Researchers Train a 1-Trillion-Token AI on Human Cell Aging — and Validate It in Living Mice
A team from UCSF, Gladstone Institutes, and NVIDIA trained MaxToki — a foundation model on nearly 1 trillion gene tokens — to model how cells change across the entire human lifespan and identify targets that could slow aging-related decline. Crucially, its predictions were validated in live mice, and it distinguished Alzheimer's disease from resilience with no disease-specific training.
Original sourceA collaboration between UCSF, the Gladstone Institutes, and NVIDIA has produced MaxToki, a biological foundation model trained on nearly 1 trillion gene tokens that models cellular aging trajectories across the entire human lifespan — not just static snapshots.
The model's key capability is predicting not just what a cell looks like at a given age, but how it is changing and where it is headed. This temporal modeling allowed the team to identify molecular targets that could interrupt or slow aging-related cellular decline, which they then tested experimentally in living mice — a bar that most AI biology papers never reach.
Perhaps most striking was an emergent capability: MaxToki could distinguish between full Alzheimer's disease and Alzheimer's resilience — patients who carry the biological signatures of the disease but never develop cognitive symptoms — with no Alzheimer's-specific training data. It inferred this distinction entirely from the patterns in its aging trajectory model.
The model was developed with NVIDIA's BioNeMo infrastructure, reflecting a deepening partnership between NVIDIA and academic medical research centers that goes beyond GPU provisioning. The work was posted to bioRxiv on April 1, 2026, and detailed coverage appeared on April 5.
The implications for drug discovery are significant. Traditional approaches to aging research study cellular states at a single point in time. MaxToki's trajectory-based approach generates what researchers describe as a 'temporal map' of cellular decline — making it possible to intervene at the moment of maximum leverage rather than after damage is done.
Panel Takes
The Builder
Developer Perspective
“1 trillion gene tokens and experimental validation in mice puts this in a different category from most AI biology papers. The NVIDIA BioNeMo integration signals that this isn't a one-off academic project — this is infrastructure-scale biology research. Watch for this model to be used as a backbone for drug discovery pipelines.”
The Skeptic
Reality Check
“Mouse models have a notoriously poor translation rate to human therapeutics — the graveyard of Alzheimer's drug trials is full of interventions that worked perfectly in mice. The Alzheimer's resilience finding is genuinely interesting but needs replication in human cohorts before anyone should get excited about clinical applications.”
The Futurist
Big Picture
“A model that understands cellular aging as a trajectory rather than a state is a paradigm shift in longevity research. The Alzheimer's resilience finding suggests there are protective biological programs that evolution preserved — and that we can now map them computationally. This is one of the most important papers of the year.”