Gladstone Institutes unveils MaxToki, an AI model that predicts cell aging

MaxToki is an attempt to teach AI to see aging not as a set of disconnected biological snapshots but as a continuous trajectory of change in a cell over time. The team at Gladstone Institutes and its partners says the model can not only estimate how "aged" a particular cell state is, but also predict which genetic interventions may speed up or slow down the process. For aging research this is an important shift: instead of describing a cell's current state, there is now 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 one cell at a given moment. That is enough to recognize a cell type or its current state, but not enough to understand the dynamics of aging, which unfold over years and decades.

MaxToki is built as a decoder-only transformer and trained in two stages. First, the model studied individual cell transcriptomes from a corpus of roughly 175 million samples. Then its context was extended and it was further trained on age trajectories: about 22 million single-cell transcriptomes collected from roughly 3,800 donors of different ages, from birth to 90+ years. In total, training covered almost 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 reduces dependence on technical noise and keeps "housekeeping" genes, which are active almost all the time, from drowning out the signal. Instead, the model does a better job of picking out regulatory genes, including transcription factors, which often determine a cell's transition from one state to another.

In the second stage, MaxToki learned to work not with a single cell but with a short sequence of cell states and the time gaps between them. This made two types of queries possible: predicting how much time separates one cell from another, and generating the expected state of a cell after a given interval.

According to the authors, the model showed a clear gain on tasks that required reconstructing age trajectories for unfamiliar examples. On held-out cell types absent from training, the correlation between predicted and actual time shift reached 0.85. On held-out ages and donors, it was 0.77. The median time-prediction error was 87 months, versus 178–180 months for simpler baseline methods.

Also important: MaxToki does not receive explicit sex or cell type labels at inference time, and instead recovers context from the data itself. This brings the model closer to in-context learning in the style of language models, only in the biological domain.

Particularly interesting is the test on diseases the model never saw 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 an age-matched control group. For lung fibroblasts in pulmonary fibrosis, the estimate reached about 15 years. In microglia from Alzheimer's disease patients, the model showed about 3 years of additional age shift. No such signal appeared in people with mild cognitive impairment or in so-called Alzheimer-resilient patients, who have similar neuropathology but no pronounced cognitive deficit. This hints that the model may be picking up not just age, but specifically pathological acceleration of cell aging.

The strongest part of the work is the attempt to move from prediction to action. The researchers used MaxToki for in silico gene screening in heart cells and looked for targets that could push the trajectory toward aging or rejuvenation. Several previously undescribed pro-aging candidates were then tested experimentally. In human cardiomyocytes, their overexpression triggered inflammation and mitochondrial dysfunction programs and worsened cell function, including disruptions in calcium cycling and contraction rhythm. The two strongest candidates were also tested in young mice: according to the authors, within just six weeks this led to measurable deterioration of heart function. For biomedical AI this matters: the model does not merely interpret data elegantly, it produces hypotheses that survive laboratory testing.

What does this mean in practice? MaxToki remains a research system and has been published as a preprint, so it is too early to talk about clinical use. But the idea itself is very strong: if foundation models learn to model cell trajectories over time reliably, they could become a tool for early discovery of targets against age-related diseases, for selecting drug candidates and for testing hypotheses before expensive experiments. Put simply, biology gets something like a "simulator of the future" for cells, and that is far more interesting than yet another model that merely describes a snapshot of the present.