ABOVE: 3D rendering of a HeLa cell: plasma membrane (brown), ER (green), mitochondria (orange), nucleus (purple), Golgi (blue), endosomes (cyan), vesicles (red), lysosomes (yellow), lipid droplets (pink), microtubules (dark sticks), and ribosomes (pink haze) COURTESY OF STEPHAN SAALFELD

To fully understand how cells work, scientists need to know how their moving parts relate to one another in space and time. However, because of their size and the amount of data involved, visualizing cellular structures in three dimensions has proven difficult. Now, in a trio of new studies, two teams of molecular scientists have aimed to make it easy for everyone to see inside cells. By incorporating painstakingly collected experimental data and partnering with computational biologists, they are bringing 3D visualizations of organelles and chromosomes into sharper focus. 

The researchers are also making their 3D data, published in separate studies in early October, freely available for anyone to explore in order to allow researchers around the globe to probe their own questions about how cellular form impacts function. As Karissa Sanbonmatsu, a structural biologist at Las Alamos National Laboratory and coauthor on one of the papers, puts it: “We’re trying to do Google Earth for chromosomes.”

Using the techniques from these three papers, computational biologist Robert Murphy of Carnegie Mellon University, who wasn’t involved in the research, says scientists might compare different cell types—for example, a cancer cell and a healthy cell—to start to understand what role the organization of cellular structures plays in physiology and disease. “That’s one of the very first things you would want to do.”

Improving organelle clarity, with a little help from AI

For the past decade, Howard Hughes Medical Institute (HHMI) scientist Shan Xu has worked to adapt focused ion beam scanning electron microscopy (FIB-SEM)—a microscopy technique originally developed for material science applications—to biological research. FIB-SEM works by taking an SEM image of objects embedded in resin, then shaving a tiny sliver of the sample using the ion beam and taking another picture. By repeating this process over and over, scientists can stack all the images to create a 3D rendering. But the machines need to shut down after 3 to 5 days to recharge the ion beam, and when they start up again, the amount of material the beam shaves off and the resolution of the image aren’t as accurate. To reduce the inaccuracy after each shutdown and achieve the clarity needed to map the insides of cells, Xu added more stable ion beam hardware that allowed a higher current, troubleshot the control of the beam, and sped up the imaging of the SEM. Xu constructs each of the machines himself. “I build them one by one, so [they’re] like my kids,” he says. 

Those tweaks, detailed in one of his team’s two October Nature papers, improved the 3D resolution to 4 nm from 8 nm, which often meant the difference between organelles looking like clear, detailed structures and fuzzy clouds.

But the resulting volume of data meant that the researchers needed a faster way to identify and map out the organelles inside the cells: enter machine learning. For two years, two people worked full-time to manually identify organelles and outline their boundaries in 3D images from FIB-SEM. Then, Larissa Heinrich, a computer scientist at HHMI Janelia Research Campus, used those annotations to train a neural network to map structures within the cells, as reported in the team’s second Nature paper. 

Heinrich says that the network uses the manually annotated images to learn rules, “trying to adjust them in a way that the output it produces is the same as what the humans did.” The network doesn’t just look at each pixel and make a call about whether it’s part of an organelle; it examines the pixels around it to determine whether the call is logical. The scientists estimate it would have taken one person 60 years to manually identify the same number of organelles the algorithms can map in a few hours.

Murphy says the results of the two studies show “a critical instrumentation advance,” adding that “the use of that technology to produce the large-scale data collections that they’ve done is important.” He says that the machine learning work is vital, but notes that the algorithms still can’t identify every type of organelle with a high level of accuracy. The discrepancy in identification is often related to how abundant the organelles are in the human-labeled training sets. For example, centrosomes are rare in each cell, so the AI doesn’t have as many chances to learn what they look like from the training sets. Heinrich says that more training sets and algorithms will help improve the mapping accuracy for all organelles and structures present inside cells.

Still, with the high-resolution images and the human- and AI-identified organelles, the researchers were able to build open-access 3D atlases of several cells and tissues, including commonly used HeLa cells, immune T-cells attacking ovarian cancer cells, and pancreatic beta-cells.

Adding the fourth dimension

In the third study, published in PNAS, a separate research group used computational approaches to infer the 3D structure of chromosomes. 

In the past, to infer 3D DNA structures, scientists have first used a technique called Hi-C to determine the 2D interactions between sections of DNA. Hi-C involves physically cross-linking interacting stretches of DNA and then fusing them together. That way, all the interacting stretches of DNA, even if they are far apart on the chromosome, will appear side by side in the sequencing data. These 2D data would then be used to build 3D models, but this required a lot of assumptions about what would happen based on the DNA sequence and proteins that hold loops of DNA together. The researchers behind the new study didn’t want to make that many assumptions. “All we wanted to do is simply make a structure that follows the rules from experiments,” says the study’s first author Anna Lappala, a polymer physicist at Massachusetts General Hospital.

So the researchers incorporated experimental 2D interaction data, simulated physical forces, and Newton’s equations of motion to predict the 3D structure of the X chromosome. They didn’t stop with 3D, however. They repeated the process at different time points during a process called X chromosome inactivation (XCI), thus adding the fourth dimension to their analysis. The high-resolution modeling, which required analysis of enormous datasets, was made possible by using supercomputers at Los Alamos National Laboratory.

When two X chromosomes are present in a cell, most of the genes on one X are deactivated through XCI to prevent developmental abnormalities. This silencing is initiated when one X expresses a noncoding RNA called Xist that coats the chromosome. 

The results of the structural modeling show that as the X chromosome undergoes XCI, it forms a dense core with a looser surface. Several genes known to escape XCI are located at the surface, which the researchers suggest might allow better access for the molecular machinery that expresses genes. “What’s really ground-breaking about this paper is that we figured out a way to visualize the 3D structure of the X chromosome based on a decidedly two-dimensional map that one gets from Hi-C datasets,” says Jeannie Lee, a molecular biologist at Harvard Medical School and coauthor on the paper. The authors were also able to track the spread of Xist RNA on the chromosome over time, which they show in a video


Sanbonmatsu, also an author on the paper, says that their computational approach to modeling 3D structure could also be applied to the rest of the chromosomes in the genome.

Computing the future of biology

“What I think is most significant about this kind of work [in the PNAS paper] is the attempt to explain something very complex, such as the structure of the genome and especially dynamic processes in the genome, based on the fundamental first principles of physics,” says biomedical engineer Vadim Backman of Northwestern University who wasn’t involved in any of the three studies. 

“Across all three papers, an important point to make is how critical computational analysis and modeling is for this area,” adds Murphy. In the three new papers, powerful computers allowed the scientists to take huge experimental datasets and learn something new about organelles, chromosomes, and other structures without years and years of manual work.

To speed up the analysis of data and progression of 3D cellular biology research even more, the authors of all the papers are committed to open access to data. “It’s much better to let the world see what we have invested in,” Xu says. Xu has patented the new microscope technology, but it is free for universities and nonprofits to use, and the atlas of 3D cell data is freely available to explore. Sanbonmatsu wants to eventually enable biologists to look at the 3D structure of whichever chromosome or gene they are interested in using a “point and click on a browser.” 

“We’re trying to democratize this whole process,” she says.