000 people with CCM, a new treatment could become available soon. The finding may also have implications beyond CCM, which causes leaky blood vessels, a symptom seen in many other diseases, including multiple sclerosis and sepsis.

Bringing Repurposed Drugs to Patients

About one in five scripts are for off-label uses, which lack FDA approval. Generally, doctors can make the determination. They are unlikely to write a script without a disease model, which some rare diseases lack, according to Julie Owen, director of chemistry at Recursion.

“A lot of academic institutions are offering up suggestions for repurposing based on a biochemical assay or protein binding assay,” said Owen. “It’s just not sufficient. It’s not going to give confidence to clinicians to give a drug to a patient as an off-label use.”

If compounds are well understood or have FDA approval for other indications, their toxicity and possible side effects have been assessed. Knowing that a drug can be safely administered to healthy patients “can save you the phase 1 trials,” said Owen. “Potentially, you can get into phase 2 trials, depending on the disease model evidence that you have.”

Three repurposed candidates for rare diseases (including CCM) identified by Recursion are currently in phase 2 studies. Repurposed candidates for cancer indications are also in phase 2 and preclinical trials.

If these efforts pan out, thousands of patients could be able to access more effective treatments for conditions that pharmaceutical companies might otherwise have ignored. And the possibilities for AI-enhanced drug repurposing are only expanding.

The “pooled version” of the Carpenter–Singh Lab’s platform, for instance, uses a strategy known as optical barcoding, which tags genetic mutations in each cell model. This allows researchers to put thousands of cells representing hundreds of different diseases into each “well” (mini test tube) in a multiwell plate. Then, they can add a different compound into each well to study how thousands of compounds interact with hundreds of diseases simultaneously. Machine learning algorithms analyze images of the cells to “see” how they change.

“That can really scale it up,” Singh said. “If you’re interested in a specific disease, we don’t know what’s going to happen, but you’re increasing your chances of finding something.”

Continue on to part 3 of AI’s Drug Revolution.