GenBio launches a virtual cell AI to simulate human cells
GenBio launches virtual cell modeling with AIDO Cell, a “world model” the Palo Alto AI startup says can simulate a human cell in its natural state and after drugs or other successive interventions. Co-founded by 2024 Nobel laureate David Baker, the early-stage system aims to let scientists interrogate biology computationally from DNA to whole-cell behavior. Lifespan.io reports the company describes it as a preview, not a solved map of the cell.
Key Takeaways
- GenBio AI says AIDO Cell can simulate a human cell from DNA and RNA through protein to whole-cell behavior.
- An early demo reproduced imatinib’s known mechanism of action on leukemia cells across several biological levels.
- The preview supports K562 and HepG2 lines, with more cell types and versions planned this year and next.
- GenBio is preparing early academic access and pitching the model as an in silico drug-testing ground.
What is AIDO Cell, and why does it matter?
Simulating life on a computer has been a decades-long goal, but earlier tools could not recreate biology’s complexity. Foundation models such as AlphaFold, ESM, GeneFormer, Evo, and Nucleotide Transformer still cover only slices of life. Palo Alto-based GenBio AI now says AIDO Cell is an early “world model” of a cell: natural state plus responses to drugs and other interventions, from DNA to the whole cell.
Baker, who received the 2024 Nobel Prize in Chemistry for computational protein design, said cellular biology is not solved. What is new, he argued, is one system that can simulate a cell across those scales in one place so researchers can interrogate it computationally. A Nature Medicine Perspective by co-founders Eric Xing, Eran Segal, and Le Song sketches three stages: modality-specific foundation models, links across scales, then joint alignment so the network behaves as one system.
What can the virtual cell actually do today?
In an early demonstration, GenBio modeled imatinib’s effects on leukemia cells and reported that AIDO Cell reproduced the drug’s known mechanism of action across several levels of cellular biology. The model is stateful, so successive perturbations can build on one another rather than remaining isolated predictions. It currently supports K562 and HepG2, immortalized lines from chronic myeloid leukemia and a liver tumor.
GenBio calls the release a preview and early functional demonstration. Additional cell types are in development, with more advanced versions planned later this year and over the next year. An early-access collaborator program is being prepared for scientists in academia, biotech, and pharma. Co-founder Emma Lundberg, a Stanford professor and co-director of the Human Protein Atlas, said virtual cell models should not stay behind closed doors: bring your data, and the AI can build a tailored virtual cell.
How could a virtual cell speed up drug and aging research?
Near term, AIDO Cell is pitched as a virtual testing ground so researchers can explore cellular drug effects in silico before physical experiments and drop weaker candidates earlier. Co-founder and CSO Ziv Bar-Joseph, who led AI for R&D at a major pharmaceutical company, said the hard problem was determining how a drug behaves in a real cellular environment, which often fed clinical trial failures. If accurate enough, he said, companies could evaluate several candidates quickly and reach better, safer treatments years earlier.
Similar systems could let scientists perturb genes, proteins, or pathways, follow predicted consequences across biological levels, and generate hypotheses faster than in the lab. That ambition matters for longevity and biohacking, where cell-level aging is still mapped slowly by hand. Separate research the same week showed five senescent brain cell types send and receive very different signals that can spread, or even discourage, senescence.
CTO Le Song said AIDO avoids a one-way chain of predictions, where a small molecular error balloons at the cell or patient level. Instead, predictions at molecular, cellular, and higher levels can be checked against real measurements, somewhat like feedback used to improve large language models. Remaining hurdles include a common language across scales and stopping small inaccuracies from accumulating as they propagate.