Neuropixels · Grey Matter

Neuropixels is a thin silicon probe that is pushed into the brain to record the spikes of hundreds of individual neurons at once along its length, and it is the standard tool for recording large populations of single cells across several brain structures in animals.

Neuropixels. Neuropixels is a thin silicon probe that is pushed into the brain to record the spikes of hundreds of individual neurons at once along its length, and it is the standard tool for recording large populations of single cells across several brain structures in animals.

The first version, published in 2017, is a single shank 10 mm long with a cross-section of 70 by 20 micrometres, carrying 960 recording sites. Amplifiers, filters and digitisers are built into the base of the probe itself, so 384 channels, chosen from the 960 sites, leave the probe already as digital data. Because the sites sit a few tens of micrometres apart, each spike is seen on several neighbouring sites, and spike-sorting software uses that pattern to assign spikes to individual neurons.

It reads what EEG cannot. Each site sees the action potentials of the few cells within roughly a hundred micrometres, so the probe records which neuron fired and when, with no inverse problem to solve.

It sees a thin line. A probe samples the neurons along one track, a few hundred out of the many millions in a region; covering the brain this way would take an impractical number of penetrations.

In humans it has been used in short recordings during brain surgery, with a probe version adapted for the operating room; chronic human use is still a research goal.

Later versions (2.0, with four shanks and a smaller base) increase site count and stability for long recordings in animals.

Neuropixels trades coverage for resolution.

It answers the question EEG cannot (which cell fired) for a few hundred cells, and leaves the rest of the brain unread.

Questions: How does a Neuropixels probe tell which neuron fired? Inside the tissue, a recording site a few tens of micrometres from a neuron picks up that neuron's action potential directly, as a sharp dip of a fraction of a millisecond. The probe's 960 sites are packed so closely along its shank that each spike appears on several neighbouring sites with a characteristic pattern of sizes and shapes. Spike-sorting software groups spikes by that footprint, and each group is taken to be one neuron, which turns 384 channels of voltage into the firing times of hundreds of individual cells. What would it take to record every neuron in a human brain? The human brain has about 86 billion neurons. Sampling each one's voltage at 1 kHz with 10 bits (a modest choice, since Neuropixels samples its spike band at 30 kHz with 10-bit converters) gives about 8.6×10148.6 \times 10^{14}8.6×1014 bits per second, close to 101510^{15}1015, against 384 channels for one Neuropixels shank. Electronics in the head also make heat, and implant designs keep tissue warming within about 1 °C, which caps how much amplification, digitisation and transmission can sit inside the skull. Sorting spikes on the device and sending only events helps, but reaching every cell would also take a way of placing sensors through the whole volume without damaging it, which no current technology has. Why must spikes be sampled tens of thousands of times a second while EEG makes do with hundreds? A spike lasts about a millisecond and its sharp edges carry frequencies of several kilohertz, and a signal has to be sampled at more than twice its highest frequency to be captured, in practice several times more to keep its shape. Neuropixels therefore samples its spike band at 30 kHz, and spike sorting depends on those shapes to tell neighbouring neurons apart. EEG records summed synaptic currents that change slowly and are filtered further by the skull, with little useful power above a few hundred hertz, so hundreds of samples a second to a few thousand are enough.