From Virtual Cell to Virtual Patient: The Missing Layer in Between

In late 2024, Cell published a position paper jointly signed by more than forty scholars, with a title that hides nothing: "How to build the virtual cell with artificial intelligence."

The author roster reads like a roll call of the strongest computational-biology forces on this planet: Stanford, Harvard, the Chan Zuckerberg Initiative, Genentech, and the Arc Institute. The vision is equally clear—build an AI Virtual Cell that uses multi-scale foundation models to learn cellular states, so that given a perturbation (knock out a gene, add a compound), it can predict with high fidelity what will happen to the cell.

Overnight, the virtual cell became the star of the moment. The initiating institutions followed with funded programs, and research organizations and companies around the world moved to stake their positions.

Yet a harder question has rarely been answered head-on:

> From virtual cell to virtual patient—what lies in between?

The intuitive answer: more cells. The human body is built from tens of trillions of cells; make the virtual cell bigger, more complete, and connect them up—is that not a virtual patient?

This essay argues exactly the opposite:

> A patient is not a bigger cell. What is missing in between is an entire layer of scale.

1. A Patient Is Not a Bigger Cell

From cell to patient is not an enlargement in quantity but a leap in quality: molecule, cell, tissue, organ, body—every transition between scales is accompanied by emergence, and the laws of one level cannot automatically derive the laws of the next.

This is exactly where the glory of reductionism fails. This year's sixteen-author review in Nature Reviews Drug Discovery (whose verdict of "data explosion, knowledge poverty" I discussed in my previous essay) puts it plainly: living systems possess holism and emergent effects; pathway redundancy, compensatory activation, and positive and negative feedback loops are pervasive, and single-target blockade is frequently offset by compensatory mechanisms—regimens that look impressive in vitro and in animal experiments fail in large numbers once they enter the human body.

That is why the industry keeps repeating its dark joke: curing cancer in mice is far too easy.

The review also flags another long-standing conceptual confusion: a ligand is not a drug. A ligand only needs to produce biochemical binding activity against a target; a true drug must clear the bar on every dimension—efficacy, safety, drug-likeness, and tolerability. Large numbers of high-affinity ligands carry excess toxicity, insufficient solubility, or metabolic defects, and hold no drug potential whatsoever.

> A molecule that binds is not a body that responds.

Between "binding" and "response" lies exactly that layer of scale.

An analogy: however perfect the virtual cell becomes, it amounts to knowing a building's electrical diagram by heart—which wire feeds which socket, which breaker lights which lamp. But what a physician faces is not a building; it is a city—water, electricity, gas, and traffic shaping one another in real time, a city that repairs and adapts itself—and every patient is a city built to a different plan.

Make the building's wiring diagram ten thousand times more precise, and it still will not yield the city's traffic dispatching.

> What is missing is not a bigger model, but a map at another scale.

2. A Map at Another Scale

In my previous essay on the quantification of knowledge, I drew the distinction between maps and travelogues: on a map every object has coordinates, the objects are joined by a road network, and the road network supports route planning.

The virtual cell builds its map in a cellular coordinate system: the axes are gene expression, protein abundance, and cell state. However finely the map is drawn, the coordinate system itself determines that it cannot answer questions such as "how will this person's hepatic metabolism compensate"—the latter requires a coordinate system at the scale of the human body.

The question then follows: how do you convert between two coordinate systems?

Our answer is to introduce an intermediate layer—the module. In our research with the Future Laboratory of Tsinghua University, we built a human-scale representation from 332 modules—the module map of capability omics (Capomics): covering the Aging Hallmarks, organ systems, immunity, and nutrition, and also incorporating Traditional Chinese Medicine (TCM) syndrome proxies and medicine-food homology targets. Molecular-scale events (binding, inhibition, agonism) are mapped onto module-scale functional changes (pathway activity, cellular processes, organ function), and the modules then compose the whole-body response.

> "Molecule → module → human": the module is the conversion layer between the two coordinate systems.

The virtual cell is good at answering "after this perturbation, what happens to this cell"; the virtual patient must answer "after this intervention, what happens to this person." The translation in between cannot be bypassed.

One more point deserves emphasis: the modules are not 332 isolated targets but an interaction network—the redundancy, compensation, and feedback relations among modules are modeled explicitly in the network structure. Perturb one module, and what you watch is the ripple across the entire map, not the reading of a single point. This is precisely the answer to the emergence problem raised in Section 1: compensation can be simulated only because it is written into the map's road network, rather than being left behind in the subordinate clauses of the literature.

3. The Two Things a Virtual Patient Has That a Virtual Cell Does Not

I once defined the world model through five elements: State, Action, Transition, Objective, and Feedback.

When the scale shifts from the cell to the human body, all five elements must be rebuilt on a new foundation. Two of them are things the virtual cell cannot supply by its very nature—and the virtual patient must grow them for itself.

The first: the medical semantics of Action.

In cell experiments, actions are simple: knock out a gene, add a compound, change the medium—repeatable, parallelizable without limit, with instant readouts.

In medicine, actions are another matter: medication, nutrition, exercise, surgery—each comes with safety constraints and tolerability boundaries, some are partially irreversible, and all are forever bound to the Objective: not "what was done," but "what it was done for."

More crucially, a medical Action must be bound to a "responding population." The same intervention works for some people and fails—or even harms—others; this is the most expensive watershed in clinical translation. So the core question the virtual patient must simulate is not "how potent is this molecule," but:

> Will this person respond to this intervention?

The second: the re-testing feedback loop.

Cells can be experimented on endlessly; the human body cannot. A person can only come back periodically to be re-tested.

It sounds like a constraint, yet it is precisely the data structure unique to the virtual patient. Predict that a person will respond; after one cycle of intervention, the person returns for re-testing: if the markers have truly improved, the prediction gains credit; if not, the model is corrected. "Predict → intervene → re-test → correct"—every re-test is a prospective, n=1 validation.

The sixteen-author review points out that since the EU's Avicenna action plan introduced the concept of in-silico clinical trials in 2016, prospective validation of efficacy prediction at the level of the individual patient has long been absent; and the "high performance" of most models derives precisely from information leakage caused by unreasonable splits between training and test sets.

> Re-testing is the virtual patient's weapon against paper metrics.

Testing yields data; re-testing is follow-up. Once this loop begins to turn, the virtual patient starts to evolve by itself—a property that simply making a virtual cell bigger can never provide.

4. The Economics: Why This Layer Is Worth the Most

With the methodology settled, one last account to settle: why is this layer worth betting on?

The yardstick offered by the sixteen-author review is unforgiving: the variable with the greatest influence on the total cost of bringing a drug to market is not the speed of preclinical discovery, but the Phase II success rate. Phase II is where efficacy first faces a real test in human beings, where patient heterogeneity is at its fiercest—the acknowledged valley of death of drug development. And behind every successfully approved drug stands a capitalized total cost approaching US$0.9 billion.

And for programs that stratify patients with biomarkers, the capitalized cost per successful approval is only half that of unstratified programs.

Half. This is not an optimization of a few percentage points in speed; it is an order-of-magnitude change in the cost structure.

Seen against this, the current state of the industry becomes somewhat ironic: AI applications are heavily concentrated in preclinical molecular discovery—because preclinical data are abundant, modeling is easier, and results arrive fast. The review calls this "looking for the keys under the streetlight": the keys were lost in the dark, but under the streetlight it is bright, so that is where everyone searches.

The virtual cell is that very bright streetlight. It has value, and it deserves to be built. But the most expensive cut in drug development is not made in the petri dish; it is made in the valley of death.

> The ruler is not under the streetlight; it is in the valley of death.

Conclusion

From virtual cell to virtual patient, what is missing is not compute, not data volume, not parameter scale.

What is missing is a layer of scale: the conversion layer from the cellular coordinate system to the human coordinate system—a module map at the scale of the human body.

What is missing is a semantics of action: not "perturb the cell," but "for this objective, do this to this person—will this person respond?"

What is missing is a structure of feedback: not endlessly repeatable experiments, but the closed loop of "predict → intervene → re-test → correct."

The virtual cell answers: after perturbation, what happens to the cell.

The virtual patient answers: after intervention, what happens to this person.

The missing layer in between is the entire battlefield.

References

1. Bunne C, Roohani Y, Rosen Y, et al. How to build the virtual cell with artificial intelligence: Priorities and opportunities. Cell. 2024;187(25):7045-7063. DOI: 10.1016/j.cell.2024.11.015.

2. Bender A, Thomas MC, Scannell JW, et al. Artificial intelligence in drug discovery — what it is, where we stand and the path forward. Nature Reviews Drug Discovery. 2026. DOI: 10.1038/s41573-026-01496-2.

3. Xiong J, Xia Q. Toward a Self-Learning AI Agent for Drug Repurposing: Building Human-Scale Representations for Virtual Patients. Preprints.org. 2026. DOI: 10.20944/preprints202608.0998.v1.

4. Xiong J. World Models for Biomedicine: A Steerability Framework. Preprints.org. 2026. DOI: 10.20944/preprints202605.0366.v1.