‘Virtual Cells: Transforming Raw Data into Living Predictive Models of Biology’

Any gamer can tell you that computers are remarkably good at mimicking almost everything—from the daily chores of running a household to the complex challenges facing a civilization spread across multiple planets. You’d think recreating life’s simplest building block—the cell—would be effortless by comparison. It isn’t.

Inside every cell lies a vibrant community of biological molecules, constantly interacting and responding to signals from their surroundings in ways scientists are still working to unravel. What holds true for one kind of cell may not apply to another. Yet beneath the apparent disorder lies a hidden logic.

“A cell is an intricate system—one that is remarkably tough and adaptable,” explains Emma Lundberg, a bioengineer at Stanford University in California. “Yet it’s also highly organized—cells possess their own internal architecture.” In recent years, scientists have been working to decode that architecture, transforming enormous collections of molecular information into “virtual cells”—digital models that recreate what happens inside cells, both during normal operations and when they face outside pressures.

Multiple research groups are now drawing from extensive pools of transcriptomic readings (showing which genes are active) and related data to construct models that might uncover the root biological causes of illness and suggest new treatment approaches. “Virtual cells should serve a specific purpose, and for me, that purpose is speeding up how we generate and test scientific hypotheses,” says Yusuf Roohani, a machine-learning specialist at the Arc Institute in Palo Alto, California.

Still, the field is nowhere near producing a fully operational virtual cell. “Nobody who’s being honest would say they’ve built a virtual cell—unless they’re trying to pitch a start-up,” remarks Fabian Theis, a computational biologist at Helmholtz Centre Munich in Germany. Today’s models can depict fixed cellular conditions but have difficulty forecasting how cells shift and adapt over time. Advancing to more sophisticated computer-simulated biology will demand ever-larger quantities of varied data and clever methods for merging them.

Building on solid ground

The surge in artificial intelligence has stoked excitement about virtual cells, but researchers have wrestled with creating computational cell models for many years. “Two decades ago, we already had ‘virtual cell 1.0,’ where scientists attempted to map out systems biology using mathematical equations,” recalls Bo Wang, an AI expert at the University of Toronto in Canada.

‘World models’ represent AI’s newest breakthrough: understanding their potential

These mathematical models benefit from being rooted in concrete, well-tested biochemical and biophysical laws—linking equations together that capture how cells carry out tasks like metabolism, signaling, and motion. “You gain genuine insight into the mechanics—you can make sense of what’s happening, and that’s highly appealing,” Lundberg notes.

For example, a team headed by Zaida Luthey-Schulten at the University of Illinois at Urbana-Champaign unveiled an advanced mathematical framework in March that accurately reproduced cell division in an engineered strain of Mycoplasma bacteria¹. Meanwhile, Paul Macklin, an engineer at Indiana University in Bloomington, and his colleagues have invested over ten years in crafting a platform named PhysiCell, designed to model how human cells and tissues react to various environmental factors. According to Macklin, this tool has proven valuable for studying cancer, including the forces that fuel tumor growth or influence how patients respond to immune-based treatments.

Despite these achievements, mathematical models face fundamental constraints tied to how much researchers actually grasp about cell biology. Projects like the Human Cell Atlas have generated massive libraries of gene-expression information and other molecular data, covering proteins and epigenetic markers—but making biological sense out of thousands of molecular connections remains incredibly challenging. This is where AI excels, according to Maria Brbić, an AI researcher at the Swiss Federal Institute of Technology in Lausanne: “AI systems are exceptional at navigating vast combinations of possibilities.”

Experts disagree on what features a genuine virtual cell should possess, but a worthwhile simulation ought to capture at least the standard condition of a particular cell type, then forecast how a specific disruption changes that condition. Numerous efforts have drawn on deep-learning “foundation models,” where AI programs detect hidden patterns within massive pools of unlabeled experimental data.

Can AI construct a virtual cell? Researchers compete to map life’s basic unit

Roohani compares this to ChatGPT, a chatbot driven by a foundation model that learned patterns from online text to generate sensible replies to nearly any question. “You develop more versatile representations applicable across a wider spectrum of cellular and biological situations,” he explains. Ideally, a biological foundation model could predict how different cell types would behave under circumstances absent from its training data, and even offer reliable forecasts for entirely unfamiliar cell types.

Single-cell gene-expression datasets are currently the favored resource for exposing biological foundation models to various cell types, and such information is plentiful. Roohani’s team has built a repository named scBaseCount, which employs AI to continuously gather and standardize transcriptomic data for training purposes. The database holds roughly half a billion cells—and continues growing. “That’s several multiples larger than the next biggest comparable collection,” Roohani states.

However, a meaningful representation cannot rest only on a cell’s core characteristics—what AI models refer to as embeddings. A virtual cell must additionally grasp how different disturbances reshape the cellular landscape. Filling in these specifics calls for carefully controlled experiments where scientists selectively disable individual genes or treat cells with a broad spectrum of chemicals. “To build models that reveal cause and effect, we need data that demonstrates cause and effect,” Wang emphasizes. One notable example is the X-Atlas/Pisces dataset, assembled by Xaira Therapeutics, a pharmaceutical firm in South San Francisco, California. Shared openly via HuggingFace—an open-source AI hub—Pisces contains gene-expression profiles from 25.6 million cells spanning multiple lineages, each subjected to precise genetic modifications.

The pitfalls of perturbation

In theory, these models could help researchers deduce which genetic defects propel a specific cancer’s development or identify drug classes capable of correcting metabolic disturbances in compromised cells. Several foundation models are approaching this level of capability.

For instance, in January, Roohani and his team presented Stack², a model trained using the scBaseCount dataset.

Using this information, the team created a comprehensive reference map—called a ‘perturbation atlas’—that forecasted how various drug treatments would impact 28 different human tissue types. In March, Xaira unveiled its X-Cell model³, developed using the company’s proprietary Pisces dataset. Wang, who leads biomedical AI research at Xaira, reports that X-Cell accurately predicted the gene expression shifts associated with T cell activation, even without specific training on that process. This capability enabled researchers to identify possible mechanisms for deactivating these immune cells—a breakthrough that could lead to new treatments for inflammatory diseases and other immune-related conditions. “We validated established T cell inhibitors like CD3 and related molecules, and also discovered several promising new candidates that may suppress T cell activity,” explains Wang.

Predicting cellular responses to interventions remains a significant scientific challenge. Wang notes that current AI models represent only the first steps toward this goal. “Currently, research has been limited to cell lines, which are simplified models of real biological systems,” he explains. These models often fail to accurately represent how actual organs and tissues behave. Moreover, gathering sufficient training data directly from human primary cells at a useful scale presents major practical difficulties.

Scientists have also found it difficult to show that large-scale gene-expression models consistently outperform traditional mathematical approaches. The Arc Institute organized the Virtual Cell Challenge in 2025, offering competing teams a chance to directly compare their models’ predictive accuracy. While the event generated impressive participation—with around 5,000 participants from over 100 countries, according to Roohani—purely AI-driven models failed to outperform hybrid approaches that incorporated classical statistical techniques.

Brbić has encountered similar challenges when evaluating the reliability of deep-learning models. She identifies a key limitation in standard evaluation methods, which often focus on detecting overall differences in gene activity between treated and control cells. This approach can cause subtle but important biological changes to be obscured by random experimental variability, complicating computational analysis. “Single-cell gene expression data contains significant technical noise,” says Brbić. “Observed differences may reflect genuine biological responses, but they could also stem from laboratory procedures or other confounding factors.”

To address this, Brbić’s team introduced Systema in 2025, a benchmarking framework that filters out technical noise to highlight true perturbation-related changes in gene activity⁴. Similarly, Roohani’s predictive model, State, is designed to account for natural variation between individual cells within populations⁵. When State was assessed using an evaluation method—like Systema—that emphasizes perturbation-specific effects rather than total gene expression changes, it correctly predicted approximately one-third of the genes most impacted by a given treatment in a validation dataset. This marks a substantial improvement over the 7% accuracy rate achieved with standard evaluation methods.

Filling in the Gaps