MountainSort (a component of MountainLab) is spike sorting software developed by Jeremy Magland, Alex Barnett, and Leslie Greengard at the Center for Computational Biology, Flatiron Institute in close collaboration with Jason Chung and Loren Frank at UCSF department of Physiology. It is part of MountainLab, a general framework for data analysis and visualization.
MountainLab software is being developed by Jeremy Magland and Witold Wysota.
The software comprises tools for processing electrophysiological recordings and for visualizing and validating the results.
Contact the authors for information on the slack team for users and developers.
Announcement: As of November 21st, 2017, a beta release of the new version of MountainSort/MountainLab is available here: https://github.com/flatironinstitute/mountainsort. Please use that version going forward.
Please use the new version of MountainSort and MountainLab (see the announcement below). But here are the old installation instructions.
Please use the new version of MountainSort and MountainLab (see the announcement below). But here are the old instructions for the first sort
-
As of November 21st, 20017, a beta release of the new version of MountainSort/MountainLab is available here: https://github.com/flatironinstitute/mountainsort. Please use that version going forward.
-
ms3 - development branch with the ms3 processing pipeline
-
2017_06 branch - snapshot with only critical bug fix updates
-
Installation demo and introduction to Mountainlab -- skip to the end of the video to see the example sort and the GUI. Also describes how the software is organized and some of its philosophy.
-
Demo of additional WIP GUI for viewing very large datasets -- spikeview
Repo of unit (and not so unit) tests
The .prv data management system
Documentation on using annotation scripts will be forthcoming.
Because one of the goals of mountainsort is to enable reproducible spike sorting, we strongly advise against manual modifications that go beyond merging bursting clusters and perhaps rejecting certain noise clusters. Instead, we suggest that you export the cluster metrics along with the sorted clusters and then set cutoffs for inclusion of data in analyses based on those metrics. This will make it easy to describe your subsequent analyses and easy to determine how those analyses do or do not depend on cluster quality.
The isolation and noise overlap metrics we provide work well for the situations we have focused on, but they can be supplemented or replaced by other objective metrics as needed. Such metric processors may be included in the pipeline as post-processing plugins as C++, matlab, or python modules. Contact us if you you would like to contribute additional cluster metrics, or need help with integration.
A guide to using MountainSort with snippets, rather than continuous data acquisition
An old guide: Cluster metrics and automated curation
Thanks to all the users on the slack team for ongoing testing and feedback.
Barnett, Alex H., Jeremy F. Magland, and Leslie F. Greengard. "Validation of Neural Spike Sorting Algorithms without Ground-truth Information." Journal of Neuroscience Methods 264 (2016): 65-77. Link to arXiv
Magland, Jeremy F., and Alex H. Barnett. Unimodal clustering using isotonic regression: ISO-SPLIT. Link to arXiv