Remove artifacts recording not used in Kilosort

[aradiawilms]

Hello! I have been using spikeinterface's run_sorter function in two ways:

sorting = si.run_sorter('kilosort4', blanked_rec, folder=outputFolder, docker_image=False, verbose=True, remove_existing_folder=False, raise_error=True, **params_kilosort4)

where (1), blanked_rec is a variable (a RemoveArtifactsRecording) that comes from running this:

blanked_rec = si.remove_artifacts(recording=recording_to_blank, list_triggers=mStim_sample_idx, ms_before=ms_before, ms_after=ms_after) 

(we want to remove the artifact of microstimulation on the recording)

and where (2), blanked_rec is a loaded recording of (1) saved (a BinaryFolderRecording):

blanked_rec.save(folder=output_path)
...
blanked_rec=si.load(blanked_filename) 

However, in both cases we found that despite the blanked recording looking good when manually scrolling through it, kilosort seems to use an incorrect recording. When doing manual curation in phy, some spikes are detected during the period that is blanked, which is visible when plotting the spiketimes relative to the TTLs that were blanked (picture attached). Many units also have suspiciously low spikecounts.

I think (1) didn't work because the RemoveArtifactsRecording is only a wrapper to the actual recording underneath. But I cannot explain why (2) didn't work either. Was there anything else wrong in my approach?

Any help is greatly appreciated!

[samuelgarcia]

Hi.
Sorry ffor this.
What happen you you do plot_tarces in the blanked zone, is it flat in the the zone ?

[aradiawilms]

Hi! Using a custom function that uses get_traces and then plots a blanked time window:

fs = blanked_rec.sampling_frequency

    for i, mStim in enumerate(mStim_TTLs):
        # Convert time window (seconds) into sample indices
        start_frame = int((mStim - 0.2) * fs)
        end_frame = int((mStim + 0.4) * fs)

        # fetch traces for a subset of channels (first 50 here)
        # updated by AW to avoid loading all data into memory at once
        traces = blanked_rec.get_traces(
            start_frame=start_frame,
            end_frame=end_frame,
            channel_ids=blanked_rec.get_channel_ids()[:50]  # first 50 channels
        )

        # Create figure and axis
        fig, ax = plt.subplots(figsize=(10, 6))

        # Plot each channel with an offset for clarity
        offset = 200  # µV spacing between channels
        for ch in range(traces.shape[1]):
            ax.plot(traces[:, ch] + ch * offset, linewidth=0.5)

        # Formatting
        ax.set_title(f"Stimulus {i} ({ms_before} ms before, {ms_after} ms after)")
        ax.set_xlabel("Samples")
        ax.set_ylabel("Channels (offset)")

        # Save figure
        fileName = blankedFolder.joinpath(f'blankedStim{i}.png')
        plt.savefig(fileName, dpi=300, bbox_inches='tight')
        plt.close(fig)

The windows are flat:

While plot_traces returns non flat traces

This is the case for both (1) and (2).

[samuelgarcia]

oups. This looks like a bug.

[alejoe91]

@aradiawilms could you share the recording and the code?

Is it possible that the recording has timestamps or a non-zero start time? in that case, the plot traces would use the recording timestamps, which might offset your expected stimuli.

[alejoe91]

Note that both (1) and (2) should give the same traces. I also use remove_artifact to remove optogenetics stimulations and get the correct blanking in the output traces!

[alejoe91]

@aradiawilms any update?

[aradiawilms]

Hi,

Unfortunately, we can't send a recording and the full code because it is unpublished at the moment.

However, I don't think nonzero start times are the cause. I also don't think the remove_artifacts is the problem as both the custom plotting of the traces and the spikeglx viewer show the recordings to be blanked correctly. That's why it seems to be a problem when passing to kilosort? If I run the same data in kilosort directly (through the GUI, selecting the file I saved using a custom save to binary function), there are no spikes in the blanked window.