NML cell and NetPyNE cell do not match

[HengyeZhu]

update!!!: The current RE folder is brand new (the previous one was too messy), so please check the issue messages from November onwards.

The ih file stores data exclusively for the insertion of ih channel.
The cal data also appears somewhat unusual—the calcium channels don't seem to be functioning. However, here the NML and NetPyNE implementations do match.

[sanjayankur31]

I'll take a look. To confirm, the files with "*.v.png" are the NeuroMl generated files, and the other PNG file is the NetPyNE one?

[HengyeZhu]

Apologies for the confusion. I've already renamed the file.

[sanjayankur31]

Plots to generate for full morphology cell later:

https://github.com/OpenSourceBrain/BahlEtAl2012_ReducedL5PyrCell/tree/master/NeuroML2#after-conversion-to-neuroml

[sanjayankur31]

Calcium dependent channels and info:

NeuroML/NeuroML2#117

[HengyeZhu]

Sir, individual ion channels (+ pas channel) match between NML and NetPyNE, but difference still exists when all are combined together.
https://github.com/OpenSourceBrain/Macaque_auditory_thalamocortical_model_data/tree/feat-neuroml-gsoc/NeuroML2/compare_MC/IT2/compare

[sanjayankur31]

OK, that sounds good. I'll take a look and try to see what's going on.

[HengyeZhu]

The single-compartment cell also exhibited different behavior, where the NML (NeuroML) model appears to show less hyperpolarization.

https://github.com/OpenSourceBrain/Macaque_auditory_thalamocortical_model_data/blob/feat-neuroml-gsoc/NeuroML2/compare_SC/IT2/compare/Figure_1.png

[sanjayankur31]

I'm looking at the different comparison charts, and I think we need to modify the input protocols a bit. For example:

On the left is the plot for cal (with pas), on the right for only pas, but they look identical---so the cal mechanism isn't actually doing anything, which implies that we aren't quite testing it properly.

I'm going through it all in a new branch feat-add-compare-readme, will open a PR when I have some changes to suggest:

https://github.com/OpenSourceBrain/Macaque_auditory_thalamocortical_model_data/tree/feat-add-compare-readme/

[HengyeZhu]

caConcExt always equlas zero. I think this is the cause.

Resolved (just not properly recorded)

[HengyeZhu]

In the hh2ad channel (HH2.mod), only ina is noticeably mismatched. I checked several times but couldn't find where the issue lies. I also reduced dt (from 0.01ms to 0.001ms).

see https://github.com/OpenSourceBrain/Macaque_auditory_thalamocortical_model_data/blob/feat-neuroml-gsoc/NeuroML2/compare_MC/RE/compare/hh2ad/compare_hh2.png

[HengyeZhu]

I suddenly realized that the logic I used for verification wasn't rigorous enough—I should have included a diff graph. Then I noticed an issue: when simConfig.recordStep = 0.1 (ms), there is a floating-point accumulation error (NetPyNE might not use integer step counts for recording), as seen in tut2.py from http://doc.netpyne.org/tutorial.html. However, when simConfig.recordStep = 1.0 (ms), there is no floating-point accumulation error (which should prove my point), as demonstrated in tut5.py.

Therefore, I adopted a 1.0 ms sampling interval and found that even the leaky current of the RE Cell (with only soma) does not fully match. I will look into the underlying implementation next.

I will post the specific code tomorrow, as the current code is a bit messy.

BTW, when dt=0.0001 (ms), the error is nearly two orders of magnitude smaller, but this time step is impractical.

Update!!!
The accumulation of floating-point errors in netpyne does not affect the recording of voltage. I mean, it only appears to have errors, but the voltage is still recorded at the exact moments of 2.5E-5, 5E-5, 7.5E-5, etc.

[HengyeZhu]

I forgot that the current simulation uses jLEMS, and since jLEMS appears to use an explicit Euler method, the presence of some error is not unexpected.

[HengyeZhu]

I created the RE cell using Neuron's Python API and found that the difference between netpyne and neuron is always zero, while neuroml and neuron do not completely match. It seems the issue lies in the inadequate model conversion.

Update:
The issue of voltage mismatch in the passive case has been resolved. The reason was that the numerical precision of the voltage output from neuroml2 differed from that of Neuron. After I applied rounding to the Neuron data, the difference became consistently zero.

[HengyeZhu]

I wrote a neuron file by hand and had it use the hoc and mod files exported from neuroml. I found no significant errors, so the issue should be narrowed down to the conversion of ion channels.

But I really can't figure out where it went wrong :C

[HengyeZhu]

The ion channels exported from NeuroML, after conversion by mod2c, use the explicit Euler method in the final .c file instead of the cnexp form. However, when I tried to modify the exported file and recompile it, although it changed to the cnexp format, it still does not match the original mod file, which is a bit strange.

I am now directly comparing Neuron and NeuroML, as I have verified that the difference between NetPyNE and Neuron is zero.

here