Workflow hurdles with SpykingCircus2/Tridesclous2: SNR filtering, Probe loss, and Parameter documentation

[gaelle09]

Hi,

I am reaching out in hopes of getting help on quite a few questions I have regarding spike sorting using Spike interface. As a quick introduction, I’m a microengineering and data science master’s student writing her master’s thesis.

For reference, I am currently trying to locate spikes in an ephys recording obtained from a rat’s motor cortex using a 64-electrode array, where electrodes have been multiplexed. The goal is to observe the neural activity in response to an electrical stimulus in the form of a current at a certain intensity. I am looking to see the increase in neural activity, quantify it and locate it as well. My approach to attain that objective, is to make use of spike sorting algorithms. A fundamental aspect to consider is that I am working with very short recordings (5 to 8 sec) that are made up of concatenated individual 1 sec, in which a stimulation event occurs at 0.1sec. So, if we have an 8 sec recording, that means it contains 8 stimulation events and that means it is formed by 8 individual 1 sec recordings that have been recorded in the same day under the same stimulation conditions and in the exact same brain region of the rat (the electrode array was never removed).

I have thus run a data sample through the conventional preprocessing steps and sorted my data using tridesclous2 and spykingcircus2. I tweaked the parameters until I could see something. Here come my first questions. First, I couldn’t find a detailed and accurate description of each parameter. That made the task difficult as I don’t even really understand what I am doing although I had a slight idea but using Gemini, so, not the most reliable source.

In order to support my case, I will include images as well as a sorting output and notebook.

First question: what is the meaning of each parameter and/or where can I find that information?
Hereby I will paste the parameters that got me to find the next results I’ll be including.

Second question/problem: in Spykingcircus2’s parameters you can set a condition on the snr. I set that condition to only keep clusters with SNR>5 but in the resulting units I get some with lower SNR

Third question/problem: I attached my probe to the recording and verified it was still present after the preprocessing steps. But when running the sorters, I got a warning that no probe was attached. That poses quite a problem for me to rely on the location of the units and activity. I would thus need help with that please. Also, when the unit is localized, what does that correspond to in real life? (see images)

Fourth question/problem: I got some waveforms that look like spikes but are quite noisy and usually have an SNR of 2.5 to 4 which is quite low. How can I be sure that these are spikes and not noise? Also, when these waveforms are plotted, why are there multiple traces plotted that usually look quite flat? (see images)

Fifth question: I have realized that whenever I plotted anything, no axis-titles or units ever appeared. I know you can add them but in some cases I have no idea what is being plotted. Is there a way to resolve this?

Please excuse me for this raffle of questions. I hope you can help me.
Thank you very much in advance,

Best regards,
Gaelle
tridesclous2_something_promising_th_4.5_circus_omp_matching.zip

tridesclous2_output_th_4.5_circus_omp_matching.zip

[yger]

I would say that 8s long recordings are not enough to perform spike sorting, at least with any template-matching based algorithms such as Kilosort, Tridesclous, Spykingcircus or Lupin. If you are interested to know more about circus 2 or even lupin (now default in spikeinterface), have a look to https://www.biorxiv.org/content/10.64898/2026.01.23.701239v1
I could give you more details on parameters, but as said, 8 1s long recordings concatenated is not the best workflow, to say the least, for these softwares. You might only pick artefacts at the concatenated times, and you might not have enough spikes to properly cluster from. (Do not worry about the warning messages for the probes in between, they do not affect the results and the probe geometry your provided correctly)

[gaelle09]

Hi, thank you so much for your response. What would be the minimum required time for these algorithms to work? Are there any algorithms you know of that might work? Also, do you think spike sorting would work if I merged all recordings done at every stimulation intensity? Because before I used to treat the recordings done in the same stimulation conditions, if that makes sense.
Do you think the waveforms are spikes or probably not?
Thank you so much in advance

[chrishalcrow]

I think you should try a simpler algorithm. Maybe first just try peak detection, there are some docs here: https://spikeinterface.readthedocs.io/en/latest/how_to/analyze_neuropixels.html#detect-and-localize-peaks

The docs are about motion correction but doing peak detection and plotting the peaks might be helpful for you.

I'd also suggest trying a non-template-matching spike sorter: maybe mountainsort or herdingspikes? These do a "simple" peak detection then a clustering of the detected spikes. These should be better for short recordings, although 8s is very short... I think I'd want about 5 minutes of recording to do a spike sorting.

You can find out more about the parameters using e.g.

import spikeinterface.full as si
si.get_sorter_params_description('spykingcircus2')

[gaelle09]

I have tried mountainsort5 but couldn't find the right way to determine the parameters to obtain any good result. I will try on my full data. Would it be possible to have a detailed description of the parameters please? So I can maybe find a way to hopefully some results. My full data now is 137s, still shorter than 5 min...
Also, would you say the waveforms I have included in my prior message are spikes?
My final question would be what reasonable number of spikes would you expect for such a recording approximately?
Thank you very much for your help