Why AI Companies Are Racing to Build a Virtual Human Cell

A human cell is a Rube Goldberg machine like no other, full of biological chain reactions that make the difference between life and death. Understanding these delicate relationships and how they deteriorate in illness is one of the central fascinations of biology. A simple error in a gene can cause the protein it produces to be shaped incorrectly. A distorted protein cannot do its job. And without this protein, the body – you – may start to break down.

Cells are so complex, however, that it is difficult to understand how the failure of one protein propagates through the system. Graham Johnson, a computational biologist and scientific illustrator at the Allen Institute for Cell Science, remembers fantasizing at a lunch table more than 15 years ago about a computer model of a cell so detailed, so complete, that scientists could watch such processes occur. At that time, “everyone was snickering,” he says. “It was just too unrealistic.”

But today, some researchers are using AI to take new steps toward the goal of a “virtual cell.” Google’s DeepMind is working on such a project, and the Chan Zuckerberg Initiative (CZI) has made virtual cells a major focus of its Biohub research network, says Theo Karaletsos, senior director of AI at CZI. There is even a new award, created by the Arc Institute, for virtual cell-type models. The goal of all these efforts is to predict how healthy and diseased cells function, in such detail that it is possible to accelerate drug development and scientific discoveries. Virtual cells could even streamline basic research, some believe, by moving biologists from the lab to the keyboard.

By the way, what is a virtual cell?

The precise definition of a virtual cell varies depending on who you talk to. Some scientists, like Johnson, hope that a virtual cell will include a visual representation that you can click on and explore. Others see it primarily as a set of computer programs capable of answering questions and making predictions about what is likely to happen. But the concept is not a new idea. For decades, biologists have been building mathematical models of cellular processes. To achieve them, researchers rely on data from experiments with real cells and propose equations describing what happens.

There is now more data than ever on human cells, thanks in part to technology that allows scientists to spy on the activities of individual cells. But finding equations for each process and putting them together is a monumental task. “The old way of doing it,” that is, manually, “has had, I would say, very limited success,” says Stephen Quake, a professor at Stanford University and former scientific director of CZI. Last year, he and other researchers published a paper laying out a vision for another approach, one that would feed cell data directly to specialized AIs. “You build models that learn directly from the data, rather than trying to write equations,” he explains.

Learn more: Why baby boxes are suddenly everywhere

Quake and his colleagues have obtained interesting first results. They used data on cells from 12 different species to train an AI. The AI ​​was then able to make accurate predictions about the cells of species it had never seen before, Quake says. He was also able to infer the relationships between different cell types within the same species, even though he had no information about these connections. “That’s what got me personally excited about this approach,” Quake says.

Another team of researchers, including some at Google DeepMind, is also exploring the use of AI to create virtual cells. They trained AIs on large datasets containing information about cells, allowing users to ask questions like: “How will this cell react to this drug?” » then receive answers about which parts of the cell may be affected.

These are just some of the approaches scientists are taking to create virtual cells. It is likely that there will eventually be many different types of virtual cells, designed for different types of researchers. The virtual cell used by a cancer biologist, for example, may be different from the one used by a cell biologist seeking to answer questions about the evolution of a given structure. And it is possible that they will use both traditional modeling approaches and AI.

What virtual cells could allow us to do

Virtual cells could facilitate and accelerate the discovery of new drugs. They could also provide insight into how cancer cells evade the immune system or how an individual patient might respond to a given treatment. They could even help fundamental scientists formulate hypotheses about how cells work, which could guide them toward experiments to be performed with real cells. “The overall goal here,” Quake says, “is to try to transform cell biology from a field that is 90 percent experimental and 10 percent computational to the other way around.”

Some scientists wonder how useful the predictions made by AI will be if AI can’t explain them. “AI models are normally a black box,” says Erick Armingol, a systems biologist and postdoctoral researcher at the Wellcome Sanger Institute in the United Kingdom. In other words, they give you an answer, but they can’t tell you why they gave you that answer.

Learn more: 11 Things Therapists Wish Every Child Knew

“Personally, the reason I ended up in this field is because I wanted to simulate the human body as a whole and how cells connect to each other and interact. So that’s my dream,” he says. Black box answers could be useful for driving drug development, but they might not be as useful to fundamental scientists – at least not in the way that many AIs are currently implemented. (CZI’s Karaletsos says some of their AIs are configured to provide explanations for their reasoning. “We want to understand, not just predict,” he says.)

Johnson, author of a 2023 paper on the importance of building virtual cells, hopes that whatever scientists end up building can be visualized. His ideal is “a visual, interactive and intuitive version of something complicated,” he says. “I think AI is absolutely essential to enabling all of this. I’m just not interested in black box predictions as the primary outcome.”