How features are calculated#

The key parameters computed in driftplots are the spike amplitudes, depths and unit templates. Below are details on how these are computed, as the methods vary between SpikeInterface and across Kilosort versions.

Glossary

  • spike : a single action potential as recorded on the probe, and detected in the data by the sorter.

  • unit : A putative neuron that generated a set of spikes. During sorting, each spike is assigned to a unit.

  • spike waveform : the waveform (i.e. voltage signal as detected on the probe) of a single spike.

  • template : the canonical waveform that represents a unit. Can be thought of as the representative spike for a neuron (e.g. often it is computed as the average of all spike waveforms). Individual spike waveforms are similar to their template waveform.

  • whitened template : A whitening transform is typically applied to the data during preprocessing, to remove correlations between channels. Whitening has the effect of concentrating a spike waveform onto fewer channels and significantly changing the data scaling. Waveforms are often extracted from whitened data, meaning the unit templates are whitened. The whitening operation can be applied via matrix multiplication with an ‘whitening matrix’ and reversed by application of the ‘inverse whitening matrix’.

  • PC features. features.npy contains the PCA scores for each spike (i.e. by how much to scale the PC basis vectors to reconstruct the waveform). Here we refer to these as PC features similar to Kilosort.

../_images/example-waveform.png

An example waveform (y-axis is channel, x-axis is time).#

When loading kilosort outputs directly, all data is presented as int16, as saved on disk (no scaling to uV). In SpikeInterface, the scaling will depend on your data and the options chosen when creating the SortingAnalzyer.

In SpikeInterface, all spike amplitudes, depths and unit templates are directly computed by the SortingAnalzyer, from the passed recording. They are not reconstructed from Kilosort’s sorting output, and so can be quite different.

In general, spike amplitudes, depths and unit templates are calculated differently in SpikeInterface and across Kilosort versions and so are not directly comparable across sorters.

Amplitudes#

SpikeInterface

In SpikeInterface, the spike amplitude is calculated as the traces value at the spike peak on the channel in which the associated unit template has the strongest signal.

Kilosort

Kilosort 1, 2

In Kilosort 1,2 the amplitudes.npy file does not contain the actual spike amplitudes. Instead, it contains a scalar used to scale the template to give the best fit to the spike waveform. Therefore, taking the amplitude of the unwhitented template scaled by this amplitude value will give the approximate spike amplitude.

This matches how amplitudes are calculated in Nick Steinmetz spikes repository and Phy

Here, the amplitude is calculated as the peak-to-trough.

Kilosort 2.5, 3

The amplitudes.npy file and waveform reconstruction in Kilosort 2.5 and 3 in theory works the same as in Kilosort 1, 2. However, the inverse whitening matrix saved in these versions is not the true inverse whitening matrix, but instead a scaling matrix (see here and here). As such, the scaling here is likely not a good representation of the true spike waveforms. While the computed amplitudes are unlikely comparable to other Kilosort versions, but should be consistent across sessions sorted by the same sorter.

Kilosort 4

In Kilosort 4, the amplitudes.npy does not contain a scaling value as in previous Kilosort versions. Instead, it calculates a proxy for the amplitude as the l2 norm over the PC features (over both all channels and all PCs) i.e. like an energy over the feature space.

To approximate the spike amplitude, we use the method described in this Kilosort issue. We are currently investigating how this method compares to SpikeInterface and other Kilosort versions.

Templates (interactive mode)#

In Kilosort 1-3, the whitened templates are always displayed. Here the template is the universal or learned canonical waveform for each unit, used in template matching. They are not unwhitened for consistency, as it is not possible to unwhiten the templates in Kilosort 2.5 and 3 due to the missing inverse whitening matrix issue mentioned above.

In Kilosort4, the unit templates are the average spike waveform (reconstructed from PCs) (after whitening, filtering, and drift correction). These are also displayed in their whitened form.

The displayed unit templates are not scaled to the clicked spike amplitude because to the authors knowledge this value is not available in Kilosort 4.

In SpikeInterface, the templates are also computed from the spike waveforms, using the operator chosen when computing the template extension. Whether the templates are whitened or unwhitened will depend on the preprocessing chain used for the recording passed to the SortingAnalzyer, on which these waveforms and templates are computed.

Spike Depths#

In general the depth (y-position) of each spike is estimated from its waveform.

In Kilosort 1-3, the center of mass of the PCA scores (found in features.npy) is used, as computed in Nick Steinmetz’s spikes repository. and Phy.

In Kilosort 4, we use the saved values from spike_positions.npy.

In SpikeInterface, the depths are taken from the "spike_locations" extension.