Researchers propose testing a computer model of a cell with the Yamanaka factors, four proteins that gave mouse fibroblasts stem cell properties
Researchers propose testing a computer model of a cell with the Yamanaka factors, four proteins that gave mouse fibroblasts stem cell properties
In a Dialectic interview on October 15, Arc Institute co-founder Patrick Hsu proposed a test for a cell model: it should rediscover known causal relationships in biology. His example was the 2006 experiment in which four Yamanaka factors converted mouse fibroblasts into induced pluripotent cells.
In the Dialectic interview, Hsu describes this experiment as a test for a “virtual cell,” a computer model that takes a description of a cell’s state and predicts its response to an intervention. In the 2006 experiment, Shinya Yamanaka and Kazutoshi Takahashi introduced four transcription factors, proteins that regulate gene activity, into mouse fibroblasts. The cells acquired the properties of induced pluripotent stem cells, which can develop into different cell types.
“Can a virtual cell model rediscover the Yamanaka reprogramming factors?”
An OpenAI model has already proposed variants of the Yamanaka factors that were subsequently tested in the laboratory. Hsu proposes a more demanding test: the model should independently identify the factors responsible for the change in cell state that biologists discovered in 2006.
This test requires data on the effects of interventions. Arc collects these data using Perturb-seq: different genes are suppressed in cells within a single batch, and RNA is then measured in each cell as a readout of gene activity. This produces paired interventions and measured responses, which the model uses to learn to predict the effects of a subsequent intervention.
In the Virtual Cell Challenge 2026, Arc tests how well these predictions transfer across six laboratory cell cultures. Models receive baseline cell profiles and the names of genes to be suppressed using CRISPRi, a method that reduces the activity of a selected gene. The measured laboratory response is withheld as a reference; participants predict how the intervention will change the RNA profile in a new cellular context.
The Yamanaka factors provide a historical test: can a model identify a causal relationship that biologists have already established experimentally? The competition tests whether predictions transfer to experiments whose results are withheld. A laboratory then measures the response to a new hypothesis proposed by the model, and that measurement feeds back into training.