What happens to a single cell when it is exposed to a drug? Today, researchers often answer this question by measuring cells in the laboratory after treatment. Artificial intelligence could eventually help anticipate some of these responses. It learns from large numbers of cell measurements and estimates how a cell state is likely to change under a particular drug.
Researchers in Zurich and Lausanne are developing models of this kind. They are not creating digital copies of human beings. For now, their models perform smaller and clearly defined tasks: predicting molecular changes in individual cells or analysing how cells are organised within tissue.
The potential benefit is considerable. Drug candidates could be tested more selectively, experiments planned more efficiently and cellular responses understood in greater detail. But the road to a virtual cell is still long.
From measurement to prediction
Cells do not all respond to a drug in the same way. Even cells of the same type can differ in their state, genetic profile or environment. Single-cell measurements reveal these differences. They also have an important limitation: many methods fix or destroy cells during the measurement process. Researchers therefore usually cannot observe the same cell before and after treatment.
Models such as CellOT learn from groups of untreated and perturbed cells. They identify likely transitions between these cellular states. The model does not predict the future of an entire person. Instead, it estimates how a measured molecular cell state might change under a particular perturbation or treatment.
CellOT: an early answer
CellOT was developed by researchers at ETH Zurich, the University of Zurich and University Hospital Zurich. In the study, the model learned from single-cell measurements and predicted responses in new cell samples that had not been measured before. It was tested, among other settings, on drug responses in melanoma cells and on data from people with lupus and glioblastoma.
This is an important but limited step. CellOT does not reproduce a complete cell. It predicts a specific change in a cellular state – for example, how the activity of certain genes or proteins might shift. It cannot determine whether a treatment will work for a particular patient. Preclinical research and clinical trials remain essential.
The cell’s surroundings matter
A cell does not respond independently of its surroundings. An immune cell may behave differently inside a tumour than in an isolated cell culture. Future models therefore aim to connect different types of data and biological levels: RNA, proteins, microscope images, individualcells and tissues.
One example from Swiss research is VirTues. The model was developed for spatial proteomics. It analyses not only which proteins and cells are present, but also where they are located and which cellular neighbourhoods form within tissue. This makes it possible to describe biological patterns across several levels – from an individual cell and its surroundings to the tissue as a whole.
VirTues is not a complete virtual cell. It is a spatial-proteomics model that adds an essential dimension to the virtual-cell vision: cells do not function in isolation. Their surroundings influence how they respond to signals and drugs.
The reality check
These models hold promise. They could help identify promising drug candidates earlier and prioritise experiments more effectively. They do not, however, make laboratory work unnecessary. Reliable predictions depend on large amounts of high-quality, comparable and carefully controlled data.
The key question is therefore not how large a model is, but how well it has been tested. Does it work with new cell types, new drugs and data from other laboratories? Does it recognise its own limitations? And does it communicate when its prediction is uncertain?
An apparently precise answer can be misleading if a model has seen little training data from a rare disease or a particular population. External validation, transparent methods and clear communication of uncertainty are therefore just as important as technical performance.
Research and regulation
The stronger the influence of an AI model on drug development, authorisation or treatment, the higher the requirements for data quality, validation, documentation and ongoing monitoring must be. Swissmedic identifies data quality, model transparency, quality control and possible bias as key challenges when AI is used in medicinal product development.
A model used only to prioritise laboratory experiments must be assessed differently from one whose results may influence an authorisation decision or a treatment. The intended use determines the regulatory requirements.
Virtual-cell models are unlikely to replace laboratory experiments entirely in the foreseeable future. Their realistic value lies in helping researchers ask better questions, test hypotheses and plan experiments more precisely. Research in Zurich and Lausanne shows what this path could look like: moving from limited predictions of individual cellular responses towards models that also account for the surrounding tissue.
References
1. Bunne C. et al. Learning single-cell perturbation responses using neural optimal transport. Nature Methods, 2023. https://doi.org/10.1038/s41592-023-01969-x
2. Bunne C. et al. How to build the virtual cell with artificial intelligence: Priorities and opportunities. Cell, 2024. https://doi.org/10.1016/j.cell.2024.11.017
3. Bunne C. et al. Optimal transport for single-cell and spatial omics. Nature Reviews Methods Primers, 2024. https://doi.org/10.1038/s43586-024-00334-2
4. Wenckstern J. et al. The Virtual Tissues foundation model resolves spatial proteomics across scales. Nature, published 5 August 2026. https://doi.org/10.1038/s41586-026-10884-y
5. ETH Zürich. Predictions of the effect of drugs on individual cells are now possible, 2023. https://ethz.ch/en/news-and-events/eth-news/news/2023/10/predictions-of-the-effect-of-drugs-on-individual-cells-are-now-possible.html
6. EPFL. AI-Powered Virtual Cell Could Become Biology’s Universal Simulator, 2024. https://actu.epfl.ch/news/ai-powered-virtual-cell-could-become-biology-s-uni/
7. Swissmedic. Framework conditions for the use of AI in medicinal product development and the regulatory process. https://www.swissmedic.ch/swissmedic/en/home/humanarzneimittel/authorisations/artificiel-intelligence.html
(Image: Google DeepMind)
