New AI Model Predicts How Cells Age

Every part of your body is aging every day. Over decades, the cells in your heart, brain, and lungs gradually change in ways that can boost your risk of heart disease, dementia, and other age-related conditions.

Scientists have long tried to map those changes, but most studies examine cells one snapshot at a time, missing the bigger picture of how aging actively unfolds.

Now, researchers at Gladstone Institutes, working with NVIDIA, have built an artificial intelligence (AI) model that tracks the molecular details of exactly how cells throughout the body change over time.

Ultimately, that could let them identify drugs or other interventions to slow or reverse the cellular changes driving age-related diseases, including heart disease and dementia.

“Aging progresses across very long timelapses that are essentially impossible to model in the lab in human cells,” says Gladstone Investigator Christina Theodoris, MD, PhD, whose team developed the model. “What makes this model exciting is that we can use it to identify the most promising anti-aging targets before committing to the kind of costly, long-term studies that aging research requires.”

Christina Theodoris consults with Javier Gόmez Ortega, first author of the study and a bioinformatics fellow in her lab.

To assess its performance, Theodoris and her collaborators used the AI model to analyze heart cells and tasked it with predicting genes that either accelerated or slowed normal aging. When the researchers tested the effects of those genes in human cells, the predictions held true: activating the pro-aging genes caused the cells to display hallmarks of aging, such as irregular beating and gene activity associated with inflammation.

Building a Model to Span Decades

Previously, Theodoris’s lab developed Geneformer, a foundational AI model that learned how genes work together by studying millions of human cells across health and disease. The tool helps predict how individual cells respond to genetic changes, but only at a single point in time.

The new temporal AI model, called MaxToki—named after a Japanese bullet train whose name is a homonym for “time” in Japanese—builds off that approach, adding the new element of following tissues throughout aging.

Theodoris’s team trained MaxToki using detailed data on nearly 170 million cells from healthy humans ranging in age from birth to over 90 years old. Rather than treating each cell as an isolated snapshot, the model learned to predict how the gene activity of any given cell type changed with aging.

A major challenge in building MaxToki was sheer computational scale. Training the model required processing the equivalent of nearly one trillion pieces of genetic data—an amount approaching what it takes to train the large language models that power tools like ChatGPT.

To make this feasible for an academic lab, Gladstone collaborated with NVIDIA to accelerate key parts of the underlying code to run more efficiently on specialized NVIDIA AI infrastructure and accelerated computing. Gladstone leveraged NVIDIA BioNeMo and CUDA-X libraries to optimize training throughput and efficiency.

Javier Gόmez Ortega is first author of the study that tested whether the MaxToki AI model would recognize something it had never been taught: the accelerated aging that accompanies some diseases.

“The sheer computational scale of training models on nearly a trillion genetic tokens is a major hurdle for even the most advanced labs, and there is an urgent need for accelerated computing to handle these massive temporal datasets,” says Anthony Costa, director of Digital Biology at NVIDIA. “By leveraging NVIDIA BioNeMo and NVIDIA CUDA-X libraries, we provided the Gladstone team with the arithmetic throughput and memory efficiency needed to train a billion-parameter model in a fraction of the time. This collaboration will enable researchers worldwide to use MaxToki to rapidly identify anti-aging targets and discover new ways to slow or prevent age-related diseases like dementia and heart disease over 400 times faster.”

Aging With Disease

Because MaxToki was trained exclusively on data from healthy individuals, the team could test whether it would recognize something it had never been taught: the accelerated aging that accompanies some diseases. When presented with lung cells from patients with pulmonary fibrosis, the model calculated a 15-year acceleration in aging in those cells compared to healthy controls. Cells from heavy smokers showed a five-year acceleration in aging.

In data from a group of patients with Alzheimer’s disease, MaxToki detected about a three-year aging acceleration in microglia, a type of brain immune cell. This aging acceleration was also seen in a second dataset provided by Gladstone Investigator Andrew Yang, PhD. However, the same aging acceleration was not seen in the microglia of patients in the second dataset who had signs of Alzheimer’s in the brain but no symptoms of dementia—a phenomenon called Alzheimer’s resilience.

“The AI was only trained on normal aging, yet it correctly perceived accelerated aging in disease cells,” says Javier Gόmez Ortega, PhD, first author of the study and a bioinformatics fellow in Theodoris’s lab. “Using the model could help us identify the specific changes in gene activity that cause age-related diseases and point to potential ways of slowing or preventing them.”

With this goal in mind, the team used MaxToki to screen for aging-related genes in ventricular cardiomyocytes—the heart’s primary pumping cells. The AI model flagged dozens of candidate genes and the scientists selected five to test in the lab in collaboration with Gladstone Investigator Michael Alexanian, PhD.

When each of the five genes was activated in human heart cells, the resulting gene activity showed patterns seen in older cells, such as increased inflammation and disrupted mitochondrial function. With further experiments, the scientists found additional hallmarks of aging in the cells, including slowed calcium cycling and irregular beating.

Senior author Christina Theodoris and her team used MaxToki to analyze heart cells and tasked it with predicting genes that either accelerated or slowed normal aging. When the researchers tested the effects of those genes in human cells, the predictions held true.

“These were targets we would not have tested otherwise,” Theodoris says. “Seeing them cause functional decline in heart cells, with exactly the kinds of changes we associate with cardiac aging in humans, confirms the model is capturing something real about the biology.”

The team conducted early tests in young mice and found that activating the top two pro-aging genes led to signs of heart dysfunction within a month, suggesting the genes play a functional role in aging. The researchers are now testing whether blocking the same genes could protect against age-related heart decline.

The pre-trained MaxToki model and code are publicly available so researchers around the world can apply them to their own questions about aging and disease.