Neuropixels: A Brief Case Study

(I’m ripping off Sasha Chapin and writing a “post” every “day” for 30 days. Kindly adjust quality expectations.)

Maybe as a legacy of the Manhattan Project, the field of physics is unusually good at large-scale collaboration to build research tools for themselves. CERN, LIGO, NIF, Arecibo — the list goes on. Other fields of science struggle to do the same.

A recent exception is the Neuropixels project started by Prof. Tim Harris while he was at Janelia.

Tim noticed that dozens of labs were recording electrical activity in the brains of live mice for their research projects, but they were doing so with subpar electrode probes that each lab (tried to) manufacture for itself. He decided that was bogus; led an effort to make a reliable, reusable, 10x-better-than-state-of-the-art probe; and made it available to researchers at-cost (~$1,000).

As of 2021, over 5000 Neuropixels have been shipped to labs around the world. Tim’s too modest to say Neuropixels has revolutionized microelectrode recording in animals, but just about everyone else you ask will say it has.

So, how did he do it?

Tim showed up to Janelia in 2008, having never read a single primary neuroscience paper. His colleagues’ probes at the time had about 64 sensors apiece, and everyone wanted more. He made some obvious improvements (his words) to his colleagues’ probes, but he knew he could do 10x better with the right technology.

Luckily the right technology wasn’t new: it was the same CMOS fabrication technology that goes into making modern computer chips. It’s just that neuroscience labs aren’t set up to fabricate electronics at the level of TSMC. But neuroscience didn’t need the fancy stuff. As Tim put it, he just needed to bring probes “from 40-year-old technology to 20-year-old technology.”

In 2013 he started hunting for a fabrication facility to build that technology. His opening offer to every one was “I don’t need a paper, I need a probe, and if you won’t manufacture them for everyone who wants one when we’re done, I’m going elsewhere.” Eventually he found IMEC in Belgium, who after some back-and-forth about the technical specifications said, “Sure, we’ll do it. That’ll be €5.5M.”

This was considerably more money than Tim had at his discretion, so he asked for some from his boss at Janelia, the legendary Prof. Gerry Rubin. Gerry did have that much money, but he told Tim, “If this project is as important as you think it is, you should be able to find a partner.” This wasn’t Gerry being withholding. It was that when you’re spending millions on a tool, you can only make one tool, so you need a strong, international consensus on what to build. And there’s no better way to get buy-in on a decision than…well…literally having people buy in.

Enter the lawyers

Two days after talking to Gerry, Tim heard that Christof Koch was leaving Caltech to go to the Allen Institute and that he wanted a better probe. Tim called Christof, and a few days later Allen signed on as a partner on the project. After some more phonecalls, Tim got Wellcome and researchers at UCL’s Gatsby interested.

Altogether that meant making a 5-way international multiparty contract between Janelia, Allen, UCL, Wellcome, and IMEC. Everyone understood IMEC would keep all the IP on the engineering and fabrication side, since they were doing all the work there. But Tim needed to make sure researchers would be free to publish papers about whatever they used Neuropixels to discover. Without that guarantee, the project was pointless. And he was insistent that “nobody gets to make money, and nobody gets to go first [once the probes were fabricated].”

At this point in the story, Tim gives all the credit to Janelia’s lawyers. We’ll have to use our imaginations about what went down, but the upshot is that thanks to “real champions” at the various institutions to curtail the “unrealistic views of academic lawyers and their opinions about the IP their faculty [would] generate,” all five parties eventually signed an agreement. The four academic partners agreed to pay IMEC to design probes that weren’t fully open-source, but were documented well enough for any researcher to use, and once completed they would be distributed by lottery to researchers, at-cost, with no IP claims on downstream discoveries.

The first 30 probes were shipped in 2018, five years after the project began. Each carries 960 senors, of which a researcher can select any 384 to simultaneously record from. A 2.0 version of Neuropixels was released in April 2021, and the team is now working on a human-usable version.

Product design for scientists

Tim gives the architecture of Janelia a lot of credit for making Neuropixels happen. It’s designed for serendipitous meetings: all glass walls, a single good coffee pot, and a single cafeteria that only serves lunch for 90 minutes. He was submerged in people who needed the tool he was building, and every day he could ask people, “Well, what if we did this?” And, critically, he could tell just by walking around who was forward-thinking and who had their heads down.

The initial Neuropixels proposal was made with the backing of only two such forward-thinking Janelia researchers. They were two of the top researchers in the field, so their backing meant something. But still, Tim is clear that he had to overcome a lot of conservatism. If you don’t offer neuroscientists a big step forward, they’re not going to change what they’re doing. Their projects already contain enough scientific risk; there’s no desire for additional risk of equipment failure. Moreover, scientists have an unreasonable affection for suffering (or for their underlings’ suffering), so you have to prove to them they’ll be able to do things they couldn’t do before, rather than just making their (underlings’) lives easier.

Once Tim got a critical mass of support, he claims his main contribution was consensus building, i.e. keeping the neuroscientists from screwing up the engineering by constantly changing specifications. The neuroscientists would change their mind about what they wanted every week, and he condensed that into small alterations given to IMEC every 6 months.

His main tip for consensus building is that meetings (1) should never last more than 30 minutes, (2) should have an agenda of very specific questions, and (3) should have cookies provided by you.


Have feedback? Find a mistake? Please let me know!

Special thanks to Prof. Tim Harris for wasting an hour of his day talking to me, and to Prof. Nick Steinmetz for his email responsiveness.