Microelectrode arrays

Once you can fabricate electrodes with photolithography, there is no particular reason to make only one. That observation is most of what a microelectrode array is, and its consequences took thirty years to work through.

What it is

A substrate — glass, silicon, later flexible polyimide — carrying tens to thousands of metal sites, each with its own trace out to an amplifier. Cultured cells grow on top, or the whole thing is inserted into tissue. Every site records the extracellular field, so you get spikes from whatever neurons are close enough, without impaling anything.

Why it matters

The trade against the patch clamp is stark, and it is a trade, not an improvement. Extracellular recording cannot see subthreshold voltage, cannot isolate a single channel, and gives you spikes filtered through the geometry of the tissue between the cell and the site. What it gives back is simultaneity.

Questions about how a population coordinates — synchrony, sequences, oscillations, the relationship between one neuron’s firing and its neighbours’ — are not answerable one cell at a time, no matter how well you measure that cell. Changing the channel count changed which questions existed.

The constraint is impedance

The engineering constraint that governs everything here is site impedance. A smaller site sees fewer cells and localises better, but its impedance rises, and with it the thermal noise referred to the amplifier input. Low-impedance coatings — platinum black, iridium oxide, later conducting polymers — exist to break that coupling: more effective area for the same footprint, so selectivity and noise stop trading directly against each other.

The noise is Johnson noise from the real part of the site’s impedance $Z$. Its RMS voltage is $v_n = \sqrt{4kT\,\mathrm{Re}(Z)\,\Delta f}$, where $k$ is Boltzmann’s constant, $T$ the absolute temperature and $\Delta f$ the bandwidth. Working from measured impedances, Viswam and colleagues found in 2019 that shrinking a bare platinum site from 100 × 100 µm to 1 × 1 µm raised that noise in the spike band, 300 Hz to 5 kHz, from about 3 to 13 µV RMS — against extracellular spikes of tens to hundreds of microvolts. Coated with platinum black, every size stayed below 8 µV.

Impedance costs signal as well. The site forms a divider with the amplifier’s input, and the capacitance of the wiring between them shunts it. On a CMOS array whose amplifiers present about 3.8 pF, some 40 MΩ at 1 kHz, the smallest bare site they measured, 11 µm², lost about 70% of the signal; coated, none of the four sizes lost more than 2%. The other cure is to put the first amplifier on the array, before the wiring can load the site.

One electrode or a dozen

In 1970 Kensall Wise and James Angell, electrical engineers at Stanford, and the Stanford neurologist Arnold Starr described a probe made the way integrated circuits are made: gold electrodes on a silicon carrier, insulated by a film of silicon dioxide that photoengraving opened only at the tips. With probes of this kind they recorded single neurons in cat cortex. They put the case for arrays in one line — once the photomasks exist, the processing is “identical whether one electrode or a dozen are needed per probe” — and proposed source-follower amplifiers on the carrier, so that the capacitance of the leads would neither attenuate the signals nor couple the electrodes together.

Two years later Thomas, Springer, Loeb, Berwald-Netter and Okun put the electrodes under the cells instead, as the floor of a culture chamber, and recorded the electrical activity of contracting heart cells from chick embryos. Single neurons in culture took longer. Guenter Gross and colleagues described a fixed array designed for long-term single-unit recording in vitro in 1977, and by 1979 Gross was recording single units from snail ganglia on 36 photoetched gold electrodes. Jerome Pine’s 1980 dishes had 32 electrodes, each about 8 by 10 µm, under neurons dissociated from the superior cervical ganglia of newborn rats. They recorded individual cells within about 40 µm of an electrode’s centre, and could stimulate through the same sites.

Wise took the silicon line on at the University of Michigan. By 1985 Najafi, Wise and Mochizuki were making shanks 3 mm long, 50 µm wide and 15 µm thick, defined by a deep boron diffusion that stops the etch. In 1986 Najafi and Wise built what the 1970 paper had proposed, and more: amplifiers and a multiplexer on the probe itself, ten sites read out over three leads, for 5 mW. The Utah array took the other geometry. In 1991 Campbell, Jones, Huber, Horch and Normann described a 4.2 mm square of silicon carrying 100 needles 1.5 mm long, each tipped with platinum and isolated from its neighbours by p–n junctions — sites spread across a patch of cortex at one depth, where Michigan’s run down a shank. It was built to stimulate, for a visual prosthesis; a commercial descendant became the recording sensor of the BrainGate trials.

Population dynamics, not just single units

What an array shows that an electrode cannot is coordination. On arrays in the floors of culture chambers, Droge, Gross and colleagues found in 1986 that networks of mouse spinal neurons, dissociated and seeded at random, often burst in rhythms coupled across electrodes — mostly in phase, occasionally in antiphase — with each electrode hearing different cells, and some patterns lasting more than twelve hours.

The retina made the point sharper. Maffei and Galli-Resta had found in 1990 that neighbouring ganglion cells in prenatal rats usually fire in strong correlation, the kind of correlation a Hebbian rule could use to sort the inputs from the two eyes. With isolated retinas from newborn ferrets and fetal cats on a multielectrode array, Meister, Wong, Baylor and Shatz recorded up to a hundred ganglion cells at once in 1991 and saw what the correlation was: bursts a few seconds long, separated by a minute or two of silence, each a wave of excitation several hundred micrometres wide sweeping across the retina at about 100 µm/s. One electrode would have shown a cell that bursts every minute or two. The wave exists only across the array.

Worth being clear that not every population result needs simultaneity. Churchland and colleagues’ 2012 study of motor cortex, titled Neural population dynamics during reaching, worked from trial-averaged firing rates and found the population state rotating in all nine of its datasets: seven assembled from single-electrode recordings made one after another, two from pairs of implanted Utah arrays. Averaging builds a population after the fact, but only from what repeats on every trial. Anything that differs from one trial to the next — which cells join a burst, the order they fire in, how two neighbours’ fluctuations covary — averages away, and keeping it is what the channel count is for.

What it costs

An array hears what fires near its sites: within about 40 µm on Pine’s dishes, and in tissue out to roughly 100 µm from where a spike starts. It also hears only what fires. Henze and colleagues estimated in 2000 that a tetrode — four fine wires — in rat hippocampus is within range of about a thousand neurons, 60 to 100 of them separable in principle. About six are usually found; a large fraction of the pyramidal cells, they concluded, are silent in any given behaviour. A recorded population is the near, active part of the real one, and assigning its spikes to cells — spike sorting — is an inference of its own.

Implanted arrays fail, and in one long series the brain was not the main reason. Of 78 Utah arrays implanted in monkeys from 1996, Barrese and colleagues found that recordings lasted a median of 182 days and most failures came within a year; abrupt mechanical failures, mostly of connectors, outnumbered those with a visible biological cause two to one. Without such interruptions, recordings declined slowly, losing spike amplitude and working channels at a rate that predicted complete loss at about eight years, with falling impedance pointing to failing insulation.

And a passive array needs a conductor per site. In a dish the traces can leave on all four sides; on a shank they all leave along one narrow end. Arrays of thousands of sites therefore put switches and amplifiers in the substrate and multiplex, which is the step Neuropixels took on a probe.