ModelDB is moving. Check out our new site at https://modeldb.science. The corresponding page is https://modeldb.science/266802.

NMDA receptors enhance the fidelity of synaptic integration (Li and Gulledge 2021)

 Download zip file 
Help downloading and running models
Accession:266802
Excitatory synaptic transmission in many neurons is mediated by two co-expressed ionotropic glutamate receptor subtypes, AMPA and NMDA receptors, that differ in their kinetics, ion-selectivity, and voltage-sensitivity. AMPA receptors have fast kinetics and are voltage-insensitive, while NMDA receptors have slower kinetics and increased conductance at depolarized membrane potentials. Here we report that the voltage-dependency and kinetics of NMDA receptors act synergistically to stabilize synaptic integration of excitatory postsynaptic potentials (EPSPs) across spatial and voltage domains. Simulations of synaptic integration in simplified and morphologically realistic dendritic trees revealed that the combined presence of AMPA and NMDA conductances reduces the variability of somatic responses to spatiotemporal patterns of excitatory synaptic input presented at different initial membrane potentials and/or in different dendritic domains. This moderating effect of the NMDA conductance on synaptic integration was robust across a wide range of AMPA-to-NMDA ratios, and results from synergistic interaction of NMDA kinetics (which reduces variability across membrane potential) and voltage-dependence (which favors stabilization across dendritic location). When combined with AMPA conductance, the NMDA conductance balances voltage- and impedance-dependent changes in synaptic driving force, and distance-dependent attenuation of synaptic potentials arriving at the axon, to increase the fidelity of synaptic integration and EPSP-spike coupling across neuron state (i.e., initial membrane potential) and dendritic location of synaptic input. Thus, synaptic NMDA receptors convey advantages for synaptic integration that are independent of, but fully compatible with, their importance for coincidence detection and synaptic plasticity.
Reference:
1 . Li C, Gulledge AT (2021) NMDA receptors enhance the fidelity of synaptic integration eNeuro
Model Information (Click on a link to find other models with that property)
Model Type: Neuron or other electrically excitable cell;
Brain Region(s)/Organism:
Cell Type(s): Dentate gyrus granule GLU cell; Hippocampus CA3 pyramidal GLU cell;
Channel(s): I K; I Na,t;
Gap Junctions:
Receptor(s): AMPA; NMDA;
Gene(s):
Transmitter(s): Glutamate;
Simulation Environment: NEURON;
Model Concept(s): Synaptic Integration;
Implementer(s):
Search NeuronDB for information about:  Dentate gyrus granule GLU cell; Hippocampus CA3 pyramidal GLU cell; AMPA; NMDA; I Na,t; I K; Glutamate;
/
nmda_models
Figure_2_synaptic_thresholds
0_kv.mod *
0_na.mod *
0_nmda.mod *
0_syn_g.mod *
BallStick_NoSpines.hoc
BallStickActive.hoc
BallStickCell.hoc *
Description.txt
init_BallStick.hoc
init_BallStick_Active.hoc
init_BallStick_NoSpines.hoc
makeSavestates.hoc *
makeSavestatesActive.hoc
makeSavestatesNoSpines.hoc
NoSpines_Threshold_Template.hoc
Threshold_Template.hoc *
                            
COMMENT
na.mod

Sodium channel, Hodgkin-Huxley style kinetics.  

Kinetics were fit to data from Huguenard et al. (1988) and Hamill et
al. (1991)

qi is not well constrained by the data, since there are no points
between -80 and -55.  So this was fixed at 5 while the thi1,thi2,Rg,Rd
were optimized using a simplex least square proc

voltage dependencies are shifted approximately from the best
fit to give higher threshold

Author: Zach Mainen, Salk Institute, 1994, zach@salk.edu

26 Ago 2002 Modification of original channel to allow 
variable time step and to correct an initialization error.
Done by Michael Hines(michael.hines@yale.e) and 
Ruggero Scorcioni(rscorcio@gmu.edu) at EU Advance Course 
in Computational Neuroscience. Obidos, Portugal

11 Jan 2007 Fixed glitch in trap where (v/th) was where (v-th)/q is. 
(thanks Ronald van Elburg!)

20110202 made threadsafe by Ted Carnevale
20120514 repl aced vtrap0 with efun, which is a better approximation
         in the vicinity of a singularity

Special comment:

This mechanism was designed to be run at a single operating 
temperature--37 deg C--which can be specified by the hoc 
assignment statement
celsius = 37
This mechanism is not intended to be used at other temperatures, 
or to investigate the effects of temperature changes.

Zach Mainen created this particular model by adapting conductances 
from lower temperature to run at higher temperature, and found it 
necessary to reduce the temperature sensitivity of spike amplitude 
and time course.  He accomplished this by increasing the net ionic 
conductance through the heuristic of changing the standard HH 
formula
  g = gbar*product_of_gating_variables
to
  g = tadj*gbar*product_of_gating_variables
where
  tadj = q10^((celsius - temp)/10)
  temp is the "reference temperature" (at which the gating variable
    time constants were originally determined)
  celsius is the "operating temperature"

Users should note that this is equivalent to changing the channel 
density from gbar at the "reference temperature" temp (the 
temperature at which the at which the gating variable time 
constants were originally determined) to tadj*gbar at the 
"operating temperature" celsius.

ENDCOMMENT

NEURON {
    THREADSAFE
	SUFFIX na
	USEION na READ ena WRITE ina
	RANGE m, h, gna, gbar
	GLOBAL tha, thi1, thi2, qa, qi, qinf, thinf
	RANGE minf, hinf, mtau, htau
	GLOBAL Ra, Rb, Rd, Rg
	GLOBAL q10, temp, tadj, vmin, vmax, vshift
}

UNITS {
	(mA) = (milliamp)
	(mV) = (millivolt)
	(pS) = (picosiemens)
	(um) = (micron)
} 

PARAMETER {
	gbar = 1000   	(pS/um2)	: 0.12 mho/cm2
	vshift = -10	(mV)		: voltage shift (affects all)
								
	tha  = -35	(mV)		: v 1/2 for act		(-42)
	qa   = 9	(mV)		: act slope		
	Ra   = 0.182	(/ms)		: open (v)		
	Rb   = 0.124	(/ms)		: close (v)		

	thi1  = -50	(mV)		: v 1/2 for inact 	
	thi2  = -75	(mV)		: v 1/2 for inact 	
	qi   = 5	(mV)	        : inact tau slope
	thinf  = -65	(mV)		: inact inf slope	
	qinf  = 6.2	(mV)		: inact inf slope
	Rg   = 0.0091	(/ms)		: inact (v)	
	Rd   = 0.024	(/ms)		: inact recov (v) 

	temp = 23	(degC)		: original temp 
	q10  = 2.3			: temperature sensitivity

:	dt		(ms)
	vmin = -120	(mV)
	vmax = 100	(mV)
}

ASSIGNED {
	v 		(mV)
	celsius		(degC)
	ina 		(mA/cm2)
	gna		(pS/um2)
	ena		(mV)
	minf 		hinf
	mtau (ms)	htau (ms)
	tadj
}
 
STATE { m h }

INITIAL {
    tadj = q10^((celsius - temp)/(10 (degC))) : make all threads calculate tadj at initialization

	trates(v+vshift)
	m = minf
	h = hinf
}

BREAKPOINT {
        SOLVE states METHOD cnexp
        gna = tadj*gbar*m*m*m*h
	ina = (1e-4) * gna * (v - ena)
} 

: LOCAL mexp, hexp 

DERIVATIVE states {   :Computes state variables m, h, and n 
        trates(v+vshift)      :             at the current v and dt.
        m' =  (minf-m)/mtau
        h' =  (hinf-h)/htau
}

PROCEDURE trates(v (mV)) {  
    TABLE minf,  hinf, mtau, htau
    DEPEND celsius, temp, Ra, Rb, Rd, Rg, tha, thi1, thi2, qa, qi, qinf
    FROM vmin TO vmax WITH 199

	rates(v): not consistently executed from here if usetable == 1

:        tinc = -dt * tadj

:        mexp = 1 - exp(tinc/mtau)
:        hexp = 1 - exp(tinc/htau)
}


: efun() is a better approx than trap0 in vicinity of singularity--

UNITSOFF
PROCEDURE rates(vm (mV)) {  
    LOCAL  a, b

:    a = trap0(vm,tha,Ra,qa)
    a = Ra * qa * efun((tha - vm)/qa)

:   b = trap0(-vm,-tha,Rb,qa)
    b = Rb * qa * efun((vm - tha)/qa)

    tadj = q10^((celsius - temp)/10)

	mtau = 1/tadj/(a+b)
	minf = a/(a+b)

    :"h" inactivation 

:    a = trap0(vm,thi1,Rd,qi)
    a = Rd * qi * efun((thi1 - vm)/qi)

:    b = trap0(-vm,-thi2,Rg,qi)
    b = Rg * qi * efun((vm - thi2)/qi)

    htau = 1/tadj/(a+b)
    hinf = 1/(1+exp((vm-thinf)/qinf))
}
UNITSON

COMMENT
FUNCTION trap0(v,th,a,q) {
	if (fabs((v-th)/q) > 1e-6) {
	        trap0 = a * (v - th) / (1 - exp(-(v - th)/q))
	} else {
	        trap0 = a * q
 	}
}	
ENDCOMMENT

FUNCTION efun(z) {
	if (fabs(z) < 1e-6) {
		efun = 1 - z/2
	}else{
		efun = z/(exp(z) - 1)
	}
}

Loading data, please wait...