Spike sorting with data from tungsten microwire arrays

[jeonghyeon0park]

I have tried to understand the situation and have reviewed the wiki multiple times, but I couldn’t find an alternative approach to writing this article.

We are encountering problems while sorting spikes recorded using the Tucker Davis Technologies Zif64 array. Each channel is spaced 250 μm apart, as shown in the image below. (I suspect that a few spikes were detected across multiple channels and may not provide enough resolution for spike sorting)
Spike data were acquired from each channel at 24,414 Hz and exported as a binary file resembling the Neuropixels output format. Periods without spikes were filled with Gaussian noise.
Time period without spike was filled with gaussian noise.

Details of probe
https://www.tdt.com/docs/hardware/zif-clip-based-microwire-arrays/#zif-clip-based-microwire-array-specifications

To address the issue, I reduced the value of "nearest chans" to minimize the region of spike detection. Values between 2 and 5 appeared to work well.

The results was like this. No single good unit could be detected.

My questions :

1. Can Kilosort4 handle this type of data (where spikes overlap across a few channels)? Or should we consider downgrading to an earlier version or exploring a different approach?
2. If sorting is possible, are there any recommended parameter settings or example cases I can refer to?

[jacobpennington]

Hello, a few things to start with:

1. Update Kilosort4 to the latest version.
2. Set dminx to 250, then check the "universal templates" checkbox under the probe preview in the GUI. With large contact spacing, you want to end up with a single white dot (template position) lined up with each green square (contact).
3. Check the "grouping centers" checkbox under the probe preview. Since your channels are so far apart, they should really be clustered separately, so you should see four columns of yellow circles pop up. If you only see two columns (one for each shank), try setting x_centers = 4.

If you still encounter issues after trying those solutions, please upload kilosort4.log from the new run (and any relevant screenshots).

[jeonghyeon0park]

I updated kilosort4 and set dminx as 500 (because setting them as 250 results in white dots between green squares)

Although this approach restricted unit detection to a single channel, I couldn't find a good unit with a refractory period.

These are my results with new parameters:


kilosort4.log

[jacobpennington]

dminx = 250 was not working because you have set max_channel_distance to 200. You should instead use dminx = 250 with the default max_channel_distance = 32, and in general you should always use the default values for all parameters unless necessary.

I also noticed your data in the GUI screenshots looks like it's dominated by synchronized activity across all channels. Given the geometry of the probe, I'm assuming that's some kind of noise / artifact. Sorting is not going to work well no matter what settings you've chosen if your data is dominated by noise across all channels.