Simulating a whole cell
No one can yet compute how a complete living cell behaves from its molecules up, let alone a human cell.
Open in the interactive tree →A virtual cell would predict, from DNA and conditions alone, how every gene, protein and reaction behaves and how the cell responds to a drug or a mutation. It would replace many lab and animal experiments. So far only the simplest cells are within reach.
As of October 2026
On 9 March 2026 Cell published a four-dimensional spatial and kinetic model of the minimal bacterium JCVI-syn3A (493 genes) that simulates a full cell cycle of about 100 minutes, including DNA replication, protein production and division; it took six days of computing on graphics processors and matched the real cell cycle within about two minutes on average. A human cell has roughly 20,000 genes, a genome thousands of times larger and far more complex regulation. Arc Institute's Virtual Cell Challenge (final submissions November 2025) tests AI models on predicting the effect of silencing single genes in 300,000 human embryonic stem cells, but only for one cell type and one readout.
What is missing
- Time-resolved measurements of all molecules inside single living cells (proteins and metabolites, not just RNA)
- Kinetic parameters for thousands of enzymes and binding reactions, measured under real in-cell conditions
- A theory that couples scales from atoms to organelles to the whole cell, including randomness and 3D space
- Far more computing power or smarter approximations: even the minimal cell needs days of computing on graphics processors
- Standard benchmarks and large perturbation datasets to validate predictions
Becomes possible once solved
- Testing drugs on a digital copy of a cell instead of on animals
- Designing synthetic cells on a computer before building them
- Predicting disease mechanisms from a patient's genome
Open steps
- Predicting unseen perturbations Medium AI leveragePredict how an unseen gene knockout or drug changes a cell type the model never saw, and beat simple baselines on real biology.
- In-cell reaction rates Medium AI leverageRate constants for thousands of enzyme and binding reactions are unknown under real in-cell conditions, yet a whole-cell model needs them as inputs.
- Fast stand-ins for cell simulation Medium AI leverageEven the 493-gene minimal cell needs six GPU-days per cell cycle; a human cell needs approximations that keep randomness and 3D space.
- Linking molecules to the whole cell Medium AI leverageNo theory yet couples atomic motion, protein complexes, organelles and the whole cell in one model with randomness and 3D space.
- Live molecule counts in single cells Low AI leverageTime-resolved measurements of proteins and metabolites, not only RNA, inside single living cells are missing as training and test data.
Where AI could help
Medium AI leverage. AI models are the main route, but the first big challenge showed they barely beat simple baselines, and the key data does not exist yet.
- Foundation models trained on perturbation screens to predict gene-silencing and drug responses in unseen cell types
- Surrogate models that replace slow stochastic simulations of sub-cellular reactions
- Active learning that picks the next most informative single-cell perturbation experiment
- Literature agents that collect kinetic parameters for thousands of reactions
Shown so far
- In June 2025 Arc Institute released State, a model trained on observational data from about 170 million cells and perturbation data from over 100 million cells that predicts how cells respond to drugs and gene changes. source
- At the end of the 2025 Virtual Cell Challenge, organizers reported that perturbation-prediction models were not yet consistently outperforming naive baselines across all metrics. source
- In March 2026 a GPU-accelerated physics-based simulation of the minimal cell JCVI-syn3A ran a full 105-minute cell cycle in six days of computing (not an AI model, but shows the compute scale needed). source