The future of precision medicine will not be won at the level of the cell only.
It will be won at the level of the patient.
Over the past two decades, biology has advanced by moving up layers of representation. First, we learned to read genes. Then, we learned to classify and map cells. Today, with the convergence of single-cell biology and AI, the field is attempting its next leap: building increasingly detailed computational models of biology.
Much of the current excitement centers on the idea of the “virtual cell.” Academic initiatives such as the Human Cell Atlas, along with industry efforts from groups like the Chan Zuckerberg Initiative and Recursion, are working to model how individual cells behave and respond to perturbations. By combining multimodal measurements with large-scale experimental data, these approaches aim to predict how molecular networks inside a cell change when genes are altered or when drugs are applied.
This work is ambitious, necessary, and already productive. But it also reveals a deeper limitation, not of execution, but of scope.
Cells are not where meaning ends
Cells matter. But cells are not where biological meaning ends.
Biology is full of shared words with incompatible meanings. A molecular response, a cellular response, and a clinical response share a word, but not the deeper meaning. A cell state, a system state, and a disease state are not merely different resolutions of the same phenomenon; they are fundamentally distinct.
Virtual-cell models excel at learning how cells respond under controlled conditions. However, disease and therapeutic success are not properties of isolated cells. They are emergent properties of biological systems operating in patients over time.
This distinction is not semantic. It is foundational.
Why virtual cells were the right first step
There is a good reason the field has focused on cells. From a practical perspective, they are tractable. Cells can be generated at scale, perturbed in controlled environments, and measured repeatedly. Cause-and-effect relationships can be probed directly. Data generation, while expensive, is feasible.
In that sense, virtual-cell efforts represent a natural continuation of the Human Cell Atlas, transitioning from cataloging cellular identities to modeling cellular behavior. They begin with what is experimentally and computationally possible.
But feasibility is not the same as necessity.
Precision medicine has not stalled simply because we lack cellular detail. It has stalled because we do not understand how therapies reshape biological systems in patients, and because we lack the data required to learn that.
The harder problem: The clinical realm and “virtual patients”
In the real world, the most important perturbation is not a gene knockout or a cytokine added to a dish. It is a drug given to a human being. The readout is not a cellular phenotype; it is a response, resistance, relapse, or survival.
Consider immune checkpoint inhibitors. The two best-selling cancer drugs in the world, Keytruda from Merck and Opdivo from BMS, both work by blocking PD-1, a brake on the immune system. In simple lab experiments, testing these drugs on isolated immune cells often shows limited or highly variable effects. Yet in patients, the same drugs can lead to dramatic and long-lasting tumor responses and have helped save millions of lives, or provide little to no benefit at all. That gap cannot be explained by a single cell. It reflects the behavior of the immune system as a whole, shaped by its environment and its history over time.
Clinical response is the most information-dense signal biology has. It reflects the integrated effect of a therapy across tissues, cell types, and time. Any model that aspires to guide precision medicine must be anchored to that signal.
This is where the concept of the virtual patient becomes essential.
A virtual patient is not a simulated cell, or even a collection of cells. It is a system-level representation that links biological state to clinical state. Building such representations is far more challenging than building virtual cells. It requires longitudinal sampling, deep clinical annotation, and large, diverse patient cohorts, often numbering in the tens or even hundreds of thousands.
This challenge is not merely logistical. It is clinical. Without outcome-linked data, models may be elegant, but they remain disconnected from the decisions that matter most.
Systems, not components
The immune system makes this limitation especially clear. Immune cells do not act independently; they coordinate. They exchange signals, form feedback loops, expand and contract, and transition between functional states. Small perturbations can cascade into system-wide shifts. Therapies succeed or fail not because they correct a single component, but because they move the system from one dynamic regime to another.
From this perspective, cells provide vocabulary. Understanding patients requires language.
Discriminative models help represent and stratify observed patients, grouping biological systems into meaningful states and predicting likely outcomes. Generative models can learn the distributions and transitions that govern system behavior, enabling the simulation of plausible patient trajectories under different interventions.
Together, these approaches shift biology from description to navigation.
What needs to change
If the next leap in precision medicine requires virtual patients, then behavior must change accordingly.
Today, much of the field, across industry, government funding agencies, pharma, and philanthropy, is still waiting for definitive proof points before investing at scale in large, clinically annotated patient datasets. Single-cell and multi-omic datasets remain sparse, fragmented, and often disconnected from robust clinical outcomes.
This creates a familiar chicken-and-egg problem. Without large, high-quality, outcome-linked datasets, system-level models cannot mature. Without mature models, investment remains cautious.
Breaking this cycle requires initiative. It requires treating large-scale, clinically annotated patient data not as a downstream application, but as core scientific infrastructure, on par with the Human Genome Project or the Human Cell Atlas.
The next evolution
Seen through this lens, the trajectory of modern biology is clear: from genes, to cells, to systems, to patients. Each step builds on the last, while also revealing its limits. Virtual cells are an important milestone in this progression, but they are not the destination.
The virtual cell atlas was a necessary step. It is now time to construct the virtual patient map.
This entails coordinated, large-scale efforts across academia, industry, government, and philanthropy to generate deeply profiled, clinically annotated patient datasets. Only then will we be able to move beyond cellular biology toward true system-level understanding.
Virtual cells can teach us cellular biology.
Virtual patients will teach us why therapies succeed or fail.
Precision medicine will be won at that level.
The writer is the CEO of the American-Israeli biotech firm Immunai.