「SpikeSorter」を使用した電気生理学的記録をソートするビジュアルガイド

The overall goal of this video, is to show electrophysiologists how to use the program, SpikeSorter, to detect and sort spikes in extracellular recordings from neural tissue made with microelectrode arrays. Our goal was to provide a simple solution to this task of spike sorting using a program with a user-friendly interface. SpikeSorter can read files in a variety of formats and offers a variety of pre-processing and event detection strategies, in addition to sorting and data export options.

The implications of this technique extends towards the analysis of neural behaviors, studied by extracellular recordings with multielectrode arrays. Generally, individuals new to this method will struggle because of the complexity of the task and the lack of standard methodology and objective validation. We first had the idea for this method, when we were faced with the problem of sorting our own recordings, done with 54 channel multi electrode arrays and found no available solution.

To begin this procedure, open the program. Go to file and open, and then select the recording file format from the drop-down list, at the bottom right of the resulting open file dialog. Next, select the file and click open.

Click OK when reading has finished. The recording signals are now shown in the voltage waveform display. If the recording is unfiltered and contains the local field potential, remove it by clicking on the filter icon on the toolbar.

Select high pass butterworth filter, a suitable cutoff frequency and the number of poles, then press Do It.Press OK, then done, then inspect the filtered waveform in the display. Voltage signals can be inspected by holding down the left mouse button and then dragging. Use the scroll wheel to zoom in on particular features and out again.

Hover the mouse over specific waveform points to display corresponding voltage values in the top left hand corner. Double-clicking on the window, brings up a dialog that can be also used to quickly examine any part of the recording and to examine specific events in channels or clusters, once event detection and sorting have been completed. Next, check for channels that may be faulty and need to be masked.

Click on the channel check icon and inspect the graph. To mask an outlying channel, either select the channel number or select it from the problem list. When finished, close the dialog, then click yes at the prompt, to save the mask values.

Masked channels are grayed out in the voltage display. In this procedure, go to pre-process event detection to bring up the event detection dialog. Use the slider on the top right to inspect the noise level on the particular channels.

Careful inspection of the voltage display, may also reveal silent or unusually noisy channels that need to be masked. Choose variable thresholding for event detection. The recommended range of thresholds is from 4.0 to 6.0 times noise.

Select 4.5. Choose the dynamic multiphasic filter detection method from the drop-down list. Set the window to be roughly half the width of a typical spike waveform, with the values in the range, 0.15 to 0.5 milliseconds.

After that, select the alignment method by choosing the option that best identifies a single, temporally localized feature of the spikes that are being sorted. Leave other options at their default values and press Start. Press Done to exit the dialog.

The voltage waveform window shows the detected events in gray on their assigned center channel. Event detection parameters can be adjusted based on the results seen at this point. For example, event detection may be rerun with a lower detection threshold, in order to detect more events.

The next step is not normally performed before routine sorting, but it is useful to do it when sorting for the first time, or when encountering unfamiliar data. Go to sort, convert channels to clusters, to create a single cluster for each unmasked electrode channel, assuming that each channel has some events assigned to it. Examine these clusters by going to review, view, clean and split clusters to bring up another dialog.

Use the spin control to select the cluster to be viewed. Select a cluster that has two or more distinct subclusters, then press the realign button to change the time of each event, so as to better match it to the shape of the template. Next, press Autosplit to cluster the events based on the principal components'distribution.

Subsequently, click on the checkboxes to examine the waveforms in the subclusters. Use the slider control to examine the corresponding waveforms. As an exercise, use one of the small split buttons to create a new cluster and examine it.

Notice how the red cluster, now reveals two subclusters more clearly, after the green one is split off. Sorting could continue manually this way, but it is preferable to use the faster Autosort procedure. Press Done and go to Sort, Autosort.

Leave the skip event detection option checked if event detection has already been done, as here. If it is not checked, event detection will be run automatically, using parameter values and choices inherited from the event detection dialog. In the clustering panel below, select a temporal window large enough to contain the entirety of the spike waveform, preceding and following the alignment point.

Next, select a realignment option to be used during clustering and choose a minimum cluster size. Decide on the number of dimensions in the PC space that will be used for clustering and leave other options at their default settings. After that, press Start to begin Autosort.

After Autosort is complete, press next to go to the manually guided merge and split stage, then press begin. In the new dialog, cluster pairs that are not clearly similar or distinct will be presented in sequence. For each presented pair, use the buttons to choose whether to merge the pair, merge and re-split it, to mark the pair as distinct or as ambiguous, if it is unclear as to what is the best decision.

Click on the checkboxes, to display a graph of either spike peak height or the first or second principal component against time or auto and cross correlation histograms. These may help with decisions. Here, we see an example of a merge decision.

Followed by a merge and split decision. If the merge and split option is unable to find clearly separable clusters, use the slider in the prompted dialog to manually vary a clustering parameter and find a split that looks satisfactory. If this cannot be achieved, use the revert button to go back to the original state of the two clusters.

When finished, press split as shown. Now go to review, post-processing. Use the alignment cleaning button to remove events from the clusters that are bad matches to their cluster template.

Then, use the recluster button to reassign unclustered events to particular templates based on match value. The reclaimed events are marked as a subcluster of each parent cluster and can be inspected, using the view, clean and split clusters dialog. These events will remain in the cluster and be exported as such, unless they are deleted.

Next, return to the post-processing dialog. Use the delete button and the spin control next to it, to delete the clusters with a signal-to-noise ratio less than the selected threshold. Subsequently, use the sort button to renumber the clusters according to a chosen criterion.

This figure shows the display, obtained by going to view, sorted waveforms for the recording that was just sorted. Common experience, is that waveforms for a cluster pair on the same channel, look identical, but the compare pairs dialog nevertheless, shows distinct clusters in the PC projection. Merging decisions following the previous stages of sorting are sometimes suggested by amplitude time plots and cross correlograms.

This illustrates a merging decision based on the presence of a strong asymmetric, cross correlation at short time intervals, coupled with a clear relationship between the first principal component values of the two units, shown here on the vertical axis. This figure shows a case where similar evidence for merging is lacking. The cross correlogram is not strongly asymmetric and the autocorrelograms are different.

Though waveform shapes are similar, the two units should arguably not be merged, because of the difference in the principal components'distributions. This figure shows a case where waveform shapes are similar and the first principal component values of two units blend when one of them stops firing and the other starts. In this case, the decision to merge seems straightforward.

We hope this brief introduction to SpikeSorter has been useful. We plan to continue to develop the program and add new features where they can be useful. Towards this end, we welcome bug reports, constructive feedback and suggestions for improvement.