A silicon probe that is inserted into the mouse brain can precisely measure the activity of about 200 individual neurons simultaneously. This tool should improve our ability to study functional neuronal circuitry. See Letter p.232
Countless videos depict neurons communicating with one another, flashing on and off as they signal to their neighbours. But these videos are a fantasy, constructed to illustrate how we think that the brain might work. Real videos of neuronal populations at work in live brains lack either the spatial resolution or the temporal precision to infer functional interactions. On page 232, Jun et al.1 describe a silicon probe called Neuropixels, which can simultaneously record the activity of more than 200 individual neurons. This technology promises to have a major impact on neuroscience.
There have been intensive efforts to increase the speed, density and number of simultaneous recordings taken from single neurons. Developments in optical approaches to measure activity have increased density and cell-count recordings2, but such approaches cannot reliably identify the precise timing of spikes of neuronal activity, and recordings are typically limited to structures near the brain's surface. By contrast, electrical recordings can resolve spikes from individual neurons with millisecond precision2, and thin electrodes can penetrate deep into the brain with minimal damage. But taking dense electrical recordings is challenging.
These challenges arise from the properties of the brain itself. Neurons are tiny (their cell bodies are about 10 micrometres across) and densely packed (about 100,000 per cubic millimetre in the brain's cerebral cortex)3. To measure electrical activity accurately, electrodes must be placed very close to cell bodies, and must be designed such that spikes of activity from different neurons are distinguishable from one another.
To meet these challenges, modern silicon-based probes contain arrays of closely spaced recording contacts. Each contact can pick up spikes from many neurons, and the spikes from any one neuron can be detected by multiple contacts2,4 — the activity of individual neurons can then be dissected by comparing signals from each contact. This approach works best if contacts are densely packed. However, more contacts take up more space and require more wires. The devices become thicker and cause more damage when inserted, potentially killing the neurons they are intended to interrogate.
Neuropixels represents a dramatic leap forward from previous probes. The advance was possible thanks to four non-profit funding agencies, whose large financial commitment enabled neuroscientists and engineers to work together.
Jun et al. used next-generation silicon-fabrication techniques to produce a probe that has 960 recording contacts. These channels are set roughly 20 μm apart, along a probe that has a cross-section of 70 μm × 20 μm and is 1 cm long, meaning that even the deepest structures of a mouse brain can be accessed (Fig. 1). Although silicon probes with a similar contact density exist, they are limited to far fewer channels (a typical probe might have 32 channels), and often require too many wires for chronic implantation. The on-probe electronics and wiring of Neuropixels can accommodate simultaneous recordings from 384 of its 960 channels, and the channels being measured can be switched. Switching offers the substantial benefit of allowing active recording sites to be altered after implantation, so that structures of interest can be selectively sampled.
The process began with a prototype design, test and redesign loop that systematically benchmarked materials, electronics and configurations to identify the best possible design for a laboratory probe. In early prototyping experiments, Jun et al. tested how the width of the probe affected tissue damage. Standard probes in the field range in width from 50 μm to more than 200 μm, depending on the number and configuration of recording sites. The authors aimed to maximize the number of recording sites, while minimizing the size of the probe to decrease tissue damage. They found little difference in the number of neurons sampled per channel between probes 50 μm and 70 μm wide, suggesting that the wider probe causes negligible differences in tissue damage. They also showed that there was no disadvantage to using probes with switches.
At the same time, Jun and colleagues evaluated recording materials. The material typically used to make recording contacts (gold covered in a polymer) is not compatible with the electronics used in Neuropixels. The authors found that probes based on titanium nitride performed as well as those made of conventional materials when used in animals for six to eight weeks — long enough to perform most behavioural experiments.
In a second phase of testing, Jun et al. examined various electronics configurations. On-site amplification of the signals recorded from each contact might have been expected to improve signal quality, but in fact provided minimal benefits. Furthermore, unamplified probes could be used in tandem with optogenetics experiments, in which genetically modified neurons are activated or inactivated using light. By contrast, amplification made probes too light-sensitive for usable recording during optogenetic stimulation. The ability to study neuronal activation in optogenetically manipulated circuits is extremely desirable; therefore, no on-site amplification was used.
Finally, the authors designed the cutting-edge electronics of their probe such that data from all 384 channels could be digitized and compressed into just two output wires. This design enables the device to be very small — compatible with long-term attachment to small animals and causing minimal interference with natural behaviour.
Jun et al. tested Neuropixels across different brain structures, and found that the device typically identified an average of about 0.6 isolated neurons per recording contact (and up to nearly 1 in some structures). A total of well over 200 neurons are typically recorded simultaneously per probe during a single experiment. Probes with fewer channels yield comparable numbers of neurons per channel5, so Neuropixels should detect more than 10 times more neurons than the commonly used 32-channel probes.
Recording from more neurons obviously yields much more data. It also confers another advantage — when many neurons are recorded simultaneously, connected pairs of neurons can be identified. If one neuron has a reliably increased probability of firing immediately after another, a direct connection between them can be inferred. However, previous work5 indicates that less than 0.2% of all possible pairs of simultaneously recorded neurons share a direct connection. Recording from 20 neurons simultaneously yields 190 possible pairs, of which less than 1 would be expected to be connected. Recording from 200 neurons yields a possible 19,900 pairs, with about 40 expected connections. Thus, Neuropixels promises to make recording from connected pairs commonplace, allowing direct tests of the relationships between connectivity and function across both local and distant brain structures.
In summary, the development of Neuropixels ushers in a new era of microelectrodes. The versatility of this technology opens the door to exciting experiments aimed at studying functional neuronal circuits. For example, multiple Neuropixels probes can be implanted in animals trained to perform complex behaviours, to record how the behaviour changes neural activity in several distinct brain regions, or to see how activity across these regions is correlated. Further, the use of optogenetics in combination with the probe enables analysis of the effects of manipulating specific cell types on larger circuits. Such data are likely to provide tremendous insights into neural mechanisms of perception and behaviour.