Virtual cells,
explained.

A cell is the basic unit of life. Inside a tiny boundary, it uses energy, builds and repairs itself, processes information, and responds to its surroundings. It is doing the work of staying alive.

Some cells live as whole organisms. In your body, cells work together as tissues and organs. A liver cell behaves differently from a nerve cell—and the same liver cell can behave differently when it is fed, stressed, or exposed to a drug.

That changing behavior is what we want to understand and predict.

A virtual cell is a model you can ask “what if?” You describe a cell and a change. The model predicts a particular response.

Imagine giving a drug to liver cells. You want to know whether they keep functioning, become stressed, or recover after the drug is removed. A model might tackle one part of that question: how RNA—the messages and working molecules copied from DNA—changes after a specified dose and time.

Predicting RNA is useful. Predicting recovery is a further task, with a different test. The phrase “virtual cell” covers models with very different abilities.

The useful question is: what can this model predict, for which cells, under which conditions?

If the prediction is reliable, it can help choose which compound, dose, or experiment to try next. The point is to make experiments more informative.

For a model that learns from data, the starting point is an experiment: change something, measure what follows, and keep a record of the conditions.

Biologists call the deliberate change a perturbation. It can take several forms.

First, collect examples. Give cells different treatments and record their responses. The dose, timing, cell type, and comparison condition belong with the data.

Then, train the model. It learns relationships between what changed and what happened. The output could be RNA levels, cell shape, growth, or another measured feature.

Finally, ask a new question. Hold an experiment out of training. Ask the model to predict it, then compare the prediction with the actual result. This is validation.

A model can fit familiar examples and still fail on a new drug or cell type. The test is what it gets right beyond the examples it learned from.

A cell is one connected system. Our instruments give us different views of it.

One method reads DNA. Another counts RNA. Another measures proteins. These large-scale studies are called omics: measurements of whole classes of biological components.

A parts list is not a cell state. The amounts matter, but so do the locations, interactions, surroundings, and activity of those parts.

Think of chess. Knowing which pieces exist does not tell you the position or what happens next. For a cell, we need to understand how its components are organized and working together.

Our data often leave us with a sharded representation—pieces of information about the same living system, collected through different methods and sometimes from different cells.

Multiomics combines several molecular views. Combining molecular measurements with a cell’s shape and behavior gives us a fuller view. The aim is to connect what a cell contains with what it does.

For our liver-cell example, a change in RNA is one clue. To understand recovery, we also need evidence about how the cells function afterward.

Return to our drug-treated liver cell. Four hours later, it is alive. Is it recovering? Adapting? Becoming injured? What happens when the drug is removed?

A final reading can miss the route the cell took to get there.

A trajectory is a path of change through time. Did a response begin immediately or after a delay? Did it keep growing? Did it reverse after treatment stopped?

Some models infer these paths from snapshots. Repeated measurements can provide direct evidence about how features change. Following the same individual cell also requires tracking it.

This is where trajectomics comes in.

Trajectomics is the name I use for studying cellular response histories: what changes, where, when, and in what order.

The proposal is to give models richer examples of living responses, then test whether those histories improve prediction. In our drug example, the question is whether the early response helps predict later function or recovery.

To learn a response history, we need measurements through time. One approach is to keep cells alive in a controlled environment, introduce a treatment, and observe what changes.

Flowing fresh medium through the chamber—perfusion—lets us control the exposure. Replacing drug-containing medium with drug-free medium is washout. The response may continue after the drug supply stops.

The concept above senses the cells and the medium before and after the chamber. It connects what we deliver with what we observe over time.

This is the measurement direction behind Cell Cinema: living-cell chemical histories that models can learn from. The experiment records the response; a model must still earn its prediction on new data.

They can already answer some well-defined questions. CellBox predicted responses to drug treatments left out of training in a melanoma cell line. PRESCIENT modeled how cells develop and checked predictions about the kinds of cells they become against records linking cells to their descendants.

These are specific successes. The larger challenge is reliable prediction across new treatments, cell types, and environments.

When you hear a virtual-cell claim, ask three things: what did it predict, what was new to the model, and how was the answer checked? Those questions turn a broad promise into something you can evaluate.

And a virtual organ? A virtual human?

Cells communicate and organize into tissues. Tissues form organs. Connecting models across those scales is the larger ambition.

Live-cell observations can help describe how cells change and interact. A virtual liver also needs tissue architecture, transport, and organ function. A virtual human adds communication between organs and the rest of the body.

The AI virtual-cell perspective sets out this direction. The path forward is to build capabilities that can be tested, then connect them.