Golgi cell model overstimulation leads to permanent hyperexcitability

[Helveg]

The above trace represents the Golgi cell firing patterns during a simulation where they are initialized with the following stimulation:

        "initial": {
          "io": "input",
          "device": "spike_generator",
          "targetting": "cell_type",
          "cell_types": ["golgi_cell", "purkinje_cell", "basket_cell", "stellate_cell"],
          "section_type": "dendrites",
          "synapses": ["AMPA", "AMPA_PF"],
          "parameters": {
            "noise": true,
            "start": 0,
            "interval": 5,
            "number": 20,
            "sd": 5
          }
        },
        "mossy_fiber_init": {
          "io": "input",
          "device": "spike_generator",
          "targetting": "by_id",
          "targets": [86, 92, 93, 100, 67, 72, 110, 103],
          "cell_types": ["mossy_fibers"],
          "synapses": ["AMPA", "NMDA"],
          "parameters": {
            "noise": true,
            "start": 0,
            "interval": 10,
            "number": 15
          }
        },

[Helveg]

@SteMasoli via mail:

To me it looks like the model is under an hyper-polarizing synaptic current that prevents it to fire correctly.

Hyperpolarizing? Which synapse or mechanism on the model could cause that? And is it logical that it causes the state change to permanent hyperexcitability?

NMDA on the apical dendrites? Issue

This synapse is not used in the granule to golgi connections, but it can be activated to some extend when the mossy fibers are firing. Since the AMPA current is way larger let's correct the NMDA issue seperately (#28)

Stimuli

initial

Effect: The initial stimulus by itself causes no problems on the Golgi cell.
Target: AMPA synapse on a random dendritic section.
Stimulus: 20 ± 5 spikes at 200Hz

          "io": "input",
          "device": "spike_generator",
          "targetting": "cell_type",
          "cell_types": ["purkinje_cell", "golgi_cell", "basket_cell", "stellate_cell"],
          "section_type": "dendrites",
          "synapses": ["AMPA", "AMPA_PF"],
          "parameters": {
            "noise": true,
            "start": 0,
            "interval": 5,
            "number": 20,
            "sd": 5
          }

mossy_fiber_init

Effect: Pushes the golgi cell into the problematic state described at start of thread. Activates 8 mossy fibers, which in turn stimulate the Golgi cells directly on 160 AMPA (and possibly 160 extra NMDA) synapses, but also activates a bunch of granule cells that contact the Golgi cell with ±6 AMPA synapses (1 PF, 5 AA) per granule cell.
Target: Any of the glomerulus_to_golgi compartments (dendrites, both basal and apical). Both AMPA and NMDA (if available on the target section).
Stimulus: 15 ± 5 spikes at 100 Hz (but see section below for full impact)

Stimulus

To calculate the size of the stimulus we need to know how many granule cells per golgi cell fire, and how many spikes they generate in response to the 15 spikes at 100Hz.

• Each golgi contacted is by 7300 granule cell PFs and by 1000 AA:
  • 7300 * 1 + 1000 * 5 = 12300 AMPA synapses coming from granule cells
• The 8 activated MF contact about 7000 out of 12000 GC
  • 7000 / 12000 = 0.58333 = 0.6 of all granule cells will be firing
• We confirmed that in response to 10 spikes @ 100Hz on 2 dendrites is 10 + 1 spikes at 100Hz again, let's assume this is slightly increased for 4 MF to 13 spikes or so.
• This means that the average Golgi cell receives 13 spikes at 100 Hz on 0.6 * 12300 = 7380 AMPA synapses

Could you try to test this @SteMasoli? Start sending at 0ms 13 noisy spikes at 100Hz to 7380 synapses on random dendritic locations on the Golgi cell and see if it gets stuck in the hyperactive state.

If it doesn't: double the amount of synapses.
If it does:

• Try the same protocol but starting at 300ms to see if the Golgi becomes immune after reaching some steady state
• Halve the amount of synapses at t=0ms and see if it doesn't get stuck anymore

If you would like to investigate the role of NMDA in the issue you could add 160 NMDA synapses, send some signals to them and see what happens

[claudiacasellato]

"Each golgi contacted is by 7300 granule cell PFs and by 1000 AA:
7300 * 1 + 1000 * 5 = 12300 AMPA synapses coming from granule cells"

why? from convergence and tables in paper , I would say 1381 * 1 + 400 * 2.8 = 2500

[Helveg]

From pizdaint.hdf5 which should be the v5 network:

>>> cs = scaffold.get_connectivity_set("ascending_axon_to_golgi")
>>> cs.convergence
1062.2413793103449
>>> cs = scaffold.get_connectivity_set("parallel_fiber_to_golgi")
>>> cs.convergence
7361.310344827586

Edit: Sorry, I checked the code and my numbers don't take into account that there are multisynaptic contacts. I'll update asap

[Helveg]

@claudiacasellato My calculations still say PF-GC convergence is 7361 and AA-GC convergence 384 (with on average 2.76 synapses)

>>> cs = scaffold.get_connectivity_set("parallel_fiber_to_golgi")
>>> len(cs) / scaffold.statistics.cells_placed["golgi_cell"]
7361.310344827586

So that brings the new count to 7361 PF synapses + 1062 AA synapses = 8423 AMPA synapses and 8423 * 0.6 = 5054 firing AMPA synapses

[AliceGem]

Mmmm we should check the method used for conv and div values, but also be sure about the file..can you apply the same method to the .hdf5 file V5 in the dropbox folder?

[Helveg]

Mmmm we should check the method used for conv and div values, but also be sure about the file..can you apply the same method to the .hdf5 file V5 in the dropbox folder?

Thx, this solved the difference in conv/div values (after a fix to the scaffold functions)

Opened a PR at https://github.com/Helveg/cerebellum-scaffold/pull/472

[Helveg]

Fixed by moving to a new Golgi cell model in #29