A Trove of Mouse Data Points Toward Brain-Computer Interfaces


The machine for figuring out what’sgoing on inside a living brain—well, it doesn’t look much fun for the mouse under the business end. At the Allen Institute for Brain Science, in Seattle, researchers first peel away a piece of a mouse’s skull—a lot of mice, actually—and replace it with a little window with holes in it. Then they lock the mouse into a frame that bolts onto its head.

The actual brain-scanning part is called a Neuropixels probe, a 10-mm long needle made with the same technology as computer chips. Six of them sit on a cartridge that descends on a robot arm through the holes in the skull window and into the mouse’s brain. Each needle is studded with almost 1,000 sites that can record neural activity, the electrical spikes of neurons talking to each other. The Allen researchers plunge those into the deepest regions of the visual cortex.

Then they show things to that mouse and use the Neuropixels to see what the brain does in response. What kind of stuff? A grid. A black-and-white grating. Another grating, but this time moving slowly. The sort of visual stimuli that every kind of mammal’s brain responds to. Also: pictures of the natural world, animals and more. They also showed the mice two clips from the long, single-take opening scene of the Orson Welles movieTouch of Evil, over and over again. All the while, the Neuropixels collected data from hundreds, perhaps thousands, of neurons localized with a brain-mapping standard the institute came up with.

It’s not so easy to peer inside a mouse’s head.

Photograph: Allen Institute

On Monday, the Allen Institute gathered up all 70 terabytes of raw neuroelectrical data, sorted it into 850 gigabytes of more useful information, and gave it—plus maps and new visual imaging of the brain at work—away. That’s the point. To provide the people studying brains, computers trying to simulate brains, and brain-computer interfaces with some actual numbers to work with. “For the last 50 years, people collected data from just 30 neurons, 40 neurons,” says Christof Koch, chief scientist at the Allen Institute. “It’s very different if you have 2,000, or 50,000, or 100,000 like we have.”

Koch says this dataset shows action from connected regions across the brain—maybe 1,000 neurons simultaneously from 10 different areas along the length of the probe, from one visual area to another, or from visual areas to regions where higher-order processing happens. That’s orders of magnitude more neurons than what most researchers can observe. “It’s really going to dramatically change how neurophysiology is done.”

Datasets like the Allen one aim at basic, fundamental knowledge about the brain. It’s supposed to give other researchers something they can hypothesize about, do math on, even train machine learning algorithms with. “It’s the whole gamut of questions you have in neuroscience,” Koch says. “What are the signatures of learning? If an animal learns a task, how does it show up?”

Allen Institute researchers looked for links between what mice saw and the neural activity in their brains.

Photograph: Allen Institute

Eventually, neurotechnologists would like to understand brains well enough to read their output and use it to treat disease, or even control computers and other tech. Right now, nobody knows how to do that. It’s true that labs and companies around the world are working on devices that can read brain activity with more accuracy and specificity, but they’re all still just rough drafts. Elon Musk’s companyNeuralink, for example, seems to be attempting to use a technology called neural lace, in which a robotstitchesfine electrodes into the brain itself. They’re still in rodents, and say they’re also working with non-human primates. One group of researchers developed aneural meshthat stayed implanted in mice for up to eight months; a company calledCTRL-labs(which Facebook bought) is trying to jump into humans by interfacing computers with output from motor neurons via signals in the wrist. Facebook also has a separate efforttargeting speech centersto translate words into something a computer can understand.

One major obstacle is measuring neurons’ electrical activity while also accurately mapping their location inside the brain. That’s how you figure out how a brain actually works, but it’s not easy. For one thing, you need monitoring that lasts for weeks, months, maybe even years. But electrodes degrade in the salty cerebral soup, or get gunked up by the subject’s immune response. Neuropixel probes, for example, can get in the way of the microscopy that’d be able to actually see the neurons at the same time. Softer,flexible electrodesmight help with that. (Allen researchers index all their data—electrophysiology, 2-photon imaging of neural activity, and so on—to the same overall 3D brain map.)

Another potential problem: mouse data is still mouse data, meaning it’s not a perfect model for people. The mouse visual cortex is kind of a mess, organizationally speaking—mice are tactile and olfactory critters in ways primates are not, according to some brain scientists. “I honestly feel personally a little bit similarly,” says Loren Frank, a neuroscientist at the Howard Hughes Medical Institute who works with bendy,polymer electrodes. “But in defense of the mouse people, it is clear that the mouse cortex does have many of the same properties.”

That’d be great, because mice are easier to work with than any primate—mouse research is cheaper, faster, better understood, and has few ethical constraints. (Sorry, mouseys.)

More importantly, perhaps, the tools for figuring it out are further along, which means places like the Allen can advance the tech and the data in mice, looking at more neurons at once, learning to read the data better. That’s all with an eye toward getting to humans sometime down the road. Koch says they’re working on it: “We’re working with a number of surgeons to get protocols approved to put this into a piece of the human cortex where the surgeons know ahead of time they have to remove it,” Koch says. That’d be for surgery to deal with, say, focal seizures. “That’s not too far away.”

And meanwhile, other researchers are working on developing a Neuropixel probe almost five times as long as the one at the business end of the mouse machine—48 mm, long enough to penetrate to the deepest layers of a primate brain. Food, as they say, for thought.


More Great WIRED Stories

Read More

LEAVE A REPLY

Please enter your comment!
Please enter your name here