Biologists are generally not fancy people, but when it comes to cryo-electron tomography, it can help to add a little bling.  

Cryo-electron tomography (cryo-ET) images cells in their native environment in three dimensions, providing 3D views of biological structures at nanometer resolution. The technique does a great job at revealing large structures, like cellular organelles, but is more complex to use for distinguishing the individual molecules that make them up — information that scientists use to understand how the larger structures function.  

To overcome this, researchers add small gold nanoparticles that attach to the molecules they want to see. But finding these small gold particles in the noisy images can be difficult and time-consuming.  

HHMI Investigator Eric Gouaux likens it to Sherlock Holmes and Doctor Watson searching for Professor Moriarty on a dark, foggy night on a crowded London street — they can track Moriarty’s lantern, but the signal is faint and it’s hard to discern the criminal mastermind’s features.  

Now, unlike Holmes and Watson, scientists can enlist artificial intelligence for help.  

Using AI to Uncover Molecular Details in Cryo-ET 

Gouaux, who leads a lab at Oregon Health and Science Universityexternal link, opens in a new tab, and HHMI Investigator Michael Rosen at UT Southwestern Medical Centerexternal link, opens in a new tab are leading a project to use AI to transform cryo-ET from its traditional use in imaging big things — like cells and the structures inside them — to a technique that can also be used to examine small things – like the molecules making up these cellular structures.  

Understanding where these molecules are and how they’re organized allows biologists to uncover how the larger structures work and what happens when they go awry.  

For Rosen, understanding how chromatin — densely packed DNA — is organized in the nucleus helps his team uncover how it contributes to functions like gene expression. Studying the organization of molecules in synapses — the sites where messages are passed between neurons — allows Gouaux and his team to understand how signal transmission happens.  

“Ultimately, we are really interested in: Does Moriarity have a knife? Does he have a gun? How can we figure that out?” Gouaux says. “And so, with respect to these assemblies, we really want to know how they are fitting together so we can understand, in detail, what’s going on.” 

Training AI Models to Find Molecules  

Using cryo-ET data obtained by the researchers, the team plans to develop and train AI models to quickly and accurately recognize gold nanoparticle labels of different shapes and brightnesses, ramping up the speed and sensitivity of the process beyond what the human eye can capture. 

The researchers also want to increase the amount of gold nanoparticle labels they can use and employ different types of labels that can be used with other imaging techniques, both of which could reveal additional molecular details, says HHMI Investigator Elizabeth Villa, who is collaborating on the project.

Tomography is the highest resolution imaging tool biologists have but it is hard to identify molecules of interest, so improving these labeling methods and making them easier to use will be a game changer for researchers, Villa says.

“I can’t think of a single person that uses tomography right now that wouldn’t want a plug-and-play version of this and I can think of a lot of people that are not experts in tomography that would love to use tomography,” she says.

The team also plans to create and train AI models on biochemical reconstitutions of dense chromatin as well as molecular simulations of chromatin created by Rosen’s collaborator Rosana Collepardo-Guevara at the University of Cambridge. Information from the biochemistry and simulations will help improve the model’s predictions of where the chromatin is in the tomograms, how different elements are connected to each other, and how changes, like mutations, disease, and aging, can affect its mechanics and function. 

These AI-powered methods both improve the quality of the cryo-ET images — providing better reconstruction of the molecules and structures in the sample — and the analysis, enabling more information to be obtained from the data, says Magdalena Schneider, a Machine Learning Researcher with the AI@HHMI Initiative at HHMI’s Janelia Research Campusexternal link, opens in a new tab, who is collaborating on the project.  

“We can see things that we haven’t seen before in the natural environment of the cell,” she says.  

Taking Tomography to the Next Level  

Ultimately, the team hopes to design AI tools that can be used by biologists in labs worldwide to understand many different types of molecular structures beyond synapses and chromatin.   

“If the AI can produce a generalizable approach to labeling multiple objects differently, and then rapidly identifying where they are and determining their structures, that will be transformative in a lot of different areas,” Rosen says.

The researchers say the AI@HHMI Initiativeexternal link, opens in a new tab allows them to accomplish much more than they could in their own labs by enabling them to collaborate with experts in AI, cryo-ET, and annotation, which are all housed under one roof at HHMI’s Janelia Research Campus. 

“It just wouldn’t have been possible without this,” Gouaux says. “It’s absolutely transformative.”