Thalamocortical network - Traub et al. 2005

Known issues » History » Version 9

« Previous - Version 9/21 (diff) - Next » - Current version
Helena Głąbska, 30 Apr 2014 14:57
rest of the result : Traub model and trying to reproduce fig 7 from the paper


Known issues with Traub et al 2005.

This is a quite complex and detailed model and as discussed in the original paper

Any model, even of a small bit of cortex, is subject to difficulties and hazards: limited data, large numbers of parameters, criticisms that models with complexity comparable to the modeled system cannot be scientifically useful, the expense and slowness of the necessary computations, and serious uncertainties as to how a complex model can be compared with experiment and shown to be predictive.
The above difficulties and hazards are too real to be dismissed readily. In our opinion, the only way to proceed is through a state of denial that any of the difficulties need be fatal. The reader must then judge whether the results, preliminary as they must be, help our understanding.

Even the published Fortran version of this model was acknowledged to be incomplete. Each conversion of this model will deviate to a small or large extent from this version.

Questions about physiological properties of model

Dependence on Fast Regular Bursting cells for oscillatory behaviour

Prevalence of gap junctions

High current threshold for deep pyramidal firing

Not tested with external synaptic input

Limitations of the conversion of the model to NEURON

It is useful to read the notes on conversion of this model to NEURON from Fortran by Tom Morse and Michael Hines

Slightly different method of running the simulation (e.g. in Neuron information about spike is sent immediately, in Fortran every 0.1 ms )

Diffrent behaviour of NMDA synapse when thalamus is disconnected (some bug in Fortran version?)

In Fortran code:

 z = 0.d0  ! thalamus disconnected
 gAMPA_TCR_to_suppyrRS = z * gAMPA_TCR_to_suppyrRS
 gNMDA_TCR_to_suppyrRS = z * gNMDA_TCR_to_suppyrRS
 gAMPA_TCR_to_suppyrFRB = z * gAMPA_TCR_to_suppyrFRB
 gNMDA_TCR_to_suppyrFRB = z * gNMDA_TCR_to_suppyrFRB
...

gNMDA_TCR_to_suppyrFRB becomes 0. Then when you compute NMDA activation
from TCR to suppyrFRB

....

! NMDA part
        if (delta.le.5.d0) then
       gNMDA_suppyrFRB(k,L) = gNMDA_suppyrFRB(k,L) +
     &  gNMDA_TCR_to_suppyrFRB * delta * 0.2d0
        else
       dexparg = (delta - 5.d0)/tauNMDA_TCR_to_suppyrFRB
          if (dexparg.le.5.d0) then
          z = dexptablesmall (int(dexparg*1000.d0))
         else if (dexparg.le.100.d0) then
          z = dexptablebig (int(dexparg*10.d0))
         else
          z = 0.d0
         endif
       gNMDA_suppyrFRB(k,L) = gNMDA_suppyrFRB(k,L) +
     &  gNMDA_TCR_to_suppyrFRB * z
        endif
c Test for NMDA saturation
       z = NMDA_saturation_fact * gNMDA_TCR_to_suppyrFRB
       if (gNMDA_suppyrFRB(k,L).gt.z)
     &  gNMDA_suppyrFRB(k,L) = z
! end NMDA part
....

It seems that this piece of code, more precisely the last three lines:

c Test for NMDA saturation
 z = NMDA\_saturation\_fact \* gNMDA\_TCR\_to\_suppyrFRB
 if (gNMDA\_suppyrFRB(k,L).gt.z)
 & gNMDA\_suppyrFRB(k,L) = z

kills completely NMDA activation of suppyrFRB cells from all the other populations, not just TCR (except from nontuftRS cells, nontuftRS - suppyrFRB NMDA conductance is calculated after this block). In Neuron version there is no such behaviour.

An updated version of this model in NEURON is being worked on here. The version allows to modify easily the network, e.g. to add new population (version commited on 26 June 2013 and later), replace one template by another e.g. tuftIB Traub cell by Hay cell ( version commited on 04 July 2013 or later). The main groucho.hoc file is simpler and much shorter (about 10 times), parameters like AMPA, GABA, NMDA conductances, connections between populations are defined in separated files.

Tests for Neuron and Fortran version . Trying to reproduce results from article

Remark: In Fortran version, compilations flag -finit-local-zero , seems to be important.

Single Cell

TO DO

Figure 2

TO DO

Figure 7

Figure 7 from the article

7A

In article they raported about consisting of 17 burst complexes that terminate spontaneously. The last 5 of the bursts are shown. Results from the Fortran version are very similar but only 14 bursts appears. In Neuron version the result is much different.

You can download the data for Fig 7A from Fortran and Neuron simulation: Fortran data and Neuron data.
For Neuron simulation you can also compare the result with simulation using the “traub_exact()” algoritm: Neuron traub_excat() data. More about “traub_excat()” algoritm you can read in notes on conversion of this model to NEURON from Fortran by Tom Morse and Michael Hines.

7B
Results from the Fortran version looks again very similar, although gives much more complex bursts, at least 6, when prolong the simulation up to 2000 ms (results not shown here - download ) .

You can download the data for Fig 7B from Fortran and Neuron simulation: Fortran data and Neuron data compare with Neuron traub_excat() data .

7C

You can download the data for Fig 7C from Fortran and Neuron simulation. Fortran data and Neuron data compare with Neuron traub_excat() data .

7D

You can download the data for Fig 7D from Fortran and Neuron simulation. Fortran data and Neuron data compare with Neuron traub_excat() data

Limitations of the conversion of the model to MOOSE

TODO…

Limitations of the conversion of the model to NeuroML

Optimal spatial discretisation for each cell needs to be investigated

Important details of the process of conversion of the cell models to NeuroML, and matching cell behaviour across simulators is present in the 2010 NeuroML paper.

The spatial discretisation of the cells influenced precise spike timing. Changing the number of compartments/points used to calculate the membrane potential changed the timing of the cell (e.g. changing the value of nseg in NEURON on all sections). See below for an example of how 3 cells with differing numbers of compartments converged at different rates. A) Nucleus reticularis thalami (nRT) cell; B) Superficial Low Threshold spiking (LTS) cell; C) Layer 6 Non-tufted Regular Spiking pyramidal cell. Traces for NEURON (black) and MOOSE (green) and GENESIS (red).

NMDA conductance wave form

The NMDA synapse model used in the network has an unconventional form, with a scaling factor rising lineally between 0 and 5ms, and decaying exponentially. This can probably be approximated by a double exponential synapse (coupled with v & [Mg] dependent blocking mechanism).

Firing rate vs. injected current of cells

Many of the cells show unusual F/I curves.

Support in NeuroML

All model elements from the neuroConstruct generated network can be exported to valid NeuroML v1.8.1.

Model can be exported to (mostly valid) NeuroML 2, but there is not yet an application that can handle such detailed NML2 models (but we’re working on it).