MaxToki, a temporal AI model trained on nearly 1 trillion gene tokens across human aging trajectories, accurately predicts cell state progression over time through in-context learning of unseen trajectories and identifies novel age-modulating targets that were experimentally validated to cause age-related gene network dysregulation, functional decline in cardiomyocytes, and cardiovascular dysfunction in vivo.

Foundational AI models have recently shown promise for predicting the impact of perturbations on cell states. However, current models typically consider only one cell state at a time, limiting their ability to learn how cellular responses unfold over time, particularly across long trajectories such as diseases of aging. Here, we develop a temporal AI model, MaxToki, trained on nearly 1 trillion gene tokens including cell state trajectories across the human lifespan to generate cell states across long timelapses of human aging. MaxToki generalized to unseen trajectories through in-context learning and predicted novel age-modulating targets that were experimentally verified to influence age-related gene programs and functional decline in vivo. MaxToki represents a promising strategy for temporal modeling to accelerate the discovery of interventions for programming therapeutic cellular trajectories.