Human cells operate like intricate Rube Goldberg devices, powered by biological chain reactions that dictate life and death. A core interest in biology lies in comprehending these intricate connections and their dysfunction in disease. A solitary genetic error can distort a protein’s structure, rendering it ineffective. Without that crucial protein, the organism—you—could begin to deteriorate.

However, cells are incredibly complex, making it difficult to grasp how a single protein’s malfunction might propagate throughout the system. Graham Johnson, a computational biologist and scientific illustrator at the Allen Institute for Cell Science, remembers envisioning, over 15 years ago during lunch, a computer model of a cell so exhaustive and precise that researchers could observe these processes in action. Back then, he notes, “everyone merely scoffed.” “It was simply too far-fetched.”  

Currently, however, some scientists are leveraging AI to advance towards creating a “virtual cell.” Google’s DeepMind is involved, and the Chan Zuckerberg Initiative (CZI) has significantly prioritized virtual cells within its Biohub research network, according to Theo Karaletsos, CZI’s senior director of AI. The Arc Institute has also established a for models akin to virtual cells. The collective aim of these efforts is to forecast the operations of both healthy and diseased cells with such granularity that it can hasten drug development and scientific breakthroughs. Some even propose that virtual cells could streamline fundamental research, shifting biologists from laboratory experiments to computer-based work.  

What is a virtual cell, anyway?

The exact interpretation of a virtual cell differs among experts. Certain scientists, such as Johnson, anticipate a virtual cell featuring an interactive visual display for exploration. Others conceive of it primarily as a suite of computer applications capable of addressing queries and forecasting probable events. Nevertheless, this concept isn’t novel. For many years, biologists have been constructing mathematical models of cellular activities. To develop these, scientists utilize experimental data from actual cells to formulate equations depicting their functions. 

Currently, an unprecedented volume of data on human cells is available, partly due to technologies enabling scientists to monitor individual cell activities. However, devising equations for every process and integrating them is an enormous undertaking. Stephen Quake, a Stanford University professor and former head of science at CZI, states, “The traditional method”—meaning manual approaches—”achieved, I would contend, only very limited triumphs.” Last year, he and fellow researchers presented a proposal for an alternative strategy, involving feeding cellular data directly into specialized AIs. “You develop models that learn straight from data, instead of attempting to codify equations,” he explains.

Quake and his associates have achieved They trained an AI using cellular data from 12 distinct species. The AI subsequently demonstrated accurate predictive capabilities regarding cells from species it had not previously encountered, Quake reports. Furthermore, it could deduce relationships between various cell types within a single species, even without explicit information about these connections. “That’s what, for me personally, made this method extremely exciting,” Quake remarks.

A separate research group, which includes members from Google DeepMind, is focused on They have prepared AIs using extensive datasets of cellular information, enabling users to pose queries such as, “How will this cell react to this medication?” and subsequently obtain responses indicating which parts of the cell are probable to be impacted. 

These represent merely a fraction of the methodologies scientists are employing to construct virtual cells. It’s probable that numerous distinct types of virtual cells will ultimately emerge, tailored for diverse researchers. For example, a virtual cell utilized by a cancer biologist might differ from one employed by a cell biologist investigating the evolution of a particular structure. It’s also conceivable they could integrate both conventional modeling techniques and AI.

What virtual cells might allow us to do

Virtual cells have the potential to expedite and simplify the discovery of novel drugs. They could also offer understanding into how cancer cells bypass the immune system or how an individual patient might react to a specific treatment. Furthermore, they might assist fundamental scientists in formulating hypotheses about cellular functions, guiding them on which experiments to conduct with actual cells. “The primary objective here,” Quake states, “is to endeavor to transform cell biology from a domain that is 90% experimental and 10% computational to the reverse.” 

Some scientists raise concerns regarding the utility of AI-generated predictions if the AI cannot offer an explanation for them. Erick Armingol, a systems biologist and post-doctoral researcher at the Wellcome Sanger Institute in the U.K., notes, “Typically, AI models operate as a black box.” This implies they furnish an answer without being able to articulate the rationale behind it.

“For me, my motivation for entering this field was to model the entire human body and observe how cells interlink and engage. That is the ultimate goal,” he shares. While opaque answers could be beneficial for guiding drug creation, they might prove less valuable for fundamental scientists—at least with how many current AIs are configured. (Karaletsos of CZI clarifies that some of their AIs are designed to offer explanations for their deductions. “Our aim is to comprehend, not merely to forecast,” he asserts.)

Johnson, who penned an article emphasizing the significance of developing virtual cells, anticipates that any eventual scientific creation will feature visualization capabilities. His vision is “a visual, interactive, intuitive rendition of a complex subject,” he states. “I believe AI is unequivocally crucial for facilitating all of this. I simply am not concerned with black-box predictions as the main result.”

Irrespective of their construction method, it could be some time before virtual cells of any type become operational. “This is not an accomplishment slated for the coming year,” Quake comments. “I anticipate it will require a complete decade to fulfill its promise.”

However, since that distant lunchtime conversation, Johnson remarks that progress in cell biology and computer science has fundamentally altered the outlook for eventually achieving a virtual cell. “I no longer perceive myself as a madman merely espousing this,” he says. “It now appears achievable.”