Note: Part 1 of this two-part series is available here.
Neuropixels probes have changed the face of electrophysiology. But they do have drawbacks. Inserting a stiff silicon probe into the brain can damage the surrounding tissue, as can movement of brain tissue around the probe. When the brain detects a foreign object, it can trigger the formation of scar tissue, potentially worsening recording capacity over time. An alternative approach is to build probes that are thin and flexible enough to integrate seamlessly into the brain. Both technologies offer unprecedented access to the brain, as well as new challenges, namely how to accurately process the reams of data they produce.
Pliable probes, built on a polymer substrate, are more difficult to insert than their silicon counterparts — it’s like piercing cloth with thread rather than a needle — but they cause little damage to surrounding tissue. And because they move with the brain, they stay anchored to the same cells day after day. “I’m still in the long run a believer in flexible devices as best for long-term recordings,” says Loren Frank, a neuroscientist at the University of California, San Francisco and an investigator with the Simons Collaboration on the Global Brain, who is collaborating on two different flexible-electrode projects, one with Lawrence Livermore and Lawrence Berkeley national labs and the other with Rice University.
Flexible electrodes are manufactured using photo and electron beam lithography, microfabrication methods that can produce tiny patterns at about the same density as silicon. Frank says these methods are cheaper and more widespread than the high-end fabrication methods used to make Neuropixels probes, making them more broadly accessible. It’s also easier to modify the design to suit a particular brain region or type of experiment, he says. The current iteration of the Livermore device, in development for about a decade, has 128 electrodes per probe, all of which can be used simultaneously. “If properly placed, we can record 300 neurons for months,” Frank says. Multiple probes can be placed at different spots in the brain, enabling researchers to study long-term interactions across structures.
In research published in Neuron in 2019, Frank and collaborators used the devices to search for signs of coordinated activity across the brain at the single-cell level. They implanted 16 of an earlier iteration of the probe into a rat’s brain and found changes in activity in multiple brain regions that coincided with sharp-wave ripples in the hippocampus, synchronous waves of activity tied to memory and planning.
A number of labs are now testing the probes, and researchers aim to move them to commercial production over the next few years. Frank’s collaborators at Livermore and Berkeley labs are developing versions for other systems, including mice, primates, gerbils and songbirds. SCGB investigator Beth Buffalo at the University of Washington is testing the devices in primates. She hopes the probes will help identify novel population codes in the hippocampus similar to those recently uncovered in mice. Her group is in the final stages of validating implantation techniques, which they will share with the broader community.
Frank is also collaborating with Chong Xie and Lan Luan at Rice University to develop devices that are even thinner and more flexible than the Livermore probes, which at 14 microns are about the thickness of plastic wrap. The new probes are an order of magnitude thinner — 0.7 to 1 micron thick — which translates to a thousand-fold increase in flexibility, Xie says.
At that exceptional thinness, “the force they exert on the brain is similar to the force a cell membrane would exert,” Frank says. Two-photon imaging of the probes suggests they integrate into tissue without scarring. “That’s probably not true for any other probe,” Frank says. “That’s the potential appeal — you may be able to get seamlessly into tissue.”
About 10 labs are now testing the current version of the device, which has 100 sensors. Frank, Xie and collaborators are one year into a four-year grant from the BRAIN Initiative to disseminate the technology. Together with Mattias Karlsson, co-founder of the company SpikeGadgets, Xie and Luan have started a new company, Neuralthread, to commercialize the technology. Karlsson says 32-channel probes will be available at the end of this year, with 64- and 128-channel versions available next year. They don’t yet know how much the probes will cost but estimate it’ll be less than $1,000.
A major challenge for flexible electrodes is how to insert them — unlike a stiff silicon device, flexible threads can’t be stuck directly into the brain. To overcome this problem, the threads are attached to a retractable needle with a water-soluble stiffener, which dissolves once the probes are inserted into tissue. For the newest versions of the Rice probes, a specially manufactured needle is threaded through a small hole in the bottom of the probe. The procedure is more complex than that of Neuropixels, which some researchers speculate could detract from their use. “Even if the probes were widely available, a lot of people would probably choose Neuropixels from ease-of-use and data-quality perspectives,” Steinmetz says.
Both the Livermore and the Rice teams are working on higher-density versions of their devices with more than 500 channels. With that many wires, the devices are too bulky to connect to a conventional head stage. Flexible electrodes require a different solution than that employed by the Neuropixels probes, which integrate both electrodes and processing hardware on the same piece of silicon. Instead, Frank’s collaborator Peter Denes and his team at Berkeley Lab have designed a special 512-channel chip, called the EChip, that can be soldered to the probes. All the channels on the flexible electrodes can be used simultaneously, versus only a subset of the channels on Neuropixels probes. The researchers aim to stack eight of the 512-channel devices for a total of 4,096 electrodes, making it possible to look at coordination of neural activity across structures, Frank says.
Unfortunately, the pandemic dramatically slowed efforts to test these new technologies. With many labs closed or operating at reduced capacity for the last year, “testing new devices was the lowest priority,” Frank says. But they hope to restart testing later this year as labs reopen.
Flexible electrodes aren’t simply the domain of academic labs — the potential for minimal damage and long-term stability make them promising for brain-computer interfaces, Silicon Valley’s latest tech craze. Indeed, Frank’s Livermore project lost steam a few years back when Elon Musk went on a hiring spree for his startup Neuralink and recruited some of Frank’s key collaborators. “Everything we had designed together went into their early-generation devices,” Frank says. Musk and collaborators described details of the technology in a 2019 bioRxiv preprint.
Neuralink has been much maligned by the neuroscience community for its splashy demos of tasks that academic neuroscientists have been doing for decades. In a video released in April, for example, a monkey plays the video game Pong using the Neuralink implant. Neuroscientists routinely train monkeys to perform much more sophisticated tasks with brain-computer interfaces and are testing implants in paralyzed people. In May, SCGB principal investigator Krishna Shenoy and collaborators developed an algorithm that broke the record for typing using a brain-computer interface. (Shenoy consults with Neuralink.)
But neuroscientists who helped develop Neuralink’s technology point out that the company has made significant progress transforming academic tools into a potential clinical device. “They are a company that excels at engineering,” says Tim Hanson, one of Neuralink’s founding team members now at the Howard Hughes Medical Institute’s Janelia Research Campus, who with Philip Sabes and Michel Maharbiz developed electrodes and a surgical robot to insert them. “A lot of effort at Neuralink focused on taking electrodes and the surgical process from a slow, mostly manual effort that worked in a few rats to something that works reliably in primates and ultimately humans,” adds Sabes, who now works at another neurotech startup. “They were taking ideas that had been demonstrated in the lab and putting it into a package you can imagine a patient getting.”
Previously, at UCSF, Sabes and Hanson developed a sewing machine robot designed to insert flexible electrodes while avoiding blood vessels, causing as little damage as possible. Neuralink streamlined the sewing machines and made them more reliable, with “significant advances in insertion, packaging and assembly,” Hanson says. Avoiding tissue damage and enabling stable, long-term recordings is especially essential for a clinical device intended for chronic use. Hanson hopes that Neuralink will publish information on long-term stability. “It would be gratifying to see them getting good stable recordings over years in monkeys,” he says.
Harris, Frank and others are impressed with Neuralink’s advances in electronics, which include packaging the components that might sit on an animal’s head into a fully implantable Bluetooth device. Brain interfaces currently in clinical testing, including the BrainGate2 device used in Shenoy’s study, have wires that ferry electrical signals from the brain to a processor. (The Neuralink device has not yet been tested in humans.) While they find the technology impressive, they note that it’s in the realm of modern high-end engineering. “A lab would have a hard time doing it, but a company with lots of money could,” Frank says. Given those advances, academic researchers are disappointed that the company has so far announced no plans to share their technology. Neuralink did not respond to queries about sharing plans.
Since Sabes is no longer at Neuralink, he says he can’t speak to the company’s plans for sharing. But he thinks the technology could have a huge impact on both clinical and basic research. “If these devices make their way into thousands of individuals, that will become an immensely important tool for research, similar to how deep-brain stimulation is an immensely important tool,” he says. “Clinical brain interfacing for medical use opens the possibility for human neurophysiology at scale that is a win for patients, industry and basic science.”
As the use of high-density recording technology spreads, researchers are also grappling with new challenges — namely, how to digest the reams of data they produce. Researchers have developed various software tools for this, targeting different steps in data analysis. The most difficult part of the process is automated spike sorting, which uses the shape of a spike to determine which neuron the spike came from.
According to a 2021 survey of Neuropixels users, the vast majority — 96 percent — use software called Kilosort to process their data. Kilosort and similar open-source software tools have greatly improved automatic processing. “These algorithms spit out high-quality spike-sorted data, which was a tremendous advance,” Steinmetz says. But each method has strengths and flaws, and there is low concordance among different sorters, making it difficult to know which to use. When recording from a small number of neurons, researchers could manually review the data to make sure the algorithm was working properly. But that’s impossible for thousands of neurons. “Now that we cannot review every neuron, we need a way to figure out which are high-quality neurons,” Steinmetz says, meaning you know exactly when each and every spike of the neuron occurred.
The SCGB sponsored a spike-sorting workshop in September to tackle this problem. A major challenge for assessing spike-sorting software is the lack of a gold-standard ground-truth dataset to test it on. One outcome of the workshop is a new project, headed by Daniel English at Virginia Tech, to develop a novel ground-truth dataset by recording from glass and silicon probes at the same time.
Researchers are also working with companies to make probes more broadly accessible and easier to use. SpikeGadgets, the company founded by Karlsson, is developing hardware and software systems for both Neuropixels and the different flavors of flexible electrodes. “A lot of technologies developed in the lab stop there,” he says. “We like the idea of working with labs and then commercializing tools in a way that’s sustainable.” A big part of Neuropixels’ success stems from the partnership with Imec, a commercial chipmaker, he notes.
SpikeGadgets specializes in high-channel data logging, developing souped-up head stages with a microprocessor that controls the digitizing microchip and packages information into a format readable by a computer. The company has developed systems that work with multiple Neuropixels probes, multiple EChips and other systems. “The other thing this approach allows is completely untethered systems,” Karlsson says. “We can log the data directly to an SD card on the head stage itself.”
Karlsson hopes efforts to develop and commercialize affordable high-density neural recording systems will grant much broader access to the technology, perhaps transforming how the field uses electrophysiology. “When a 128-channel system cost $150,000, labs could only afford to have one, and we were really limited to studying phenomena that were super stereotyped,” Karlsson says. “But so many things happening in the brain are not stereotyped from animal to animal — we can’t look at them in a statistically significant way until we can study enough animals.” Karlsson hopes that automating much of the recording process will change that calculus. “You can in principle have hundreds of animals doing these experiments on their own, via automation,” he says. That could “open the door for drug discovery companies to develop new therapies with electrophysiology as a biomarker, especially for diseases like Alzheimer’s and Parkinson’s that take a while to progress.”