Skip to main content

Why it takes a supercomputer to map a mouse brain

The structure of a mouse mind is incredibly complex

Share this story

Neurons and synapses of a mouse brain.
Kasthuri et. al., Cell 2015

Inside a 25,000 square foot room within Argonne National Laboratory one of the most formidable supercomputers in the world — Theta — is applying its incredible computing power to the largest batch of data ever recorded or analyzed. It’s information that researchers hope might one day contribute to our understanding of intelligence itself.

And in this case, all that data fits inside the skull of a mouse.

Theta is currently mapping the structures of mouse brains, using a data set that’s being gathered piecemeal by Narayanan “Bobby” Kasthuri, neuroscience researcher at Argonne National Laboratory and assistant professor in neurobiology at the University of Chicago. When the entire set, soup to nuts, is procured, the end result is predicted to be a million terabytes, a monstrous, impossible to comprehend amount of raw information.

Kasthuri explains his aspirations for those vast amounts of data after he hustles down Drexel Avenue to escape the cold on the University of Chicago’s medical campus near his lab. For him, the future hopefully holds an unparalleled understanding of the brain.

blobs of color moving past each other with the cyclic psychedelia of a Broad City title

Kasthuri’s lab is using Theta to create colorful diagrams, blobs of color moving past each other with the cyclic psychedelia of a Broad City title. What they are diagraming is as trippy as it looks; small slices of mouse brain, pinned down and color coded to show the brain’s structure on a minuscule level. The color blobs are neurons and synapses, the pathways through which the mind moves, every one of them documented in a nanometer-scale map called a connectome.

Researchers think a complete connectome could help to unlock the mysteries of cognition and mental health disorders, and might deepen our understanding of the differences and similarities between the brains of various organisms. On the base level, it will provide a knowledge of the working units of the brain that is crucially missing from current research.

Currently, Kasthuri is interested in creating and comparing partial connectomes from different types of animal brains. Right now, the lab is working on sections associated with addiction in mice. A mouse-addicted-to-cocaine connectome compared to a non-addicted mouse connectome could identify what neurons are impacted by addiction.

“We’re already finding — it seems like there are structural changes in the addicted brain,” Kasthuri says.

Brains on a conveyor belt

Studying the structure of a mouse brain is a delicate process, with huge potential for error. The brains are plucked from the mouse as fast as possible, which has been preserved with aldehyde fixatives in a race against the damage of death. A section of the brain is dunked in heavy metal stains for the scanning electron microscope (SEM) at Argonne. After being dehydrated and plasticized, the sample is sliced with a deli cutter-like device by a diamond knife a couple atoms wide. With typical cutting systems, the slicing and moving of the samples can lead to imperfections, which could be magnified when the supercomputer gets involved.

the sample is sliced with a deli cutter-like device by a diamond knife

Those tiny flaws can mean huge problems when researchers start analyzing the data, explains Animashree Anandkumar, a Bren professor of computer and mathematical science at Caltech who is not involved with the project. “These distortions can lead to spurious correlations,” she says.

Kasthuri’s solution is simple — a proprietary conveyor system which carts off the brain slices quickly, minimizing human error. After their ride on the conveyor, the samples are scanned by the SEM, resulting in stacks of image data. The individual neurons, synapses and other structures are recognized by shape, traced, then colored like they would be in Photoshop, tedious yet crucial work done by students like Anastasia Sorokina and Katrina Norwood, graduate students in Kasthuri’s lab.

Kasthuri lab members Anastassia Sorkina, Katrina Norwood, and Rafael Vescovi
Kasthuri lab members Anastassia Sorkina, Katrina Norwood, and Rafael Vescovi
B David Zarley

Unfortunately, it takes a human forever to digitally color in each individual structure, and analyzing the results creates connectomics’ main bottleneck. Kasthuri’s dream—a complete mouse brain; the “mouseshot”— would be practically impossible with people alone. He’s done the math. Every person on Earth working perfectly, eight hours a day, six days a week would still take 500-1,000 years to finish the project. A hundred, if you could draft everyone who has ever lived.

A coloring book for computers

The answer is an algorithm. Specifically, flood-filling networks developed by Google AI and the Max Planck Institute of Neurobiology in Germany.

“When you’re dealing with billions and trillions and hundreds of trillions of pixels in these datasets, there’s just no way that human analysis is going to be feasible” says Viren Jain of Google AI.

there’s just no way that human analysis is going to be feasible

According to Jain, one of the architects of the algorithm being used, the flood-filling algorithm approaches the data in the way a person would a coloring book. It starts with a certain structure — a particular neuron, say — that it fills in before moving on to others. That’s where Theta comes in.

Haritha Siddabathuni Som, team lead at the Argonne National Laboratory’s Leadership Computing Facility, lists off the supercomputer’s impressive stats — it’s the facility’s most powerful supercomputer yet, occupying 24 server racks, and when it’s not mapping mouse brains, it’s working on other huge data sets, including some from CERN’s Large Hadron Collider, peering into the mysteries of particle physics.

Argonne National Laboratory’s Theta supercomputer
Argonne National Laboratory’s Theta supercomputer
B David Zarley

Theta’s Intel-Cray XC40 hardware is capable of 11.69 petaflops, which, for the layperson, is really freaking fast. Humans can analyze one cubic micron of mouse brain in roughly 2 minutes; there are one trillion cubic microns in a mouse brain. While it would take all of humanity centuries to complete the task, Kasthuri believes that Theta, using the algorithm, can get it done in just five years.  

The flood-filling algorithm harnesses Theta’s horsepower for its tracing, coloring, and compiling of the mouse brain data, sending it to Argonne’s analysis and visualization cluster, Cooley, which produces the acid trip connectomes. The data sets will be open source, available for any and all to look at and study.  

The algorithm’s approach is an order of magnitude better than previous options. But even with a state of the art algorithm and a powerful supercomputer, progress towards a whole mouse connectome is still slow.

A complete connectome is rare; there was one in the late 1980s, of the worm C. elegans, for example. The next complete brain will likely be the fruit fly, a science warhorse; the FlyEM project at Janelia Research Campus aims to have a full connectome of about 1/3 of the fruit fly brain published within a year, Stephen Plaza, project scientist, says by phone.

Kasthuri expects to have his mouse brain finished within five years of the fly’s

Roughly 30 years separate the full connectome of the C. elegans worm from the fruit fly brain; in comparison, Kasthuri expects to have his mouse brain finished within five years of the fly’s, funding pending, of course. The final goal is audacious, even compared to the mouseshot: a full connectome of the human brain. That effort will take more time, and an even larger, more powerful supercomputer, Aurora 21, which is currently being built at Argonne.

Kasthuri leans in when he talks about the possibilities a full human connectome may unveil. He imagines the capability to find, and fix, issues caused by mental health disorders and traumatic brain injuries; he imagines proving if cognition comes from building up connections or, as a sculptor does to marble, honing them away (he likes the latter). He, and the other researchers in the field, imagine answers, which of course will inspire more curiosity.

“We’re constantly finding that looking at the connectome basically seems to be a Pandora’s box,” says Plaza of the FlyEM team. “Not for answers, but for even more questions.”