Welcome to Kilosort2, a MATLAB package for spike sorting electrophysiological data up to 1024 channels. In many cases, and especially for Neuropixels probes, the automated output of Kilosort2 requires minimal manual curation.
There is currently no preprint or paper for Kilosort2, so please read the wiki to find out how it works, and especially the drift correction section. Kilosort2 improves on Kilosort primarily by employing drift correction, which changes the templates continuously as a function of drift. Drift correction does not depend on a particular probe geometry, but denser spacing of sites generally helps to better track neurons, especially if the probe movement is large. Kilosort2 has been primarily developed on awake, head-fixed recordings from Neuropixels 1.0 data, but has also been tested in a few other configurations. To get a sense of how probe drift affects spike sorting, check out our "eMouse" simulation here and its wiki page.
To aid in setting up a Kilosort2 run on your own probe configuration, we have developed a graphical user interface where filepaths can be set and data loaded and visually inspected, to make sure Kilosort2 sees it correctly. The picture above is another GUI visualization: it shows the templates detected by Kilosort2 over a 60ms interval from a Neuropixels recording. The final output of Kilosort2 can be visualized and curated in the Phy GUI, which must be installed separately (we recommend the development version). Since Phy is in Python, you will also need the npy-matlab package.
Required toolboxes: parallel computing toolbox, signal processing toolbox, MATLAB >=R2016b
You must run and complete successfully mexGPUall.m in the CUDA folder. This requires mexcuda support, which comes with the parallel computing toolbox. To set up mexcuda compilation, install the exact version of the CUDA toolkit compatible with your MATLAB version (see here). On Windows, you must also install a CPU compiler, for example the freely available Visual Studio Community 2013. Note that the most recent editions of Visual Studio are usually not compatible with CUDA. If you had previously used a different CPU compiler in MATLAB, you must switch to the CUDA-compatible compiler using mex -setup C++. For more about mexcuda installation, see these instructions.
Navigate to the Kilosort2 directory and run kilosort:
>> cd \my\kilosort2\directory\
>> kilosort
See the GUI documentation for more details.
- Make a copy of
master_kilosort.mand\configFiles\StandardConfig_MOVEME.mand put them in a different directory. These files will contain your own settings, and you don't want them to be overwritten when you update Kilosort2. - Generate a channel map file for your probe using
\configFiles\createChannelMap.mas a starting point. - Edit the config file with desired parameters. You should at least set the file paths
ops.fbinary,ops.rootandops.fproc(this file will not exist yet -kilosortwill create it), the sampling frequencyops.fs, the number of channels in the fileops.NchanTOTand the location of your channel map fileops.chanMap. - Edit
master_kilosort.mso that the paths at the top (lines 3–4) point to your local copies of those GitHub repositories, and so that the configuration file is correctly specified (lines 6–7).
If you are unhappy with the quality of the automated sorting, try changing one of the main parameters:
ops.Th = [10 4] (default). Thresholds on spike detection used during the optimization Th(1) or during the final pass Th(2). These thresholds are applied to the template projections, not to the voltage. Typically, Th(1) is high enough that the algorithm only picks up sortable units, while Th(2) is low enough that it can pick all of the spikes of these units. It doesn't matter if the final pass also collects noise: an additional per neuron threshold is set afterwards, and a splitting step ensures clusters with multiple units get split.
ops.AUCsplit = 0.9 (default). Threshold on the area under the curve (AUC) criterion for performing a split in the final step. If the AUC of the split is higher than this, that split is considered good. However, a good split only goes through if, additionally, the cross-correlogram of the split units does not contain a big dip at time 0.
ops.lambda = 10 (default). The individual spike amplitudes are biased towards the mean of the cluster by this factor; 50 is a lot, 0 is no bias.
A list of all the adjustable parameters is in the example configuration file.
Kilosort2 provides a results file called rez, where the first column of rez.stare the spike times and the second column are the cluster identities. It also provides a field rez.good which is 1 if the algorithm classified that cluster as a good single unit. To visualize the results of Kilosort2, you can use Phy, which also provides a manual clustering interface for refining the results of the algorithm. Kilosort2 automatically sets the "good" units in Phy based on a <20% estimated contamination rate with spikes from other neurons (computed from the refractory period violations relative to expected).
Because Phy is written in Python, you also need to install npy-matlab, to provide read/write functions from MATLAB to Python.
Detailed instructions for interpreting results are provided here. That documentation was developed for Kilosort1, so things will look a little different with Kilosort2.
Kilosort2 by Marius Pachitariu
GUI by Nick Steinmetz
eMouse simulation by Jennifer Colonell
Please create an issue for bugs / installation problems.
This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with this program. If not, see http://www.gnu.org/licenses/.