Modules for processing extracellular electrophysiology data from Neuropixels probes, originally developed at the Allen Institute for Brain Science. This fork has been modified to run with SpikeGLX data, including integration of CatGT (preprocessing), C_Waves(calculation of SNR and mean waveforms) and TPrime (synchronization across data streams). The data can be sorted with any version of Kilosort; a version of IBL's pykilosort is also available.
Code including modifications for SpikeGLX https://github.com/jenniferColonell/ecephys_spike_sorting
Original repo from the Allen Institue https://github.com/AllenInstitute/ecephys_spike_sorting
The general outline of the pipeline is preprocessing, spike sorting by Kilosort 4, Kilosort 3.0, Kilosort 2.5 or Kilosort 2.0, followed by cleanup and calculation of QC metrics. The original version from the Allen used preprocessing specifically for data saved using the Open Ephys GUI. This version is designed to run with data collected using SpikeGLX, and its associated tools (CatGT, TPrime, and C_Waves). The identification of noise clusters is unchanged from the original code. Calculation of QC metrics has been updated to work with any Neuropixels probe type, rather than assuming NP 1.0 site geomtery; also, the metrics code can now be run on phy output after manual curation.
The spikeGLX_pipeline.py script implements this pipeline:
This code is still under development, and we welcome feedback about any step in the pipeline.
Further documentation can be found in each module's README file. For more information on Kilosort2, please read through the GitHub wiki.
-
catGT_helper: Concatenates trials, applies filters, removes artifacts in neural data. Finds edges in sync and auxiliary channels.
-
kilosort_helper: Generates config files for MATLAB versions of Kilosort based on SpikeGLX metadata and launches spike sorting via the Matlab engine. [ks4_helper] runs the python-based Kilosort 4. [pykilosort helper]((ecephys_spike_sorting/modules/pykilosort_helper/README.md) runs the IBL version of pykilosort; the only changes in fork used in this pipeline are to allow completely skipping filtering and CAR, because these functions are handled in CatGT.
-
kilosort_postprocessing: Removes putative double-counted spikes from Kilosort output. The algorithm has been changed from the original to delete all between cluster duplicates from the cluster with lower amplitude.
-
psth_events: Reformat list of events from an auxiliary channel for phy psth plots.
-
noise_templates: Identifies noise units based on their waveform shape and ISI histogram or a random forest classifier.
-
mean_waveforms: Extracts mean waveforms from the raw data, given spike times and unit IDs. Also calculates metrics for each waveform. In this version the mean waveforms are calculated using Bill Karsh's efficient C_Waves tool.
-
quality_metrics: Calculates quality metrics for each unit to assess isolation and sorting quality.
-
tPrime_helper: Maps event times (edges in auxiliary channels, spike times) in all streams to match a reference stream.
-
depth_estimation: Uses the LFP data to identify the surface channel. Updated to use site geometry from SGLX metedata. Currently does not feed this result to kilosort. Can be run in any part of the processing after CatGT, if LFP processing has been performed.
-
extract_from_npx: Calls a binary executable that converts data from compressed NPX format into .dat files (continuous data) and .npy files (event data)
-
median_subtraction: Calls a binary executable that removes the DC offset and common-mode noise from the AP band continuous file. CatGT CAR replaces this function for SpikeGLX data.
(Not used) automerging: Automatically merges templates that belong to the same unit (included in case it's helpful to others).
These modules have been tested with Python 3.8.10 and 3.9.
If you only plan to use only the MATLAB version of Kilosort, you can install and run using the procedure recommended by the original authors at the Allen Institute, which uses pipenv.
If you want to run Kilosort4 or pykilosort, or just prefer Anaconda, please skip down to Installation with Anaconda and Kilosort4 or Installation with Anaconda and pykilosort .
All of the components of the SpikeGLX pipeline are available in Windows and Linux, but the pipeline has only been tested in Windows. These instructions are for Windows 10.
If the computer doesn't already have python, install it; the current version of the pipeline environmet requires at least 3.8. The currently tested version is 3.8.10. Download the Windows x86-64 executable installer and run the exe, selecting the "Add Python to PATH" checkbox at the bottom of the dialog.
If you forget to check the the "Add to PATH" box, it can be added afterward by editing the Environment Variables (under Advanced system settings). The two paths to add are to the Python folder containing the exe, and the scripts folder, e.g.:
C:\Users\labadmin\AppData\Local\Programs\Python\Python38 C:\Users\labadmin\AppData\Local\Programs\Python\Python38\Scripts
If you have another version of Python installed, this version can be installed side by side. To use these installation instructions, version 3.8 will need to have priority in the environment PATH variable.
Open the Windows command prompt as administrator, and install pipenv:
$ pip install --user pipenvThe pipenv executable will be in:
C:\Users\labadmin\AppData\Roaming\Python\Python38\Scripts
Add this path to the PATH environment variable. If you have paths to other versions in PATH, this one will need to be first in the search list for pipenv to use the correct version.
Close the command prompt, and reopen as a user (not as administrator) for the next steps.
Clone (or download and unzip) the repo. (https://github.com/jenniferColonell/ecephys_spike_sorting)
In the command window navigate to the ecephys_spike_sorting directory at the top level of the repo, e.g.:
cd \Users\labadmin\Documents\ecephys_clone\ecephys_spike_sorting
Build the environment -- it will use the Pipfile located in this directory, and create the virtual environment in the local directory. Currently (April 2024) the latest version of setuptools appears to not function with installation of MATLAB, so after the install, we activate the environment and use pip to uninstall setuptools and install 59.8.0. Finally, install the ecephys code in the environment.
$ set PIPENV_VENV_IN_PROJECT=1
$ pipenv install
$ pipenv shell
(.venv) $ pip uninstall setuptools
(.venv) $ pip install setuptools==59.8.0
(.venv) $ pip install phylib
(.venv) $ pip install .The python version and MATLAB version need to be compatible. For Python 3.8, this requires MATLAB 2020b or later. The code has been tested only with MATLAB 2021b.
Install MATLAB 2021b. Side by side installations of MATLAB are fine, so there is no need to delete earlier versions, and running code specific to an earlier version should be possible.
Open MATLAB 2021b, and enter the command gpuDevice(). You make get a message that there are no GPU devices with compatible drivers. Later versions of MATLAB also require more recent drivers for the GPU card. MATLAB 2021b requires version 10.1 or later of the Nvidia drivers.
If you get that message, quit MATLAB. Update the drivers for the GPU card; this can be done with the Device Manager in Windows 10, and will also happen automatically if you update the CUDA Toolkit. The pipeline has been tested with CUDA Toolkit 11.2 (by the way, all CUDA toolkit versions are backward compatible to older hardware). After updating, restart MATLAB and enter gpuDevice() again to make sure it is recognized.
The MATLAB engine for python must be installed in the local instance of python run by the virtual environment. Open the command prompt as administrator, navigate to the ecephys directory, and enter:
$ pipenv shell
(.venv) $ cd <matlabroot>\extern\engines\python
(.venv) $ python setup.py installReplace with the root directory of your MATLAB 2021b installation, for example:
C:\Program Files\MATLAB\R2021b
For more details about installing the python engine, see the MATAB documentation:
https://www.mathworks.com/help/matlab/matlab_external/install-the-matlab-engine-for-python.html
NOTE: This install needs to be repeated whenenver the virtual environment is rebuilt (e.g. after creating a new clone or download of the repo).
After completing the install, close the command window and reopen as a normal user (not administrator) to run scripts.
These instructions are to build an environment compatible with KS4 and MATLAB versions of Kilosort. The instructions are adapated from the Kilosort4 github (https://github.com/MouseLand/Kilosort)
If not already present, install Miniconda with python 3.9 (https://docs.conda.io/en/latest/miniconda.html).
Install Kilosort 4 and pytorch (you can pick your name for the environment:
conda create --name ks4_ece python=3.9
conda activate ks4_ece
python -m pip install kilosort[gui]
pip uninstall torch
conda install pytorch pytorch-cuda=11.8 -c pytorch -c nvidiaAs with pipenv, to be compatible with versions of MATLAB < R2021, install an earlier version of setuptools:
pip uninstall setuptools
pip install setuptools==59.8.0It's a good idea at this point to run a small test dataset through the kilosort gui. There are tips for debugging issues with the pytorch installation in the KS4 readme.
Next install ecephys. To force the correct versions of some components, they must be uninstalled and reinstalled manually. This will be corrected in a later verison. From the anaconda prompt, navigate to the ecephys_spike_sorting directory (containing setup.py) and run the commands:
pip install -e .
pip uninstall argschema
pip install argschema==1.17.5
pip uninstall marshmallow
pip install marshmallow==2.19.2
pip install h5py
pip install phylibTo run the MATLAB versions of Kilosort in this environment, follow the instruction below in Set up to run MATLAB from Python in Anconda
Ensure that CUDA Toolkit 11.2 or later is installed. The pipeline is currently tested with 11.2.
If not already present, install Miniconda with python 3.9 (https://docs.conda.io/en/latest/miniconda.html).
Create a folder to hold the environments. Use git bash to clone ecephys repo (https://github.com/jenniferColonell/ecephys_spike_sorting) and pykilosort repo (https://github.com/jenniferColonell/pykilosort).
Open an Anconda prompt and navigate to the pykilosort folder. Create an environment to hold both pykilosort and ecephys using the yml, and activate when complete.
conda env create -f ./ece_pyks2.yml
conda activate ece_pyks2As discussed in the pipenv inatallation above, the current version of setuptools appears to be buggy. Replace it:
pip uninstall setuptools
pip install setuptools==59.8.0To install pykilsort and other components, navigate to the pykilosort directory and run the commands:
pip install -e .
pip install cython
conda install -c conda-forge pyfftw
pip install git+https://github.com/int-brain-lab/ibllib.git
pip install -U phylibNext install ecephys. To force the correct versions of some components, they must be uninstalled and reinstalled manually. This will be corrected in a later verison. From the anaconda prompt, navigate to the ecephys_spike_sorting directory (containing setup.py) and run the commands:
pip install -e .
pip uninstall argschema
pip install argschema==1.17.5
pip uninstall marshmallow
pip install marshmallow==2.19.2
pip install h5pyThe python version and MATLAB version need to be compatible. To be compatible with python 3.9, the MATLAB version must be 2021b or later.
Install MATLAB 2021b. Side by side installations of MATLAB are fine, so there is no need to delete earlier versions.
Open MATLAB 2021b, and enter the command gpuDevice(). You make get a message that there are no GPU devices with compatible drivers. Later versions of MATLAB also require more recent drivers for the GPU card. MATLAB 2021b requires version 10.1 or later of the Nvidia drivers.
If you get that message, quit MATLAB. Update the drivers for the GPU card; this can be done with the Device Manager in Windows 10, and will also happen automatically if you update the CUDA Toolkit. The pipeline has been tested with CUDA Toolkit 11.2 (by the way, all CUDA toolkit versions are backward compatible to older hardware). After updating, restart MATLAB and enter gpuDevice() again to make sure it is recognized.
The MATLAB engine for python must be installed in the local instance of python run by the virtual environment. Open an Anaconda prompt as administrator, activate the environment, and then navigate the setup script for the MATLAB engine:
conda activate ece_pyks2
cd <matlabroot>\extern\engines\python
python setup.py installReplace with the root directory of your MATLAB 2021b installation, for example:
C:\Program Files\MATLAB\R2021b
For more details about installing the python engine, see the MATAB documentation:
https://www.mathworks.com/help/matlab/matlab_external/install-the-matlab-engine-for-python.html
NOTE: This install needs to be repeated whenenver the virtual environment is rebuilt (e.g. after creating a new clone or download of the repo).
After completing the install, close the Anaconda window and reopen as a normal user (not administrator) to run scripts.
CatGT, TPrime, and C_Waves are each available on the SpikeGLX download page. To install, simply download each zipped folder and extract to a convenient location, see the instructions here. The paths to these executables must then be set in create_input_json.py.
NOTE: The pipeline is now compatible with the latest CatGT. If you are updating the pipeline, make sure you also get the most recent versions of CatGT, TPrime, and C_Waves.
Parameters are set in two files. Values that are constant across runs�like paths to code, parameters for sorting, etc � are set in create_input_json.py. Parameters that need to be set per run (run names, which triggers and probes to process�) are set in script files.
In create_input_json.py, be sure to set these paths and parameters for your system:
- ecephys_directory: parent directory that contains the modules directory
- kilosort_repository
- KS2ver -- needs to be set to '2.5' or '2.0', and be correct for the repository
- npy_matlab_repository
- catGTPath: contains the CatGT.exe file
- cWaves_path: contains the C_Waves.exe file
- tPrimePath: contains the TPrime.exe file
- kilosort_output_temp (see note below)
Note: The kilosort_output_temp folder contains the kilosort residual file and also temporary copies of the config and master file. With kilosort 2.5, this "temporary" file--which has been drift corrected--may be used for manual curation in phy. If you want it to be kept available, set the parameter ks_copy_fproc=1; then a copy will be made with the kilosort output and the params.py adjusted automatically.
Other rarely changed parameters in create_input_json.py:
- Most Kilosort and pykilosort parameters.
- kilosort post processing params
- quality metrics params
Read through the parameter list for create_input_json.py to see which parameters are already passed in and therefore settable per run from a calling pipeline script. These currently include the threshold parameter for Kilosort, switches to include postprocessing steps within Kilosort, and radii (in um) to define the extent of templates and regions for calculating quality metrics. These radii are converted to sites in create_input_json.py using the probe type read from the metadata.
The scripts generate a command line to run specific modules using parameters stored in a json file, which is created by the script. Create a directory to hold the json files, e.g.
\Users\labadmin\Documents\ecephys_clone\json_files
There are two example scripts for running with SpikeGLX data:
sglx_multi_run_pipeline.py Meant to process multiple SpikeGLX runs, especially with multiple probes. The threshold for kilosort and the refractory period for the quality metrics are set per probe by specifying a brain region parameter for each probe. A first pass through all the probes in a run generates json parameters files for CatGt and sorting+post processing, and a second loop actually calls the processing. Finally runs TPrime. See comments in the script file for parameter details.
sglx_filelist_pipeline.py Meant for running sorting/postprocessing modules on collections of preprocessed data, independent of the standard SpikeGLX run structure.
For either script, edit to set the destination for the json_files, and the location of the input run files. Edit the list of modules to include those you want to run. For the full pipeline script, you also need to set the CatGT and TPrime parameters.
These scripts are easy to customize to send the output to different directories.
To run scripts in pipenv, open a Windows command line, navigate to the ecephys_spike_sorting\scripts directory and enter:
pipenv shell
(.venv)$ python <script_name.py>To run scripts in Anaconda, open an Anacodna prompt, activate the environment, navigate to the ecephys_spike_sorting\scripts directory and enter:
conda activate ece_pyks2
python <script_name.py>If you manually curate your data in phy, you can recalculate mean waveforms and quality metrics for the curated clusters. You'll need to run a script that skips preprocessing and sorting, and just runs the mean_waveforms and metrics modules. The required changes in sglx_multi_run_pipeline.py are:
- Set variable run_CatGT = False
- Set variable runTPrime = False
- Only include mean_waveforms and quality_metrics in the list of modules to be called, e.g.
modules = [
#'kilosort_helper',
#'kilosort_postprocessing',
#'noise_templates',
#'psth_events',
'mean_waveforms',
'quality_metrics'
]
When the mean_waveforms and metrics modules are re-run the first time, these output files are preserved with their old names:
- metrics.csv
- waveform_metrics.csv
- clus_Table.npy
These output files are renamed with an added "_0":
- mean_waveforms.npy -> mean_waveforms_0.npy
- cluster_snr.npy -> cluster_snr_0.npy
The new output files are numbered by the latest version. Output files from the first re-run are named:
- metrics_1.csv
- waveform_metrics_1.csv
- clus_Table_1.csv
- mean_waveforms_1.npy
- cluster_snr_1.npy
Another re-run will create a full set with _2, etc
These modules require Python 3.5+, and have been tested with Python 3.5, 3.6, and 3.7.
Three of the modules (extract_from_npx, median_subtraction, and kilosort_helper) have non-Python dependencies that will need to be installed prior to use.
We recommend using pipenv to run these modules. From the ecephys_spike_sorting top-level directory, run the following commands from a terminal:
$ pip install --user pipenv
$ export PIPENV_VENV_IN_PROJECT=1
$ pipenv install
$ pipenv shell
(ecephys_spike_sorting) $ pip install .You can now edit one of the processing scripts found in ecephys_spike_sorting/scripts and run via:
(ecephys_spike_sorting) $ python ecephys_spike_sorting/scripts/batch_processing.pySee the scripts README file for more information on their usage.
To leave the pipenv virtual environment, simply type:
(ecephys_spike_sorting) $ exitIf you don't have it already, install homebrew. Then, type:
$ brew install pipenv
$ export PIPENV_VENV_IN_PROJECT=1
$ pipenv install
$ pipenv shell
(ecephys_spike_sorting) $ pip install .You can now edit one of the processing scripts found in ecephys_spike_sorting/scripts and run via:
(ecephys_spike_sorting) $ python ecephys_spike_sorting/scripts/batch_processing.pySee the scripts README file for more information on their usage.
To leave the pipenv virtual environment, simply type:
(ecephys_spike_sorting) $ exit $ pip install --user pipenv
$ set PIPENV_VENV_IN_PROJECT=1
$ pipenv install
$ pipenv shell
(.venv) $ pip install .Note: This will work in the standard Command Prompt, but the cmder console emulator has better compatibility with Python virtual environments.
You can now edit one of the processing scripts found in ecephys_spike_sorting\scripts and run via:
(.venv) $ python ecephys_spike_sorting\scripts\batch_processing.pySee the scripts README file for more information on their usage.
To leave the pipenv virtual environment, simply type:
(.venv) $ exitThis code is an important part of the internal Allen Institute code base and we are actively using and maintaining it. The implementation is not yet finalized, so we welcome feedback about any aspects of the software. If you'd like to submit changes to this repository, we encourage you to create an issue beforehand, so we know what others are working on.
See Allen Institute Terms of Use
© 2019 Allen Institute for Brain Science