MaxToki: AI Model Predicts Cellular Aging and Disease Trajectories

Researchers at the Gladstone Institutes have introduced MaxToki, a temporal foundation model designed to forecast how human cells age. Unlike traditional biological models that view cells as static snapshots, MaxToki treats cellular biology as a dynamic process, allowing it to predict gene network trajectories across the human lifespan. By analyzing single-cell RNA sequencing data, the model can identify how cells progress over time and pinpoint targets to potentially slow the development of age-related diseases.

A New Approach to Biological Time

MaxToki functions as a transformer decoder model, similar to the architecture powering large language models. The research team trained the model using a rank value encoding strategy, which represents transcriptomes as a ranked list of genes rather than raw counts. This approach emphasizes transcription factors—the master regulators of cell state—and provides robustness against technical batch effects. The model was trained in two stages on nearly 1 trillion gene tokens, utilizing data from diverse human tissues to capture the nuances of cellular aging from birth to age 90 and beyond.
The model’s most significant innovation is its temporal prompting strategy. By providing a context trajectory of cell states and a specific timelapse, MaxToki can predict the resulting transcriptome of a cell at a future point. Because the model uses continuous numerical tokenization, it understands time as a continuum rather than discrete categories. This allows it to perform in-context learning, enabling the model to generalize to cell types and donors it never encountered during its training phase.

Uncovering Disease Signatures

One of the most striking capabilities of MaxToki is its ability to infer age acceleration in disease states without prior disease-specific training. When tested on lung tissue from smokers or patients with pulmonary fibrosis, the model identified significant age acceleration compared to healthy controls. Similarly, in Alzheimer’s disease research, the model successfully distinguished between patients with the disease and those who were Alzheimer-resilient. While patients with Alzheimer’s showed clear signs of age acceleration in their microglia, resilient individuals did not, suggesting the model can identify protective biological signatures.
The utility of MaxToki extends beyond computational prediction. The model successfully nominated novel pro-aging drivers in cardiac cells, which were subsequently validated in living mice. These findings demonstrate a direct path from in silico screening to biological verification. By making the model and training code publicly available, the researchers aim to provide a framework that the scientific community can use to explore new tissue types and identify intervention points for age-related conditions before they manifest.