Key Insights
- Spatial tools offer up biological insights with huge numbers of parameters.
- To turn raw data into insight, researchers rely on algorithmic data processing and machine learning.
- Builders of generative artificial intelligence models of biological systems are also eyeing these high-context data sources hungrily.
Spatial biology is a catchall term for the analytical techniques that can identify biomolecules in their tissue context. The powerful technologies that power these developments—such as spatial transcriptomics, light-sheet microscopy, and cryo-electron tomography—can collect a lot of information about the many molecules that make up a tissue. And where there are data, there are artificial intelligence models and applications.
People in the field say this moment is a synergistic one between the tools for observing complex biology and the tools for modeling and predicting it.
New instruments for integrative molecular imaging
According to Serge Saxonov, cofounder and CEO of 10x Genomics, the company’s founding premise is that to advance biology, researchers will need new tools for “measuring at very large scale at high resolution.” The key, he says, is to capture “the right objects in very large numbers.”
Last month, 10x launched a spatial-transcriptomics instrument dubbed Atera that can capture almost the whole human transcriptome at single-molecule scale in cross-sections of tissue. The instrument detects between several hundred and a few thousand transcripts per cell, depending on the cell size, using a probe set that covers roughly 93% of human protein-coding genes. Using that information, researchers can identify cell types and then figure out how they interact in space.
“Whole transcriptome [spatial analysis] is fundamentally very challenging to do,” Saxonov tells C&EN. “Doing it at high throughput was previously not thought of as being possible.”
“To build artificial intelligence that can accurately represent the full complexity of biology and accelerate scientific research, we need orders of magnitude more data than exists today.”
The challenge is that a single cell contains—depending on cell type—thousands of transcripts, often closely packed together in space. “That was one of the big challenges for us to solve using optics algorithms and chemistry all together,” Saxonov says.
Saxonov says there was no single breakthrough improvement that made the new instrument possible—instead, he says, the optimization was end to end, including probe chemistry, fluidics, and optics for superresolution imaging. 10x also engineered on-platform software to analyze the raw data that emerge.
The new 10x platform is just one example of how AI is being integrated into bioanalysis to manage the huge datasets that these instruments produce. According to Michael Snyder, a Stanford University biochemist who co-leads the Human BioMolecular Atlas Program (HuBMAP), most instrument vendors have some early analysis built in to their machines. “Cell annotation, defining a cell type, cell segmentation, building 3D maps: all of these are complex computational tasks,” Snyder tells C&EN. “We use AI for every one of those steps.”
Harmonizing different instruments’ pre-analyses can be tricky, Snyder says. But algorithms are indispensable for making sense of the giant spatial datasets that these instruments produce. To take 10x’s instrument as an example, Saxonov estimates that “a single run of Atera generates as much data as streaming Netflix at 4K for 12 years straight.”
Making sense of massive datasets
In a plenary lecture at the American Society for Biochemistry and Molecular Biology’s annual conference this spring, Eric Betzig, the slightly maverick 2014 chemistry Nobel laureate, described working with biologists who were imaging their systems with the light-sheet microscope that his research group built at the Howard Hughes Medical Institute (HHMI) Janelia Research Campus.
“They all go back with big smiles on their faces, and 10 TB of data on a drive—and 2 weeks later, they call us up crying because they have no idea what to do with 10 TB of data,” Betzig said, drawing a chuckle from his audience of biochemists.
Light-sheet microscopes like Betzig’s in his present lab, at the University of California, Berkeley, can collect images of very large tissues—even whole organisms—in three dimensions, rapidly and with minimal harmful illumination. That capability means the instruments can capture huge amounts of biological context. But they also produce a data-analysis challenge.
As Betzig said during his lecture, “The problem of where we’re at right now is that you can’t handle the complexity.”
Betzig has previously proposed a solution of centralized facilities that offer both vast spatial data collection and AI models that present the data in interpretable ways. “I was the last guy on earth who wanted to have anything to do with AI,” Betzig said. He doesn’t like to do what everyone else is doing. “But goddamn it, I’m forced to because there’s just nothing else possible. We have to build minds that can see across vast data and then report it back to us meat bags in the end, in a way that we can understand.”
In addition to using machine learning to make sense of their datasets, researchers also are using those data to train AI models.
Generating data to train generative models
“AI is the perfect tool for resolving the complexity of biology,” Pushmeet Kohli, the vice president for research at Google’s DeepMind, told the newsletter Decoding Science recently. But, he said, a key challenge is giving the model adequate information.
Crowdsourced protein-structure databases like the Protein Data Bank powered DeepMind’s famous AlphaFold model, and the team has relied on similar information in genomics. “But if you go up the complexity hierarchy and think about cells or whole organisms, the same kind of data does not exist,” Kohli said. “We will need to invest much more in curating high-quality data for those domains.”
Last month, Biohub, a research institute funded by Priscilla Chan and Mark Zuckerberg, launched the Virtual Biology Initiative, a 5-year project aiming “to galvanize a global effort to create the technologies and multi-modal datasets needed to build predictive models of life.”
Buried in the $500 million bet on AI for biology is a huge commitment to tooling.
“To build artificial intelligence that can accurately represent the full complexity of biology and accelerate scientific research, we need orders of magnitude more data than exists today,” Alex Rives, the Biohub head of science, says in a statement accompanying the launch.
To collect such data, Biohub has earmarked $400 million for technology development—a figure almost equal to the annual budget of the National Institute of Biomedical Imaging and Bioengineering.
In an email, Biohub spokesperson Allison K. Wilson tells C&EN that that commitment includes “developing cryo-electron tomography that will resolve atomic-level details in the cell, and microscopy that can observe millions to billions of cells in living tissues and organisms. We’re also building molecular, cellular, and tissue engineering technologies that will enable better experiments and the measurement of more parameters.”
Another $100 million will go to collaborative data collection. Biohub says it has already brought on board a variety of groups with biological atlas programs—research institutes that, like HuBMAP, aim to collect complete molecular descriptions of various tissues.
A bustling market for spatial data collection and modeling
Spatial data collection and modeling have captured analytical-instrument vendors’ attention. Major manufacturers best known for mass spectrometry, DNA sequencing, and fluorescence microscopy are converging on molecular tissue imaging to cash in on the spatial-biology gold rush.
“They all go back with big smiles on their faces, and 10 TB of data on a drive—and 2 weeks later, they call us up crying because they have no idea what to do with 10 TB of data.”
“The appetite for researchers to get and have more data is not slowing down,” Rob Tarbox, a vice president at Complete Genomics, a sequencing company that also offers spatial tools, told C&EN earlier this spring. In March, the AI company Bioptimus announced it would conduct multimodal profiling of 100,000 patient specimens to train its model.
And access to the resulting models is also fueling deals. For example, GSK paid $50 million to license foundation models that the start-up Noetik built based on spatial cancer data in January of this year.
In a separate, open-source project, researchers at Genentech and Stanford launched SpatialAgent about a year ago. The AI model is trained on spatial transcriptomic data and can identify cell types and tissue niches based on transcriptome images, design gene panels for new assays, and help researchers zero in on the specific cell subtypes or interactions that cause problems.
In a preprint published before peer review, the authors conclude that their model “matched or outperformed human scientists across key tasks” in hypothesis generation and discovery in spatial biology and may resolve analytical bottlenecks.
Saxonov says researchers can use AI tools to run analyses that once “required deep bioinformatics expertise.”
With instruments that demand fewer trade-offs between throughput, sensitivity and choice of analytes offering ever-more-complex data, Saxonov says, he hopes that machine-assisted interpretation these complex data will clear the way to more-straightforward insights. “Once you actually see the full picture, my view is that interpretation will actually, in many ways, become more straightforward.”
