On April 6, 2026, David Sinclair highlighted a new preprint that brings AI directly into the heart of cellular aging. Researchers at Gladstone Institutes, in collaboration with NVIDIA and others, have built MaxToki — a temporal AI foundation model trained on massive single-cell gene expression data spanning the entire human lifespan.
The model doesn’t just observe how cells change with age. It learns the trajectories of cell states over time and predicts perturbations that can restore youthful identity. Experimental validation in vivo already shows the predicted targets influence age-related gene programs and functional decline.
This is biology meeting modern foundation modeling at scale.
Cells Lose Identity Over Time — MaxToki Learns How to Restore It
Think of cellular differentiation as Waddington’s epigenetic landscape: young cells sit stably in their developmental valleys. Over decades, noise and damage push them off course. The cell doesn’t lose its DNA code — it loses clear instructions on which genes to express. This drift in cellular identity is central to David Sinclair’s Information Theory of Aging (ITOA): aging as progressive loss of epigenetic information.
MaxToki was trained to model exactly these long-term trajectories. The team used nearly 175 million single cells — approximately 1 trillion gene tokens — from healthy humans ranging from birth to over 90 years old. By treating gene expression as a temporal sequence, the model learns to generate plausible cell states across extended time periods — essentially simulating how a cell “ages” or, crucially, how it can be nudged back.
The name MaxToki comes from the Japanese Shinkansen bullet train and is a homonym for “time” in Japanese — fitting for a model that moves biology along the time axis.
What MaxToki Can Actually Do
The model ships with four core capabilities that set it apart from static single-cell approaches:
- Temporal generation: Produces cell state trajectories over long timelapses of human aging
- In-context learning: Generalizes to unseen trajectories without retraining
- Perturbation prediction: Identifies novel age-modulating targets — genes or pathways — that can shift cells toward more youthful expression profiles
- Experimental validation: Several predicted targets were tested in vivo and shown to affect age-related gene programs and reduce functional decline
The preprint credits Shinya Yamanaka — Nobel laureate for induced pluripotent stem cells (iPSCs) — as a co-author, linking the work directly to cellular reprogramming concepts. NVIDIA contributed critical acceleration of the underlying code for efficient training on specialized AI infrastructure.
Why Temporal Modeling Changes the Question
Most single-cell models today are static snapshots. MaxToki is explicitly temporal — it treats aging as a dynamic process rather than isolated states. That shift unlocks new kinds of questions:
- What minimal set of interventions can reset a drifted cell without full reprogramming?
- Can we simulate thousands of potential perturbations in silico before picking candidates for wet-lab testing?
- How do different cell types — neurons, cardiomyocytes, immune cells — age on their own trajectories, and what shared or unique rejuvenation levers exist?
The model’s ability to generalize via in-context learning suggests it could become a powerful co-scientist tool: feed it a starting cell state and a desired youthful endpoint, and it proposes candidate drivers. Pretrained MaxToki models are already available on Hugging Face, and code is released via NVIDIA’s Digital Biology Research GitHub — continuing the open-science momentum the field has built in recent months.
Programming Therapeutic Cellular Trajectories
At BioComputer we follow moments when biological data becomes truly programmable. MaxToki is one of those moments. It turns decades of single-cell atlases into a generative engine for hypothesis generation in aging and rejuvenation.
Combined with partial epigenetic reprogramming — OSK or full Yamanaka-factor approaches — temporal models like this could help optimize dosing, timing, and cell-type specificity, reducing risks while maximizing functional restoration.
Sinclair’s thread frames the work as reinforcement for the Information Theory of Aging: identity doesn’t vanish with age; it drifts. The logical next step is learning how to push the marble back into the right valley.
With models like MaxToki, that push can increasingly happen in silico first.
References
- Gómez Ortega et al. (2026). Temporal AI model predicts drivers of cell state trajectories across human aging. bioRxiv. https://doi.org/10.64898/2026.03.30.715396
- Sinclair, D. (2026). X thread on MaxToki and ITOA. https://x.com/davidasinclair/status/2041044367347986483
- Gladstone Institutes. (2026). New AI Model Predicts How Cells Age. https://gladstone.org/news/new-ai-model-predicts-how-cells-age
Related: The Information Theory of Aging · Single-Cell AI Foundation Models · AI-Biology Convergence
Feature image: AI-generated using Grok.