How AI is advancing cryoelectron tomography

Biologists aren’t generally flashy people, but when it comes to cryo-electron tomography, it helps to add a little glamor.

Cryo-electron tomography (cryo-ET) images cells in their native environment in three dimensions, providing a 3D view of biological structures at nanometer resolution. The technique is great for revealing large structures like organelles, but it’s more complex to use to distinguish the individual molecules that make them up, information that scientists use to understand how large structures work.

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

HHMI investigator Eric Gouault likens it to Sherlock Holmes and Dr. Watson searching for Professor Moriarty on a dark, foggy night in the crowded streets of London. They are able to track Moriarty’s lantern, but the signal is weak and it is difficult to discern the criminal mastermind’s features.

Unlike Holmes and Watson, scientists can now turn to artificial intelligence for help.

Uncovering the molecular details of Cryo-ET using AI

Mr. Guo heads the lab at Oregon Health & Science University.External link, opens in new taband HHMI investigator Michael Rosen of UT Southwestern Medical Center.External link, opens in new tab They are leading a project that uses AI to transform cryo-ET from its traditional use of imaging large things, such as cells and their internal structures, to a technology that can also be used to examine small things, such as the molecules that make up these cellular structures.

Understanding where these molecules are located and how they are organized allows biologists to uncover how larger structures work and what happens when they fail.

For Rosen, understanding how chromatin (the densely packed DNA) is organized in the nucleus helps reveal how chromatin contributes to functions such as gene expression. By studying the composition of molecules within synapses (where messages are passed between neurons), Guo and his team are now able to understand how signal transmission occurs.

“Ultimately, what we’re really interested in is whether Moriarity has a knife, whether he has a gun, and how do we figure that out,” Gouor said. “So when it comes to these gatherings, we really want to know how they work together to understand in detail what’s going on.”

Training an AI model to find molecules

The research team plans to use the cryo-ET data acquired by the researchers to develop and train an AI model that quickly and accurately recognizes gold nanoparticle labels of various shapes and brightnesses, increasing the speed and sensitivity of the process beyond what the human eye can capture.

The researchers also hope to increase the amount of gold nanoparticle labels available and employ different types of labels that can be used with other imaging techniques, both of which have the potential to reveal additional molecular details, said Elizabeth Villa, an HHMI investigator collaborating on the project.

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

“I don’t think there’s anyone who uses tomography today who wouldn’t want a plug-and-play version of this, and I think there are a lot of people who don’t need tomography experts who want to use it,” she says.

The team also plans to create and train an AI model for the biochemical rearrangement of dense chromatin, as well as molecular simulations of chromatin created by Rosen’s collaborator Rosana Collepardo-Guevara of the University of Cambridge. Information from biochemistry and simulations can help improve models’ predictions about where chromatin lies in the tomogram, how its various elements are interconnected, and how changes such as mutations, disease, and aging affect its mechanisms and function.

These AI-powered techniques improve the quality of cryo-ET images, better reconstruct the molecules and structures in the sample, and improve both the analysis and the ability to get more information out of the data, said Magdalena Schneider, a machine learning researcher in the AI@HHMI initiative at the HHMI Janelia Research Campus.External link, opens in new tabpeople who are collaborating on the project.

“In the natural environment of a cell, you can see things you’ve never seen before,” she says.

Taking tomography to the next level

Ultimately, the team hopes to design AI tools that biologists in labs around the world can use to understand many types of molecular structures beyond synapses and chromatin.

“If AI can generate a generalizable approach to label multiple objects differently, quickly identify where they are, and determine their structure, it will be transformative in many different fields,” Rosen says.

According to researchers, the AI@HHMI initiativeExternal link, opens in new tab Having the ability to collaborate with experts in AI, cryo-ET, and annotation, all coming together at HHMI’s Janelia Research Campus, will allow us to accomplish much more than we could in our own labs.

“Without this, it wouldn’t have been possible,” Guo says. “It’s completely transformative.”