News Article

Building a virtual 4D cell

November 12, 2024
Conceptual virtual cell

“The dream as we’ve collected all this data has always been to build the first spatial model of a human cell. All of a sudden – right now – I feel like we might actually be able to do it.”
— Emma Lundberg, PhD, Stanford University

Emma Lundberg has never shied away from big, complex problems – not just big ideas, but the unimaginably vast amounts of data they require. For the last two decades, in addition to her academic research in the growing field of spatialomics, she’s spearheaded the open science Human Protein Atlas project with the goal of mapping all human proteins and sharing millions of high-resolution images with scientists around the world. 

“The dream in the back of my head as we’ve collected all this data has always been to build the first spatial model of a human cell,” she says. But until relatively recently, creating that kind of model simply wasn't feasible. “There were too many pieces in human cells, too many emergent properties,” she says. “But these are the exact features that make the models so valuable, so we could provide reliable predictions about how cells will behave.”

Handling cells’ many scales, enormous number of components and types of interactions and confoundingly non-linear responses – where a small change can have massive downstream effects – all add up to a seemingly impossible data processing challenge. But AI is changing the game entirely, offering a real path toward creating the virtual cell. 

“I’m very excited about where we are now, where we have both AI tools and better capacity for generating multimodal data. It’s really changed my view on being able to reach that vision. All of a sudden—right now—I feel like we might actually be able to do it.”

Putting biology in place

At the core of creating the virtual cell is understanding not just how proteins are shaped, but where they are located within the cell. For a long time, scientists operated with a tacit assumption that a protein’s shape has a one-to-one relationship with its job. This should mean that knowing a protein’s structure is enough to tell us where and how it functions within the cell. 

But it turns out that many proteins are more versatile than we first thought. 

“One of the most exciting discoveries in our field is that more than half of human proteins are in multiple places in the cell,” Lundberg explains. “If you think of the cell as a house, it would be like me moonlighting in both the kitchen and the laundry room at the same time, doing both of those different kinds of tasks. And if more than half of our proteins are capable of performing multiple functions, that really diversifies the functionality of proteome – and makes cells even more complex than we thought.” 

As if this 3D complexity weren’t enough, especially when it comes to using virtual models to design and test better drugs, we also need to factor in a fourth dimension: time. 

“A cell might react to a drug perturbation within a couple of minutes – or it could be days, or weeks, or a month,” Lundberg says. “I’m so used to thinking about the spatial axis and how cells are spatially organized. But the temporal dimensions are even bigger than the spatial dimensions.”

Bringing focus and clarity to cancer treatments

One especially pressing area for spatialomics is cancer drug development. Tumors are highly variable, not just from tumor to tumor but within each tumor itself. This variation in the tumor microenvironment leads to unpredictable clinical outcomes, including high failure rates during clinical trials. 

The research team at the Danaher Beacon for Spatialomics is applying the latest in spatial biology with cutting-edge AI to create the next generation of smart microscopes, all with the aim of making more precise and more predictable cancer drugs. Lundberg, alongside microscopy experts at Leica Microsystems and using reagents from Abcam, is hoping to develop an analysis engine that can pick up on small but crucial changes in the tumor microenvironment, whether spatial, proteomic or metabolic. 

“It’s important to build these spatially resolved models of cells so we can understand where the functions are happening and which function is disrupted, ultimately so we can make more informed predictions about how those tumors will respond to potential therapies,” says Lundberg. “It’s an exciting and important application of spatial biology and structural cell modeling.”

Data to AI and back again 

The burgeoning field of AI-powered big data projects is prompting a fundamental shift in the relationship between AI and scientific data. Instead of collecting data and analyzing later – and perhaps realizing the data aren’t suitable or are mismatched to the analysis tools – data collection and AI-driven analysis can inform one another. 

“Maybe we start out doing in silico experiments and generate synthetic microscope images before we go out to the microscope and generate real images or predict how certain mutations would cause dysfunction of cells before we go and measure them,” Lundberg offers. “It’s closing the loop between data generation and data analysis so we can get meaningful results faster.”

As the magnitude of our modeling challenges grows from folding proteins to creating virtual models of entire cells, so do the technical challenges – and the ways we think about bringing them together.

“Science is only becoming more complex in the spatial biology field,” Lundberg says. “And I think this requires us to think outside of the box also in terms of how we do science.”