Recording many neurons at once with a silicon probe

Neurons communicate with action potentials, brief electrical impulses lasting about a millisecond. An electrode placed in the brain outside the cells picks up a small trace of each action potential from the neurons nearby: a spike of a few tens to a few hundred microvolts. Silicon probes such as Neuropixels carry hundreds of recording sites along a thin shank, so they record the spikes of many neurons at once. In the first report of Neuropixels probes, two probes recorded more than 700 well-isolated neurons at the same time in an awake mouse (Jun et al., 2017).

The recording does not say which neuron fired each spike. Every site picks up several neurons, and every neuron appears on several neighbouring sites. Spike sorting works this out: the signal is filtered, spikes are detected with a threshold, each spike is described by a few numbers (its features), and spikes with similar features are grouped into clusters, one per neuron if all goes well. Clusters are then checked by hand, because automatic methods both merge neurons and split them.

Every decision of a real recording and sort is made here: how the probe enters the brain, the spacing of the sites, the filter, the reference, the threshold, the number of clusters and which clusters count as single neurons. The simulation records the true source of every spike, so the effect of each choice on the result is shown at the end. The recording lasts 8 seconds.

Learning objectives

• Explain why an extracellular spike is small and negative, and how its amplitude falls with distance from a recording site
• Separate the local field potential from spikes with a band-pass filter, and explain why spikes are not detected on the wideband signal
• Estimate the noise from the median absolute deviation and choose a detection threshold, weighing noise crossings against missed spikes
• Use the footprint of a spike across neighbouring sites and principal components to cluster spikes, and merge or split clusters
• Judge clusters with refractory-period violations and isolation distance, and label single units, multi-unit activity and noise
• Explain why neurons from one animal are not independent replicates, and choose the unit of analysis from the question

Ethics and safety

Recording with probes in living mice involves regulated procedures: fitting a head-plate and making a small opening in the skull under general anaesthesia with pain relief, and either a recording under anaesthesia or repeated recordings in an awake, head-fixed animal that has been accustomed to the set-up. They need approval by an Institutional Animal Ethics Committee; in India such work is governed by the CCSEA under the Prevention of Cruelty to Animals Act 1960, with equivalent frameworks in the European Union (Directive 2010/63/EU), the United States (Animal Welfare Act and PHS Policy) and the United Kingdom (Animals (Scientific Procedures) Act 1986). The procedures follow the approved protocol and are learned under supervision. The 3Rs apply: recording hundreds of neurons per animal supports Reduction, slow insertion, habituation and good post-operative care are Refinements, and simulations such as this one can replace some practice. Studies are reported following the ARRIVE guidelines, developed by the NC3Rs in the United Kingdom and endorsed by journals internationally (Percie du Sert et al., 2020), giving the numbers of animals, sessions and units and the curation criteria. Probes are fragile and expensive; the electrical safety of the recording rig is checked by trained staff.

Key terms

• Silicon probe: A thin shank of silicon carrying many recording sites, made with the methods used for computer chips. Neuropixels probes have nearly a thousand sites on a shank about 1 cm long.
• Recording site: A small metal contact on the probe that measures the voltage in the fluid next to it.
• Extracellular spike: The trace of an action potential recorded outside the cell: a brief negative deflection of tens to hundreds of microvolts.
• Local field potential (LFP): The slow part of the extracellular signal, below a few hundred hertz, produced mainly by the summed synaptic currents of many neurons.
• Band-pass filter: A filter that keeps a range of frequencies, here about 300 to 6000 Hz for spikes, and removes slower and faster signals.
• Common-average referencing: Subtracting, at every moment, the average or median of all sites from each site, which removes noise shared by all of them.
• Median absolute deviation (MAD): The median of the absolute deviations of a signal from its median. Divided by 0.6745 it estimates the SD of Gaussian noise, and spikes hardly affect it.
• Detection threshold: The voltage, set as a multiple of the noise SD, that the signal must cross for an event to count as a spike.
• Footprint: The pattern of a neuron's spike amplitudes across neighbouring sites, which depends on where the neuron lies relative to the shank.
• Principal component analysis: A method that finds the few combinations of measurements along which the data vary most, used to describe each spike by a handful of features.
• Cluster: A group of spikes with similar features, ideally all from one neuron. k-means divides the spikes into k clusters.
• Refractory-period violation: An interval between two spikes of one cluster shorter than a neuron's refractory period (about 1.5 ms here), showing that the cluster holds more than one source.
• Isolation distance: A measure of how well a cluster is separated from the other spikes in feature space; larger values mean better isolation.
• Single unit and multi-unit: A single unit is a cluster of one neuron's spikes; multi-unit activity is a cluster holding spikes of several neurons that could not be separated.