Gladstone Institutes Present MaxToki — an AI Model That Predicts Cell Aging

MaxToki is an attempt to teach AI to see aging not as a collection of disconnected biological snapshots, but as a continuous trajectory of cell changes over time. The Gladstone Institutes team and partners claim that the model can not only assess how "aged" a particular cell state has become, but also predict which genetic interventions might accelerate or slow this process. For aging research, this represents an important shift: instead of describing the current state of a cell, there emerges a tool for modeling where it is heading next.

Most biological foundation models today work with static data, most often single-cell RNA-seq — a snapshot of which genes are active in a single cell at a particular moment. This is enough to recognize cell type or its current state, but insufficient to understand aging dynamics that unfold over years and decades. MaxToki is built as a decoder-only transformer and trained in two stages.

First, the model studied individual cellular transcriptomes from a corpus of approximately 175 million samples. Then it was given expanded context and retrained on aging trajectories: about 22 million single-cell transcriptomes collected from approximately 3,800 donors of different ages — from birth to 90+ years. In total, training covered nearly 1 trillion "gene tokens."

One of MaxToki's key ideas is to represent a cell not through raw gene expression levels, but through their ranked order. This approach helps reduce dependence on technical noise and prevents "housekeeping" genes, which are active almost always, from drowning out the signal. Instead, the model better isolates regulatory genes, including transcription factors, which often determine the transition of a cell from one state to another.

In the second stage, MaxToki learned to work not with a single cell, but with short sequences of cell states and time intervals between them. This allowed for two types of queries: predicting how much time separates one cell from another, and generating the expected cell state after a given interval. According to the authors, the model showed notable gains on tasks where it was necessary to recover aging trajectories for unfamiliar examples.

On held-out cell types that were not in training, the correlation between predicted and actual time shift reached 0.85. On held-out ages and donors — 0.

77. Median time prediction error was 87 months compared to 178–180 months for simpler baseline methods. It is also important that MaxToki does not receive explicit labels of sex or cell type during inference, but recovers context from the data itself.

This makes the model closer to in-context learning in the spirit of language models, only in the biological domain. Particularly interesting is validation on diseases that the model had never seen during training, because it was trained only on "normal" aging. In lung epithelial cells from people with a heavy smoking history, MaxToki estimated aging acceleration of about 5 years relative to the control group of the same age.

For lung fibroblasts in pulmonary fibrosis, the estimate reached approximately 15 years. In microglia from Alzheimer's disease patients, the model showed about 3 years of additional aging shift. However, this signal was not present in people with mild cognitive impairment and in so-called Alzheimer-resilient patients, who have similar neuropathology but no pronounced cognitive deficit.

This hints that the model may be capturing not just age, but specifically pathological acceleration of cellular aging. The strongest part of the work is the attempt to move from prediction to action. Researchers used MaxToki for in silico screening of genes in heart cells and searched for targets that could shift the trajectory toward aging or rejuvenation.

Several previously undescribed pro-aging candidates were then tested experimentally. In human cardiomyocytes, their overexpression triggered inflammation programs and mitochondrial dysfunction, as well as impaired cellular function, including calcium cycling disruptions and contraction rhythm abnormalities. The two strongest candidates were also tested in young mice: according to the authors, within just six weeks this led to measurable deterioration of cardiac function.

For biomedical AI, this is an important moment: the model not only interprets data beautifully but also produces hypotheses that withstand laboratory testing. What does this mean in practice? MaxToki remains a research system for now and has been published in preprint format, so it is too early to talk about clinical application.

But the idea itself is very strong: if foundation models learn to consistently model cellular trajectories over time, they could become a tool for early target discovery against age-related diseases, drug candidate selection, and hypothesis testing before costly experiments. In simple terms, biology is getting an analog of a "future simulator" for cells — and that is far more interesting than another model that simply describes a snapshot of the present.