Skip to content
dsp-kitchen

Kilosort4

Kilosort4 sorts spikes from high-density extracellular probes (Neuropixels, polytrodes). It detects spikes with universal templates at many virtual positions on the probe, clusters them in local feature spaces, learns multi-channel templates, and resolves overlapping spikes by template deconvolution, with drift correction.

In dsp-kitchen it lives in the dspk-sorter-kilosort crate. It is written from the paper and the published defaults, with upstream’s code read to match its behaviour and no code copied. So the crate is GPL-3.0, like upstream.

What is implemented

Section titled “What is implemented”
Stage Status
Whole-recording run (Kilosort4::run, dspk sort): halo windows streamed on the device implemented
Preprocessing fitted on the recording (common reference, 300 Hz high-pass, local whitening) implemented (device)
Universal templates learned from the recording (wPCA, wTEMP) implemented (device)
Predefined universal templates (wTEMP.npz) implemented
Universal-template spike detection, wPCA features, position implemented (device)
Graph-based clustering of the detected spikes (bipartite k-NN graph, merging tree, bimodality splits) implemented (device)
Learned templates: the units’ templates aligned and near-duplicates merged implemented (device)
Learned-template matching with matching pursuit, background-subtracted features implemented (device)
Clustering of the matched spikes into the final units, with the refractory (CCG) criterion implemented (device)
Export as a SortingOutput (one unit per cluster, with its waveform template) implemented
Global merges (waveform similarity + refractory CCG), duplicate-spike removal, good / mua labels implemented
Drift correction not yet

Universal templates

Section titled “Universal templates”

Two arrays drive detection:

  • wPCA (n_pcs × nt): a single-channel temporal basis; features are projections onto it.
  • wTEMP (n_templates × nt): single-channel waveform shapes (“universal templates”).

By default (templates_from_data = true) both are learned from each recording; Kilosort4 also publishes a predefined set (6 × 61 each) used when templates_from_data = false. There are no other “pretrained weights”.

Citation

Section titled “Citation”

Pachitariu M., Sridhar S., Pennington J., Stringer C. Spike sorting with Kilosort4. Nature Methods (2024). doi:10.1038/s41592-024-02232-7

@article{pachitariu2024kilosort4,
  title   = {Spike sorting with {Kilosort4}},
  author  = {Pachitariu, Marius and Sridhar, Shashwat and Pennington, Jacob and Stringer, Carsen},
  journal = {Nature Methods},
  year    = {2024},
  doi     = {10.1038/s41592-024-02232-7}
}
Code MouseLand/Kilosort, GPL-3.0; defaults followed from v4.1.3
Predefined templates wTEMP.npz — osf.io/download/6807fb5958b763aae139aa60, 3432 bytes, SHA-256 cae1c96f8f4150be0a39627515750b70c4bc3548177cf487ae3c013f1ca6abd8 (catalog id kilosort4/wtemp-v1)

Run it

Section titled “Run it”
Terminal window
dspk sort kilosort4 recording.bin -o sorted --format phy

Every setting and its default: Kilosort4 settings. How it works: pipeline. What to change: tuning.