Unexpected behavior on HD-MEA simulation ?

[b-grimaud]

Hi,
I'm trying to simulate data from in vitro neuronal cultures growing on a CMOS-based HD-MEA, specifically the 3Brain Accura chips.
I understand that this might be a bit beyond what is originally the aim of this tool, but it might be easier on some other points (e.g. in terms of drift, jitter).

I first made the following custom probe with MEAutility :

dim: 64
electrode_name: '3Brain_Accura'
description: "3Brain Accura CMOS"
pitch: 60
shape: 'square'
size: 10.5
plane: 'xy'
sortlist: null
type: 'mea'

Then I used it with the generate_templates_and_recordings notebook, I only changed the n_jobs parameter to -1 during template generation to speed things up a bit.

The output is very "patchy", and quite unusual :

plot_templates(recgen, single_axes=False, ncols=4) returns something like this :

plot_templates(recgen, single_axes=True, cmap='rainbow') returns :

Zoomed in :

plot_recordings(recgen, start_time=0, end_time=1, overlay_templates=True) :

Ideally, I'd like to reach an output like this :

Taken from this HerdingSpike example notebook.

Is this reasonably achievable ? Some additional changes would involve sampling frequency and bit depth, but from what I've seen that's doable, my main concern is wether in vivo cell models can be more or less accurately adpated to this setup.

Thanks !

[samuelgarcia]

@b-grimaud : spikeinterface.generation has now a full machinery to generate recording.
You should have a look.

[alejoe91]

Hi @b-grimaud

The issue is the plane. By default, cells are oriented as in a in vivo settings, and the plane should be yz. If you use xy, then the cells will be perpendicular to the MEA, which is not what you want.

So I suggest you try again with plane='yz' in your probe and also to set the rot field in the templates_params to xrot, so that cells will be randomly rotate around the x axis (perpendicular to the probe).

Let me know if that works!

[b-grimaud]

@samuelgarcia I wasn't aware of this, I'll check it out !

@alejoe91 Thanks for the advice !
I switched those parameters, things changed a bit but not as much.

From the logs during template generation, I saw that some already calculated models are being loaded from disk. Should I delete those to get the cells properly modelled again ?

[alejoe91]

Something is really weird in the traces...can you paste you're entire code? I'll try to reproduce the issue

[b-grimaud]

Here it is, almost entirely based on the example notebook :

import MEArec as mr
import MEAutility as mu
from pprint import pprint
import matplotlib.pylab as plt

%matplotlib widget

default_info, mearec_home = mr.get_default_config()
pprint(default_info)

mu.add_mea(r'./accura.yaml')

With accura.yaml as :

dim: 64
electrode_name: '3Brain_Accura'
description: "3Brain Accura CMOS"
pitch: 60
shape: 'square'
size: 10.5
plane: 'yz'
sortlist: null
type: 'mea'

# define cell_models folder
cell_folder = default_info['cell_models_folder']
template_params = mr.get_default_templates_params()
pprint(template_params)

template_params['n'] = 10
template_params['probe'] = '3Brain_Accura'
template_params['rot'] = 'xrot'
# the templates are not saved, but the intracellular simulations are saved in 'templates_folder'
tempgen = mr.gen_templates(cell_models_folder=cell_folder, params=template_params, n_jobs=-1)
# this will take a few minutes...

print('Templates shape', tempgen.templates.shape)
print('Sample locations', tempgen.locations[:3])
print('Sample rotations', tempgen.rotations[:3])
print('Sample cell types', tempgen.celltypes[:3])

# save templates in h5 format
mr.save_template_generator(tempgen, filename='data/test_templates_yz.h5')

recordings_params = mr.get_default_recordings_params()
pprint(recordings_params)

recgen = mr.gen_recordings(templates='data/test_templates_yz.h5', params=recordings_params)

print('Recordings shape', recgen.recordings.shape)
print('Selected templates shape', recgen.recordings.shape)
print('Sample template locations', recgen.template_locations[:3])
print('Number of neurons', len(recgen.spiketrains))
print('Sample spike train', recgen.spiketrains[0].times)

# save recordings in h5 format
mr.save_recording_generator(recgen, filename='data/test_recordings_yz.h5')

# plot templates
mr.plot_templates(recgen, single_axes=False, ncols=4)

# mr.plot_recordings(recgen)
mr.plot_recordings(recgen, start_time=0, end_time=1, overlay_templates=True)

[alejoe91]

Can you try the plot_templates with the tempgen?

[b-grimaud]

Here's the output for mr.plot_templates(tempgen, single_axes=False, ncols=4), I keep zooming into it but it doesn't look like there are hidden details, just those blocks.

For mr.plot_templates(tempgen, single_axes=True, cmap='rainbow') :

[alejoe91]

@b-grimaud actually everything seems to work on my end, and the problem could be in the overlay_templates. Here's a snippet of the data loaded and plotted with spikeinterface:

[alejoe91]

also the overlay templates seems to be ok..since you have so many electrodes on both axis, you need to zoom in in order to see them properly, since at each electrode location you will see the entire trace:

Can you make sure you can reproduce these plots? In that case, I think that it's expected behavior

[b-grimaud]

Sorry for the delay, here's the output :

Zoomed in :

It seems that the templates are mostly matching as expected ?
Regarding the template spread and density, I guess this can be fine tuned in the parameters ?

[alejoe91]

Ok great! Not too much fine tuning is possible, except for choosing some boundaries for the distance from the probe (x limits). In general the spread will depend on the model and it's positioning with respect to the probe.

What do you mean with density? If it's density of neurons, you can just increase the number of n_exc and n_inh in the recording generation so that you reach the desired density