Distributed control strategies for conductance regulation in dendrites

Timothy OLeary and David JA Wyllie

Introduction
Neurones are known to regulate their intrinsic properties homeostatically (Davis, 2006). However, the details of the underlying mechanisms are poorly understood. We developed a simple model of how neurones might regulate their conductance densities and explored the consequences of distributed, autonomous control processes that act locally in the dendrites of a model cell with realistic morphology. A canonical control system (right) is the most general model of homeostasis. The process being controlled (the plant in engineering terminology) transforms inputs into outputs and is typically subjected to external perturbations. The control system drives the plants output toward a target output profile by delivering a control signal to the plants inputs. Two generic control paths exist: feedforward (uff) and feedback (ufb); homeostatic processes consist of negative feedback controls. A subset of the plant outputs is accessible to the feedback control process; in the neuronal context this could be the resting membrane potential, spiking rate or intracellular calcium concentration. These outputs are compared with a target to generate an error signal from which the control signal is derived.
external disturbance

Centre for Integrative Physiology, University of Edinburgh, UK

Results: emergent conductance distributions

The model was initialized with uniform conductances and allowed to reach steady-state (simulation time = 103 s). This resulted in a change in intrinsic properties such as input resistance and spiking behaviour (right) and some heterogeneity in the conductance profile of all three conductances (delayed-rectifier potassium conductance and passive leak distributions are shown below). Following excitatory synaptic input to the apical dendrites, the intrinsic properties adapted, making the cell less excitable. In addition, greater heterogeneity emerged in the distribution of the conductances. This heterogeneity was not confined to the dendrites that received synaptic input, as a distinct profile was observed in the basal dendrites.
120 100 80 60 40 20 0 -1 0 1 2 3

Rin

start steady-state post-stim

[M]

feedforward control uff

++ +
ufb
feedback control error

plant

[log10 Hz]

output

start [mV]
0 -60

steady-state

target

+
gpas
start post-stim steady-state 100 m

250 pA, 100 ms

gKdr

Biologically plausible models of control systems that underlie neuronal homeostasis require sensor mechanisms that are known to exist in the cell. Intracellular calcium concentration is often favoured because many calcium-dependent signalling pathways are present in neurones, and calcium buffers are known to interfere with homeostatic compensation in a variety of experimental settings (Turrigiano et al., 1994; Berridge et al., 2003; Davis, 2006). Error signals can be transformed by the control process in a variety of ways. Some of the possibilities are illustrated in the table (right). Homeostasis requires a negative feedback control; three common examples are proportional control (which delivers a signal proportional to the error signal), integral control (signal is proportional to the accumulated, or integrated error) and bang-bang control (which generates a fixed on or off signal once the error crosses a target threshold). All three mechanisms are biologically plausible, however integral control is favoured because it eliminates steady-state error. In addition, integral control can be implemented with a straightforward model of the dynamics of conductance regulation (outlined below). Bang-bang control is an idealized model of processes which exhibit threshold behaviour, and might plausibly describe processes such ion channel expression. We focussed on integral control, but explored bang-bang control in the context of noise.
m ho

profile path

time external disturbance

t ic sta eo ?

Controller

Characteristics

target No control output

Changes are due to external disturbances only.

Feedforward

Reaches target state if tuned, does not compensate for external disturbances. Moves toward target state but never reaches it. Partially compensates for external disturbances. Reaches target state with a lag and/or overshoot. Compensates for external disturbances Fast response but oscillates about target state. Compensates for external disturbances.

Proportional

Which controller?
The error signal and the control mechanism itself exist in a noisy environment. This is especially true in dendrites, which are subjected to a large amount of electrical noise and where noise due to the stochastic nature of biochemical signalling is substantial (Cannon et al., 2010; Kotaleski and Blackwell, 2010). We investigated the performance of an integral controller with the same parameters as that used in our neuronal model in the presence of noise (Ornstein-Uhlenbeck process, cut-off frequency = 100 Hz). Interestingly, the integral controllers performance compares poorly with a much cruder bang-bang controller under conditions of high noise (below) and exhibits far stronger frequency dependence (below right).
10
0

Integral

time external disturbance

Bang-bang

We used a multi-compartment model of a reconstructed CA1 pyramidal cell (Migliore et al., 2005) implemented in NEURON (Hines and Carnevale, 1997). From this model, we kept the passive leak and voltage-gated sodium and (delayed-rectifying) potassium conductances and introduced simplified calcium dynamics. Membrane conductances were modified to model activity-dependent homeostasis by introducing first-order calcium-dependence in the maximal conductance with a slow timeconstant (100 s). The details of the model are summarized below. Standard Hodgkin-Huxley style model:
dV Cm = g pas (V E pas ) + g Na (V ) + g Kdr (V ) dt

integral

The model

bang-bang

no noise

noise present
SNR = 2 (~ 3 dB)

10

r ( x ) = g r 10kr x ,

6 r ( x ) 2

Mean-squared error

10

-5

Mean-squared error

Controller transformation:

Simplified calcium dynamics:

d [Ca]i Ca = kCa (V Veq ) + C [Ca]i dt

Model parameters (as in Migliore et al., 2005) except:

g Na = 10 2 ; g Kdr = 10 2; g pas = 10 5 kNa = kKdr = k pas = 0.1 [s] [nS cm2 ]

10

-10

10

-10

Conductance regulation:
dg r = r ([Ca]i Cr ) (implicit integral control) dt

Ca = 100 ms mM1

C = 10 4 mM Veq = 65 mV E pas = 60 mV
10

10
-15

-20

10

-2

10

-1

10

10

10

-5

10

-3

10

-1

10

Random excitatory synaptic input was introduced to the apical dendrites at a mean (Poisson) rate of 10 Hz and with a lognormal distribution in conductance (mean = SD = 100 nS, Erev = 0 mV). The figure (below) shows a sample of activity, with the dendrites of the model cell pseudo-coloured according to membrane potential

Noise/signal ratio

Target frequency [Hz]

Conclusions and further work

We draw two main conclusions from this work: (1) simple, local homeostatic rules can generate heterogeneous conductance distributions in dendrites; (2) noise has a substantial impact on the performance of plausible homeostatic controllers. The current NEURON model has several limitations: only three conductances are included, calcium dynamics are simplified and parameters are unconstrained. Further work will combine distributed homeostatic regulation with realistic models of noise and neuronal morphology, however, this will require substantial computational resources.

40

[mV]

-60

Vm
10 ms

References
Berridge MJ, Bootman MD & Roderick HL. (2003) Nat Rev Mol Cell Biol 4, 517-529. Cannon RC, O'Donnell C & Nolan MF. (2010) Plos Comp Biol. 6. Davis GW. (2006) Annu Rev Neurosci 29, 307-323. Hines M & Carnevale T. (1997) Neural Computation 9, 11791209 Kotaleski JH, Blackwell KT (2010) Nat Rev Neurosci. 11:239. Migliore M, Ferrante M & Ascoli G. (2005) J. Neurophysiol. 94:6. Turrigiano G, Abbott LF & Marder E. (1994) Science. 264, 974-977.

[Ca]i C