Loading ...

Old drugs, new ideas. That’s one of the core drivers of using artificial intelligence (AI) for drug development.

Researchers using complex AI-fueled methods are now taking aim at what has become a brick wall for precision-focused Alzheimer’s drug development: the complex heterogeneity of both patients and the disease.

A new paper in the journal Cell took a novel multilayer evidence approach to identify existing — and possibly surprising — drugs that might be repurposed for AD treatment. Using cellular mechanism and patient population heterogeneity as foundational guidance, University of California San Francisco (UCSF) researchers hypothesized that different cells respond differently to disease pathology and sought to identify candidates that therapeutically changed the behavior of multiple types of cells.

They choreographed an arsenal of research techniques from basic statistics to mouse lab work supercharged by AI. The resulting funnel of outcomes pointed with striking consistency toward two cancer drugs: letrozole and irinotecan.

It’s not the first time drugs have reversed memory problems and disease pathology in mice, but the approach is notable for its combination of computational biology, medical record data, and animal models.

Functional genomics expert Caghan Kizil, PhD, of Columbia University, New York City, called the results “compelling.” Feixiong Cheng, PhD, director of the Cleveland Clinic Genome Center, Cleveland, called the findings “a milestone” and “an amazing achievement.”

“This good example sets the standard for next studies about how computational biology, big data, clinical data, translational neuroscience, and animal models can come in to create faster or smarter therapeutics,” Kizil said.

How They Did It

The researchers used human single-cell data sets as a starting point to develop a gene expression profile.

“We think since glial cell and neuronal cells are so different, they probably require different medication through different mechanisms to change their disease state into the healthy state,” said co-senior author Marina Sirota, PhD, the interim director of the UCSF Bakar Computational Health Sciences Institute. “And that’s why we really need single-cell data to enable to do that because we need to systematically characterize the molecular changes of these different cell types under disease condition.”

They then identified 82 drug repurposing candidates from a database of 1300 FDA-approved or previously studied options that could affect the profile of at least one cell type, and 25 candidates could affect more than one cell type.

Next, they leveraged medical record data to further refine the candidate list to 10, and a further deep dive revealed five options that showed patients who took at least one of the drugs were less likely to develop AD than people with similar disease who did not take the drug.

The combination therapy approach of letrozole and irinotecan targeted all five major brain cell types with the potential to change gene expression signatures for AD.

The next step was testing the drugs in mice. In the lab of co-senior author Yadong Huang, MD, PhD, the team used an aggressive disease model that involves both amyloid deposits and tau tangles. The combination therapy outperformed single drug therapy, reducing pathology and restoring memory. Analysis of mice brain samples confirmed the restoration of normal single-cell expression.

Kizil, who was not involved in the study and is also working to broaden AD research such as in the area of vasculature, lauded the UCSF team’s multilayer approach. 

“They start from an idea of correcting the dysregulated gene networks in a cell-type specific manner because if a single protein goes awry in disease, it affects a lot of things. So, we may not be able to correct by just correcting the function of that one gene or protein,” Kizil said. “We diagnose people when they are 60, 70, 80 years old, and the pathology might kick in much, much earlier — decades ago. So, what they look at in this paper is what changes in a larger network scale and whether there are any drugs that could partially restore this.”

A Mystery Mechanism

Casting a wide net means sacrificing precision in some areas. The treatment mechanisms are unclear. There’s a gap between the mouse model and the cancer patients. The stage of disease needs to be examined, too, and of course, the mechanism, Kizil noted.

And the data may be confounding despite the UCSF team’s “solid” methods and efforts to overcome such effects because of survival rates of breast cancer and colon cancer patients, noted Cheng, who was not involved in the study. Also, the mouse model didn’t show strong effects from single drug therapy, which contradicts the human data from medical records.

The side effects of the chemotherapy drugs pose challenges, particularly among older AD patients who may not be able to tolerate them, Cheng added.

“We need more safe, repurposable drugs or combination therapies beyond cytotoxic chemotherapy drugs to be tested in aged individuals with Alzheimer’s disease,” he said.

“We’re just scratching the surface, but there is something worth exploring here,” Sirota acknowledged.

Yadong is optimistic that letrozole and irinotecan dosage can be modified for AD compared to its cancer applications. Those tests are currently underway in the mouse models at his lab in the Gladstone Institutes Center for Translational Advancement.

If the clinical trial gets underway, letrozole and irinotecan will join a large group of already FDA-approved drugs that show promise in treating AD. One third of AD pipeline drugs are repurposable drugs among 182 active clinical trials among phases 1-4, Cheng noted, adding that the benefit of the repurpose approach is that it’s cost effective.

AI Joins the Dream Team

It’s a common refrain in the AD drug research space: Accommodating failures brings new perspectives.

Past learnings have fueled the multidisciplinary team approach, such as the combined expertise of the UCSF bioinformatics and mouse labs. Machine learning is now well underway in the AD space, and though no one is saying the technology places the field on the brink of a breakthrough, researchers appear poised to begin learning from new types of failures.

Solita said a gamechanger for their latest study was having “a unicorn” on the team.

Lead author Yaqiao Li, PhD, worked on every phase of the study, from locating the then-hard-to-find single-cell data sets, analyzing the data, aiding in the mouse lab and conducting validating single-cell analysis, to now developing a generative AI model for follow-up work. A trained mouse biologist, Li’s entire bioinformatics doctoral study period coincided with the research. She now specializes in single-cell RNA-sequencing analysis, electronic health record analysis, and wet lab work.

Li said that her biology background impressed upon her how common it is for experiments to fail. She likened the study’s methodology to a mock clinical trial.

“The really critical point was using the electronic medical record to kind of borrow clinical evidence to point us to a few drugs that seemed to show beneficial effects in humans,” Li said. “And from there, I felt like the chance of failure was really, really low because you have all this supporting evidence already.”

Another key to the study’s design, Li said, was using a type of mouse whose amyloid and tau metrics could be well matched to their human data.

“We wanted to mimic that. That's why we chose to have these transgenic animals and a very aggressive cross,” Li said. “They have pathology very early in their lifespan. Usually, a mouse can live to 2 years old, but these mice start to die around 8-10 months old.”

Solita said that forthcoming generative AI models will generate their own hypotheses and analyze clinical notes.

“We are having a shift in momentum towards AI and more systems biology and computational aspects,” Kizil said. “Backing this up with strong biology, we can get to smarter therapeutics that address a lot of different aspects of the disease rather than just one. I think we need to systemically treat people rather than correcting only one aspect. The authors show very beautifully one way to do that in this new paper.”