Virtual Cells: Testing Medicines on Simulated Patients First

In truth, most medicines fail, and they fail late. A compound can look perfect in a dish, survive animal studies and still stumble in the first human body. Researchers want to move that moment of failure much earlier, ideally before anything leaves the computer. Meanwhile, virtual cells are their name for the attempt. The idea behind the name is older than the name itself. Their bet is that computation can absorb some of that risk.

On paper, the pitch sounds like science fiction and reads like a business plan. Simulate a cell well enough, and you could trial drug ideas on it overnight. This guide explains what virtual cells actually are and where they came from. It also tracks the distance between the renderings and the reality.

Virtual cells: the 30-second answer

Virtual cells are high fidelity AI simulations of cells and cellular systems. A 2024 vision paper in the journal Cell set the goal, with researchers including Charlotte Bunne leading the argument. The claim is simple: AI can predict how biology responds to drugs or mutations. In practice, that means previewing a drug response on a simulated cell before any real one.

Overall, it amounts to preclinical testing moved inside the computer. In one line, virtual cells are patient models you can experiment on safely. The rest of this guide explains who is building them and what remains unproven. Keep that definition nearby, because the term gets stretched a lot.

What you’ll learn

Why a cell is so hard to simulate

In effect, a single cell is a busy city with no street signs. Signals flow, genes switch on and off, and molecules collide on schedule. Every drug response emerges from that crowded choreography. Capturing it in software means choosing which details matter. Simulation is the attempt to copy that choreography in code.

Complexity compounds at every layer. A cell manages energy, builds proteins, answers signals and copies itself, sometimes all at once. Each process pulls in many molecular players. No single equation captures the whole performance, which is why simulation stayed hard for so long.

Scientists have modeled pieces of this for decades. The long running Virtual Cell (VCell) platform grew from that computational cell biology tradition. It lets researchers simulate cellular mechanisms piece by piece. Those tools were exacting and useful. They were also narrow, built for specific questions rather than broad prediction.

A true cell simulation would need to predict, not just replay. It would take a perturbation nobody has tested and guess the outcome. That leap, from replay to prediction, is exactly what the new wave promises.

Why bother at all, when dishes and lab benches exist? Because dishes are slow, costly and limited to what we can observe. A simulation is patient, parallel and reversible. You can break it many times before breakfast, and nothing suffers. That flexibility is why researchers keep the faith.

None of this makes the goal foolish. It makes the goal expensive. Games could fake a world because the rules were written by their makers. Cells write their own rules, and they edit them constantly.

The 2024 paper that set the target

In 2024, the journal Cell published a vision paper titled How to Build the Virtual Cell with AI. Charlotte Bunne was among the leading authors, alongside researchers from biology and computing. The authors proposed building high fidelity simulations of cells and cellular systems. Such models would predict how biology responds to drugs or mutations before anyone runs an experiment. The title doubled as an instruction manual. Reading it today, the confidence is striking and the humility is instructive.

High fidelity deserves a definition here. It means the simulation tracks what a real cell would do closely enough to trust for decisions. Nobody claimed such models existed in 2024. The paper drew the blueprint and named the milestones instead. Blueprints, in other words, not finished houses.

That paper did something else, too. It gave the field a shared name and a shared target. The name was virtual cells, and the target was prediction across biology. Funding, startups and lab agendas began to organize around the phrase. That is how directions get set in modern science. Names matter, because they let scattered teams recognize each other.

Digital twins: the metaphor that stuck

A digital twin is a simulated model that mirrors a real system. The idea is already standard in engineering. Cardiology research uses digital twins too, simulating how a heart might respond. That track record outside biology is why cell researchers borrowed the concept. The name sounds fancy, yet the idea underneath is simple.

Swap the heart for a cell and the ambition scales up sharply. A cell twin would need to react to compounds nobody has tried. That is where AI enters, because learned models can generalize where hand built equations stall. The phrase AI biology now covers this growing toolbox. Virtual cells are the cell sized branch of that family tree. Generalization is the gift that machine learning brings to the table.

What labs and startups are building now

Blueprints for virtual cells are turning into software. In September 2026, UC San Diego researchers reported using 4D AI models of cells. Their goal was predicting how cells respond to drugs before experiments begin. Four dimensions, in this context, means watching cells change across space and time.

The work leans on digital twin thinking, tracking a living system over time. In fact, coverage describes the effort as early, so read it as a signal rather than a solution. Details remain thin, which is normal for fast moving research.

Startups and labs are building virtual cell signaling models as well. Cell signaling is the internal messaging that tells a cell how to behave. The aim is bold, predicting how any cell reacts to a compound before lab work. Company claims deserve the usual hedge, since marketing outruns validation in hot fields. The direction, however, looks real. A shortlist from a simulation still deserves a lab appointment.

Why does this race attract builders? Because prediction is the rarest currency in biology. Screens tell you what happened, while models promise what will happen. Companies that sell foresight never lack for customers.

Testing a drug idea, two ways

Every row is an ambition, not an inventory item. Labs still run the traditional path daily, and simulations only assist so far. The honest summary is that both columns coexist in 2026 pipelines. Its right column grows as the models earn trust. Both columns will coexist for years, and that is fine. Reading it that way keeps expectations calibrated.

The promise: fail cheap, protect patients

Follow the logic behind virtual cells and the appeal becomes obvious. Run preclinical testing inside computers first, then send only survivors to the lab. Animal studies could shrink as simulations take the first passes. Most importantly, patients would meet fewer compounds carrying avoidable toxicity. None of this happens automatically, and none of it is instant.

Cheaper failure changes research culture in a quiet way. When a guess costs little, teams guess more and fear less. As a result, more failure on screen can mean better success in clinics. The trash can fills faster, and the pipeline gets cleaner. Researchers call this failing early, and it is a compliment. That habit sounds small and compounds quickly.

Drug discovery history explains why early kills matter so much. The push toward the first AI-designed drugs met the same gauntlet of screens and safety checks. AI proposed the candidates, and screens judged them. Simulation aims to make that gauntlet kinder and quicker.

The economics rhyme with the ethics. A failure caught on screen costs almost nothing. In a human, a failure costs far more than money. That asymmetry is the entire business case, and the moral one too.

Hype check: what virtual cells cannot do yet

Now for the cold water. Virtual cells are not immune to that pattern. Current computer simulations cover fragments of biology, not whole living cells. Fragments are still useful, yet they are fragments. The models also need massive data, and biology’s datasets are messy and incomplete. Every prediction still needs experimental confirmation, which means the computer only half retires the lab bench.

1. Simulate fragments of biology reliably.
2. Connect those fragments into whole cell models.
3. Confirm predictions experimentally, every single time.

This is where simulations meet their oldest trap. A convincing rendering of reality is not the same as reality, and confident output can outrun evidence. Our look at AI stock prediction real or myth makes the same point for markets. There, models sound certain while reality votes last. Treat early cell simulations with that same respect. Markets and cells differ, but the psychology is identical.

Still, none of this argues against the field. It argues for grading the field on a curve. The phrase experimental confirmation keeps appearing around this field, and that repetition is the tell. Reading the caution sections first is a habit worth stealing. Caution here is a feature, not a bug.

Where this leaves patients and researchers

For patients, the promise of virtual cells is indirect but real. Safer preclinical filters mean fewer toxic surprises in early trials. Patient models would take that further, reflecting one biology instead of averages. Meanwhile, trials themselves are getting smarter, and AI in clinical trials already speeds matching volunteers to studies. Simulation at the cell level would slot neatly into that pipeline. The trial stage is where years and budgets actually disappear.

Genome models belong in this picture too, and the Evo 2 DNA model shows why. Reading DNA at scale tells simulators which mutations to test and which cells to mimic. Drug response prediction and genome reading are converging skills. The labs building one often want the other. Convergence, not competition, is the healthiest way to read this field. Biology, after all, begins with the genome’s instructions.

Researchers get a new division of labor. Models propose and rank, experiments confirm and refine. The scientists who thrive will be the ones who can question a simulation, not just run it. That skill has always separated careful science from lucky science. The craft is changing shape, not disappearing.

Where to go next in this series

This article is one spoke in our AI in medicine cornerstone, which maps the entire cluster for curious readers. The cluster grows as the field does, which is the fun part. Each spoke stands alone, so begin wherever your questions pull hardest. The cluster also tracks what remains unsolved, because every field needs its skeptics. Spokes land here as the science moves, so revisiting pays off.

If this piece raised your skepticism antennae, hold onto them. Curiosity plus caution is the reading mode this series assumes. It is also how the best labs read their own results. Skepticism, applied fairly, is how every field earns its keep.

One last angle on why patience matters here. Cell behavior involves thousands of interacting parts working at once. Capturing even a slice of that faithfully is a genuine achievement. Capturing all of it may take a decade or even longer. The science deserves time to prove itself properly. Meanwhile, partial models already guide some lab decisions today. For patients waiting on safer medicines, that steady build is genuinely good news. That is worth remembering when headlines rush the story.

Frequently asked questions

What are virtual cells?

They are AI simulations of cells and cellular systems that predict how biology responds to drugs or mutations. The goal is testing drug ideas on simulated biology before real experiments. In one line, it is prediction for biology instead of prediction for text. The term covers a research program more than one product.

Who proposed the virtual cell idea?

A 2024 vision paper in the journal Cell set the agenda, with researchers including Charlotte Bunne leading it. It builds on decades of computational cell biology, including the long running Virtual Cell platform. Vision papers matter because they give scattered efforts one target.

Are virtual cells the same as digital twins?

Not exactly. Digital twins mirror a real system, while virtual cells aim to predict biology broadly. The two ideas blend together in current research, which is why they share so much vocabulary. Shared language does not mean shared maturity, though. The overlap is real, but so is the difference.

Can virtual cells replace lab experiments?

No. Every prediction still needs experimental confirmation before anyone trusts it. Think of simulations as a filter that saves time, not as a replacement for evidence. Filters save time, and time is what trials never have enough of.

What has been achieved so far?

UC San Diego researchers reported 4D cell models predicting drug responses in September 2026. Startups are building cell signaling models that aim to predict compound effects. Both remain early stage results rather than finished tools. Treat both as direction of travel rather than arrival.

When will virtual cells reach real medicine?

Nobody can date that yet, and anyone who does is guessing. The honest sequence runs from fragments to cells to organs, over years. Watch for validated predictions, not demo videos. Patience here is not passivity, it is the correct reading of the evidence.

The bottom line

Virtual cells are a research program, not a finished product. The 2024 paper set a target, and labs are reporting their first steps. Avoidable toxicity is the enemy, and cheaper failure is the weapon. That combination deserves patience and attention in equal measure. Honest framing keeps the ambition intact. The language of twins, signals and simulations is now shared vocabulary.

Overall, keep the renderings in perspective and watch the validation instead. When a simulated prediction survives a real experiment, the field will have earned its headline. Until then, curiosity is the right amount of belief. The first validated simulations will be worth celebrating, quietly.

Sources

• Bunne et al., How to Build the Virtual Cell with AI, Cell journal, 2024 – cell.com
• UC San Diego coverage, 4D AI models of cells and digital twin drug response work, September 2026 – sciencedaily.com
• VCell, the Virtual Cell modeling platform, project documentation and history
• Industry coverage of virtual cell signaling foundation models, 2026
• Digital twin applications in engineering and cardiology research, review coverage
• Computational cell biology background, decades of mechanism modeling before AI