Okay, the “sewing machine” is pretty cool. But if the device that Elon Musk’s neurotechnology startup Neuralink developed to implant thousands of electrodes into brains (of rats and monkeys so far, and humans eventually) were its only accomplishment, Tuesday night’s big reveal would have been a big meh. Instead, six independent experts in the kind of brain-computer interfaces that Neuralink is developing told STAT they are impressed for the most part, though caveats abound.
“Overall, the concept is impressive and so is the progress they’ve made,” said neurobiologist Andrew Schwartz of the University of Pittsburgh Medical Center, a pioneer in the technology. “But a lot of this still seems to be conceptual. It’s hard to tell what’s aspirational and what they’ve actually done.”
The immediate aim of the San Francisco startup is a system enabling people who are paralyzed to use their thoughts to operate computers and smartphones. That has been done before, including by Schwartz’s group and one at Brown University, where in 2011 two tetraplegic patients who had been implanted with the “BrainGate” neural interface system were able to control robotic arms with their thoughts, including lifting a bottle of coffee and drinking it.
But despite decades of research, the systems have so many drawbacks they’re still not in widespread use. That’s where Neuralink’s tech acumen might make a difference.
Dr. Leigh Hochberg of Brown University, who helped develop the brain-computer interface given to the patients with tetraplegia, called Neuralink’s “a novel and exciting neurotechnology…. I’m excited to see how they’ll be translating their system toward initial clinical studies.”
A key feature of Neuralink’s system is the sheer number of electrodes it plans to implant via its “sewing machine,” in which a stiff needle rapidly shoots thin-film polymer probes containing arrays of electrodes into the brain. In a white paper released on Tuesday night, by “Elon Musk & Neuralink,” the company said it had implanted 3,072 electrodes in rat brains.
There appeared to be discrepancies, however, between what Musk and his team presented in their splashy unveiling event in San Francisco on Tuesday night and what they said in that research paper, which has not yet been published or peer-reviewed and was posted on Wednesday morning to a preprint server.
“The paper, I would say, is much less ambitious than the overall presentation,” said Jacob Robinson, a neuroengineer at Rice University.
During the presentation, Musk and his team spoke with excitement about a brain-computer link that could go in both directions — recording neural activity and also stimulating it. In the model, the interface could stimulate the somatosensory cortex in a way that would make paralyzed patients feel as though they had touched something.
Pittsburgh researchers achieved that in 2016. But Neuralink’s paper doesn’t demonstrate a similar ability. That’s an important omission, Robinson said, because “given the electrodes they are working with, stimulation is going to be a lot more challenging than recording.”
Musk, in the presentation, also spoke dreamily of using the technology to merge human brains with artificial intelligence. In the white paper, there is no data to support such a vision.
An unsolved longevity problem
In his own research, Robinson is working on developing an interface to stimulate and record brain activity. He’s particularly interested in “the actual electrodes that have to live inside the tissue — and listen to brain activity and exist for a long period of time and hopefully not screw things up,” he said.
Which is why Robinson was pleased to hear Musk and his team allude to the field’s still-unsolved longevity problem — the stubborn reality that at some point, whether slowly or all at once, the electrodes implanted inside the brain will stop working.
Such a breakdown would likely require patients to get a new implant — and to have the old one taken out. Removal could be particularly challenging for Neuralink, because it’s using remarkably thin electrode-holding “threads,” Robinson said.
There’s also a safety risk associated with the failure of electrodes, if that meltdown is caused by an infection, likely in the brain tissue or in the wire implanted in the brain. Robinson praised Neuralink’s scientists for being thoughtful about trying to dramatically reduce that risk, by setting a goal to develop a device wirelessly powered through the skin so as to close up the implantation site.
But Neuralink doesn’t have that technology yet. In the system described in the paper, “there actually is a path for infection,” Robinson said.
Aside from infection, safety risks from a brain-machine interface could include a stroke, an aneurism, or an immune reaction.
“The big risk, of course, is if something goes wrong… one more major adverse event could cause a stricter regulatory crackdown,” said Andrew Hires, a neurobiologist at the University of Southern California, whose research focuses on how the brain processes the sense of touch and emotion in the cortex.
By way of a cautionary tale, Hires pointed to gene therapy, where an 18-year-old’s death in a clinical trial in 1999 chilled investment and research for a decade. Still, Hires is optimistic about Neuralink’s prospects, calling the technology it unveiled “serious and credible.”
A matter of design
Neuralink says it hasn’t figured out yet how many electrodes it would need for a brain-computer interface in patients, but it seems to have the ability to soar past existing systems. Prototypes tested in patients use hundreds of electrodes, while some used for research in monkeys approach 2,000. “Neuropixels” developed by physicist Tim Harris of the Howard Hughes Medical Institute’s Janelia Research Campus and colleagues have 960 electrodes per “shank,” a device that resembles the teeth of a comb, with electrodes on each tooth. Researchers routinely put four or even eight into their lab rats, but Neuropixels aren’t used in human brain-computer interfaces yet.
The more electrodes, the more neurons’ activity a system can monitor, including the “spikes” that indicate a neuron has fired. The more neurons that are monitored, the more information that gets fed into the software that analyzes their meaning and translates it into an output such as “move the robotic arm 3 inches up and left and grab what’s there.”
“With single-spike resolution and recording from more neurons, you can control more things more precisely,” Schwartz said. With existing EEG systems, patients spend months learning that they have to think “now I want to move my right toe” and then “I want to wiggle my nose” in order to move a computer cursor up to the top right corner of a webpage, say. “If you could record individual spikes [the brain-computer interface] would work as soon as you turned it on,” he said.
Neuralink’s electrode-holding threads are about 1.6 millimeters long, according to the company’s paper. “That would cover most of the cortex in a human,” said Harris, and therefore reach neurons below the surface that most such systems record from. “That’s the appropriate geometry you need.”
Perhaps not surprisingly for a Musk company, the system’s back-end electronics, which receive, amplify, and decode the brain activity detected by the implanted electrodes, impressed experts more than the in-brain part. The miniaturization, Harris said, is “significantly advanced over prior art.”
The quality of the brain end of the system isn’t as clear. According to data in the paper, no electrode detected neuronal activity from rat brains as strong as 100 microvolts, which Harris calls the minimum signal “that you can trust.”
If he saw such weak signals from Neuropixels, he said, “I would think something is wrong. These are very low amplitude signals. What happened? If we had these data I would think something had gone south, like our probe missed its target or we did damage inserting it.” Damage to brain tissue can heal, however, so if that’s the cause of the weak signals it might be temporary. Still, he said, “I look at the Neuralink paper and think, that’s not what I want my data to be.”
On the plus side: Neuralink’s “sewing machine” can implant the threads holding thousands of electrodes into the rat brain in an hour or less. It might take longer in people, but the best existing system, tested in rat brains. requires 48 hours of surgery. “That’s a big difference, driven by the robot that Neuralink built,” Harris said.
On the minus side: brains may not take kindly to having threads pounded into them. “My guess is that the rapid insertions do substantial long term damage, such that they may not get high quality recordings for very long, but that’s just a guess at this point,” said neuroscientist Loren Frank of the University of California, San Francisco. “We go into the brain much more slowly, and that’s likely an important part of why we were able to get such long lasting recordings” in rat experiments.
Challenges gobs of money can’t solve
Neuralink’s vow to start testing its system in patients by next year seems ambitious, if not delusional, to every expert STAT consulted. A more realistic timeline, Hires said, might project that Neuralink could start safety testing in humans within a few years — after it has produced data showing it can safely implant its technology in non-human primates and that there are no long-term consequences to doing so.
But regardless of the time frame, said Schwartz, “to me, they have the right concept, and are doing what needs to be done.”
With $158 million in funding — $100 million from Musk — Neuralink will have luxuries that academics and less richly funded companies can only dream of.
For example, when Neuralink’s scientists design their chips, they might be able to try twice as many designs as a typical academic lab can afford, helping them move faster, Robinson said.
But throwing money around may not overcome other obstacles.
Consider testing how long an electrode can live functionally inside the brain. “You can’t accelerate that process. You just have to wait — and see how long the electrodes last. And if the goal is for these to last decades, it’s hard to imagine how you’re going to be able to test this without waiting long periods of time to see how well the devices perform,” Robinson said.
The same may go for navigating the Food and Drug Administration.
“As much as I believe that with the right kind of resources, they can really push engineering at an incredibly accelerated rate,” Robinson said, “the biological processes and the regulatory approval… are going to be harder for them.”