MaxToki is a temporal AI model for predicting the drivers of cell state progression over time, providing a generalizable framework to decode and control dynamic cellular trajectories.

MaxToki is a temporal AI model designed to generate past, intervening, and future cell states across dynamic trajectories. The model undergoes a two-stage training, first learning to generate single cell transcriptomes and then expanding the input size to model multiple cell states along a context-specific trajectory. By learning how gene network states direct future cell state transitions, the model can be prompted to generate cells along the continuum or predict how perturbations would impact the trajectory.

MaxToki was 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. Overall, MaxToki represents a promising strategy for temporal modeling to accelerate the discovery of interventions for programming therapeutic cellular trajectories.

The repository includes MaxToki-217M and MaxToki-1B, variants with 217 million and 1 billion parameters, respectively, pretrained on 175 million human single-cell transcriptomes. These models can be further trained with context-specific cell state trajectories during the second stage training, such as across human aging. The models can then be used to infer trajectory acceleration or deceleration in experimental samples compared to age-matched controls or leveraged for in silico perturbation analysis to predict targets that would modulate the trajectory.

Please note that GPU resources are required for efficient usage of MaxToki.

J Gόmez Ortega, PhD; C V Theodoris, MD, PhD