AI cell model spans 1.5 billion years of aging biology
One model that understands the biology of virtually every cell on Earth. That sounds far-fetched, but researchers have built something close to it with TranscriptFormer. For aging research, it offers a new way to compare cell behavior across species and disease states.
The model, published in Science, was trained on cell data from organisms separated by 1.5 billion years of evolution, from algae to humans. It learns patterns in gene activity (transcription) and can predict how cells behave, even in situations it has never encountered before. The researchers call it a generative cell atlas: a model that does not merely store what has been measured, but can describe new cell states.
What makes a generative cell model different
A conventional cell atlas is like an encyclopedia: you look up what is known about a cell type. A generative model is more like a language model for biology: it grasps the underlying grammar and can construct sentences it has never seen. TranscriptFormer can simulate cell states, including those associated with aging, disease, or treatment.
For longevity research, that is promising because aging is a gradual and heterogeneous process. Cells in the same tissue do not all age at the same rate or in the same way. A model that understands the range of possible cell behaviors could help predict which cells deviate early from a healthy pattern and how they respond to an intervention.
Early days, but the direction is right
TranscriptFormer is primarily a research instrument. It has not yet been deployed in direct clinical applications, and prediction quality depends heavily on the training data. It is also unclear how well the model handles specific human disease processes that are underrepresented in the atlas.
Still, tools like this mark a shift in aging research: from measuring isolated biomarkers to modeling entire biological systems. In principle, that makes it possible to calculate the effects of lifestyle changes, drugs or genetic variants at the level of individual cells, though that remains a future ambition rather than a current reality.
Search terms to explore further: generative cell model transcription atlas, single-cell RNA sequencing evolutionary comparison, cell type classification machine learning aging