A Look at Virtual Cells—and How They Could Change Experimental Cell Biology
AIDO Cell was presented by GenBio AI in August 2026. In this AI-centric era, it is an attempt to build something biologists have wanted for a long time: an AI-driven digital organism that you can experiment on inside a computer. This is much more complex than many of the prediction models that are already becoming part of our day-to-day work in life sciences. Instead of focusing on one feature, AIDO tries to integrate many different aspects of cellular biology and recreate how a cell responds to perturbations.
I actually remember playing around with a virtual lab during college. It was during a physiology class, you could run a very basic experiment where you administered different doses of a diuretic to a mouse and observed how that translated into increased urine volume. Obviously, the simulation was extremely simple, but the underlying experimental logic was similar: you perturb a biological system and observe how it responds. What AIDO is proposing feels like a much more sophisticated extension of that idea, at the cellular level, where you can perturb a cell and explore how that change propagates across many different layers of biology.
If you have ever worked in a biology lab, you know that there is almost never just one experiment you could do next. There are ten, fifty, or even hundreds if you are screening for something in particular. You could knock out one gene, overexpress another, inhibit a pathway, combine two perturbations, change the dose, change the timing, or repeat everything in another cell type. Very quickly, even a fairly simple biological question becomes a huge experimental decision tree. In practice, we navigate that tree using previous data, the literature, and our own experience, while also being constrained by time and money. What a virtual cell could offer is the possibility of exploring much more of that experimental space computationally and then deciding what is actually worth validating in the lab. We are already moving in this direction in other areas of biomedical research, for example when computational methods are used to identify molecules likely to bind a particular protein or to help design better therapeutic antibodies.
There is a broader effort in the field to make biology something that can be simulated computationally, and AIDO Cell is one attempt to turn that idea into something more concrete. Instead of predicting one isolated biological feature, it tries to give you a persistent virtual cell that you can perturb repeatedly, observe in different ways, copy, rewind and send down different experimental paths. It is still very early and, as of now, access is limited: GenBio says the platform is being used internally and with alpha collaborators, while broader academic and industry access is being prepared through a waiting list. I think it is important to emphasize that point because it sets the right expectations. This is not yet a mature digital human cell ready to replace wet lab. The current version is a prototype platform, but the concept behind it is interesting enough that I think it is worth asking what could happen if this type of system becomes genuinely good.
What makes AIDO different from another biological AI model?
We already have an extraordinary number of AI tools in biology. We can predict protein structures, analyze regulatory DNA, estimate how gene expression might change after a perturbation, model molecular interactions and extract phenotypes from images. What AIDO is trying to do differently is stop treating all of those as completely separate prediction problems and instead maintain something closer to an internal representation of the cell itself.
Imagine that you start with a virtual cancer cell and knock down Gene A. The system updates the state of that cell. You then apply Drug B, but Drug B is not being applied to a fresh copy of the original cell; it is being applied to the cell that has already experienced the knockdown. You might then introduce a mutation, look at what has happened to gene expression or protein localization, clone that state and try different drugs on the two copies. GenBio organizes this around five main operations: observe, perturb, simulate, branch or restore, and design. The virtual cell carries a history rather than answering each question independently.
That may sound like a technical distinction, but I think it is one of the most important aspects of AIDO. Biologists take a cell’s history into account when designing experiments; now this model itself is designed to retain that history too. Instead of treating each perturbation as an independent prediction, it updates the virtual cell after each intervention, and that new state becomes the starting point for whatever you do next. A drug treatment can be followed by a knockout, a resistance mutation, a rescue experiment or another drug, with each step building on the previous one. That makes it closer to simulating an experimental trajectory than simply asking a series of unrelated prediction questions.
AIDO also tries to connect different layers of biology inside that same state. The current version can generate simulated readouts including chromatin accessibility, histone modifications, transcription-factor binding, RNA isoforms, protein structures and interactions, protein abundance and localization, pathway information and cell morphology. The idea is not simply to assemble a menu of separate AI tools, but to make those outputs different views of the same underlying cell. If a perturbation changes transcription, that change should propagate into downstream protein abundance, localization, pathways and eventually phenotype. We normally separate transcriptomics, proteomics, imaging and epigenetics because our experimental methods force us to measure them separately; in a virtual cell, at least in principle, you can observe all of those predicted readouts as part of the same simulated state.
The branching option that the simulator offers is another feature I find especially useful. If you reach an interesting state after several perturbations, you can save it, make copies and explore ten or a thousand different next steps. You can return to an earlier state, replay the experiment and take another branch. The cell can essentially be stored, restored, copied and resumed without consuming the original state. If you think about how we normally work in the lab, this is a big deal because every decision closes other doors: you use the cells, consume reagents, sacrifice samples or commit weeks to a particular experimental direction. With software, you can theoretically explore branches you would otherwise never have the time or resources to test.
To me, this is probably the most realistic short-term value of a virtual cell, and it is similar to how we are already beginning to use other AI predictions in biomedicine. It does not replace my experiment; it helps me explore a much larger space of possibilities before I commit time and resources to the bench. That alone will change how we design experiments.
AIDO also asks the question backwards
Most experiments begin with an intervention: I inhibit/remove this protein; what happens? AIDO is also designed to approach the problem from the opposite direction: this is the cellular state I have, this is the state I want — what intervention could move me from one to the other?
GenBio calls this in-context molecular design, and in version 1.0 the system includes functions intended to propose small molecules, antibodies or nanobodies that could drive the virtual cell toward a desired state. Think about this, for example, in the context of aging and the idea of pushing a cell away from an aged phenotype and toward a younger one. If this type of prediction becomes reliable, the implications for drug discovery are obvious. Instead of always starting with a target and asking which compounds affect it, you could begin with a disease-associated or otherwise undesirable cellular state and ask which intervention is predicted to move the system toward a healthier state.
I would put a very large asterisk next to that today. Designing a molecule computationally is one thing; showing that it binds properly, reaches the right cells, has reasonable pharmacology and actually changes disease biology is something else entirely. Still, conceptually, this turns the virtual cell from a system that simply predicts outcomes into something that could actively participate in experimental design.
Where I think this could really change science
One narrative around technologies like this one is that one day we will stop doing wet-lab experiments because AI will simulate everything. I am much less interested in that story because I think it misses the more realistic and potentially more important change.
The real value, is prioritizing experiments rather than replacing experimental proof.
Suppose you are studying a disease pathway and there are 500 plausible perturbations you could test. In the laboratory, you might choose ten based on the literature and your own biological reasoning. A good virtual-cell system could potentially let you explore all 500, perhaps including combinations, and then bring the most informative twenty back into the laboratory. Maybe those are the interventions predicted to have the strongest effect, or maybe they are the ones where the model produces an unexpected result that is worth checking. The wet lab still tells you what actually works; the simulator helps you decide where it is most valuable to look.
What I can imagine is a workflow that becomes increasingly iterative: hypothesis → many virtual experiments → prioritized predictions → physical validation → new data → improved model → another round of virtual experiments. At that point, the role of the experiment starts to shift slightly. We may use fewer experiments simply to explore what might happen and more experiments to establish ground truth, resolve uncertainty and test the places where the model is most informative.
This is also where we need to be careful
The current AIDO release should not be interpreted as a faithful digital replica of human cellular physiology. Version 1.0 includes virtual models of only two cell lines, K-562 and Hep-G2. GenBio explicitly states that these are prototypes designed primarily to demonstrate what a persistent, stateful virtual cell could enable, rather than definitive, high-fidelity simulations of those cell lines. It is also worth remembering that K-562 and Hep-G2 are highly characterized transformed cell lines, which makes them useful starting points for a proof of concept, but also limits how far we can generalize the results. Accurately simulating these cells would not necessarily mean that the same model could reproduce the behavior of primary cells, differentiated tissues or patient-specific biology.
The examples in the white paper also rely heavily on biology we already understand well, such as BCR-ABL signaling and imatinib resistance in K-562 cells. That is exactly what you would expect from a proof of concept because well-characterized systems give you something against which to compare predictions. But reproducing known biology and correctly predicting biology that nobody has measured yet are very different achievements.
This is where independent benchmarking becomes essential. Dr. Silvana Konermann and colleagues at the Arc Institute have been very explicit that virtual cells need rigorous, standardized testing. Their Virtual Cell Challenge was created in part to establish shared benchmarks and more rigorous standards for evaluating how well these models simulate cellular behavior. In the 2026 Virtual Cell Challenge, models are asked to predict CRISPRi responses in six cell lines in which they have never seen perturbation data, with the experimental measurements withheld as ground truth. There is good reason to be demanding here. In 2025, Constantin Ahlmann-Eltze, Wolfgang Huber and Simon Anders compared several deep-learning and single-cell foundation models with deliberately simple baselines for predicting transcriptomic responses to genetic perturbations, and none of the deep-learning models they tested outperformed the simple baselines. That does not mean AIDO will fail, because it is a different and broader system, but it is a useful warning that technical sophistication is not the same as biological understanding.
A cell is also much more than what we currently know how to measure
There is another limitation that is easy to underestimate. A model of a cell can only be as complete as the biology we are able to show it, and our measurements are still incomplete. Transcriptomics is extraordinarily powerful, but RNA is not a cell. Protein abundance matters, protein localization matters, post-translational modifications matter, metabolism matters, protein turnover matters, cell-cell interactions matter, and all of these processes are dynamic and change over time.
AIDO already tries to go beyond transcriptomics, which is one of the reasons I find it interesting, but GenBio also acknowledges that important components such as post-translational modifications, degradation and turnover, signaling dynamics and metabolism still need to be expanded. That is not a minor limitation; those are fundamental pieces of cellular behavior. So when we use the phrase “virtual cell,” I think we should be careful not to interpret it too literally. Right now, AIDO is better understood as an ambitious framework for integrating many different models and data types into a persistent computational representation of cellular state. Whether that representation eventually becomes accurate enough that we routinely trust it to stand in for particular experiments is still an open question.
How I would like to see it evolve
What I would really like to see next is the model making predictions about biology that we do not already know, followed by independent experimentalists testing those predictions, feeding the results back into the system and repeating the cycle. That iterative loop is where I think these models could become genuinely powerful: predict, test, learn, retrain and try again.
I would also want to see clearly where the model works, where it fails, which cell types it generalizes to, which kinds of perturbations it handles well and where its confidence starts to become unreliable. A model that can tell us when it does not know may ultimately be much more useful than one that confidently produces an answer to every question. We already see this problem all the time with large language models: a confident answer is not necessarily a correct one, and in biology the consequences of that distinction are important.
I don’t think this future makes experimental scientists less important. If anything, it makes strong experimental biology even more valuable, because these models ultimately depend on high-quality data, carefully designed perturbations and scientists who can judge when a prediction is biologically plausible and when it needs to be challenged. A model may be able to explore a much larger experimental space than we ever could at the bench, but real experiments will still - for now, at least - be essential to validate those predictions, reveal what the model has missed and provide the data needed to improve it.
Would you like to stay updated on the latest breakthroughs in biomedical science? Subscribe to my blog and join me in exploring the next frontier of medicine!