1

Spiking neuron models of the basal ganglia: dopaminergic modulation of selection and oscillatory properties

Kevin Gurney, Mark Humphries, Rob Stewart

Adaptive Behaviour Research Group

University of Sheffield, UK

2

Rationale: basal ganglia and action selection

Aim: to understand underlying function of basal ganglia.

While learning is crucial – what is being learned?

Hypothesis: Main computational role of basal ganglia is to perform action selection

Supported by high (systems) level model Simple leaky integrators to represent population

dynamics

BUT…..

3

Beyond the systems level

Do more realistic models support the selection hypothesis? Constraints provided by:

Specific neuronal properties Physiological phenomena displayed by BG in toto…. If the price of a model performing selection is its failure exhibit these

phenomena, the selection hypothesis is in question

In particular, can models display oscillatory phenomena in BG?

If so, then we can use the model to explore possible function of these oscillations

Function or artifact?!

4

Systems level – the model architecture

cf, Hazrati and Parent, 1992, Mink and Thach 1993, Nambu et al 2000, Sato et al 2000

• Assumes relatively diffuse projection from STN• Emphasises STN’s role as input nucleus

striatum STN

Cortex (‘salience’ input)

output nuclei

- + Striatum

input

STN

output

Diffuseprojection

3 ‘channels’

5

New functional architecture:selection and control pathways

Interpret GP efferents as control signals for modulating selection pathway

Gurney et al, 2001

Selection pathway Control pathway

Striatum (D2)Striatum (D1) STN

EP/SNr GP

Cortex/thalamus

Diffuseprojection

6

Oscillations in basal ganglia: matching mechanisms to phenomena

Basal ganglia display a wide range of oscillatory phenomena – from <1Hz to >100Hz

These are probably associated with a correspondingly wide range of underlying mechanisms

We focus on four BG features. Intrinsic nature of STN-GP coupling Dopaminergic modulation of this coupling Rebound bursting in STN Synaptic patterning

7

Constructing the model

8

Rebound bursting in STN

Time

Current

IK

Beurrier et al 1999

ILIT

IK > IL burst ends

inhibition

9

Importance of synaptic patterning Inhibition at soma or proximal dendrites acts divisively (rather than ‘subtractively’)70% of GPe input is proximal or somatic (Bevan et al 1997)

cortex STN

GP

Proximal dendrites somaDistal dendrites

Captured phenomenlogically: use inhibition in proximal dendrites/soma to explicitly ‘gate’ more distal input

10

Dopaminergic action in striatum

Glu

Ctx

Striatum

Glu

Ctx

StriatumDAD1

Increased PSP

DAD2

decreased PSP

W = W0(1 + λ) W = W0(1 - λ)

λ < 1

11

GABA

GP

STN

Dopaminergic action in STN

Glu

Ctx

STN

DAD2 DA

D2

W = W0(1 – k1 λ) W = W0(1 – k2 λ)

K1, k2 < 1

Similar story in GP…

12

Dopamine: hypotheses

Low levels of dopamine serve to couple STN and GP more tightly and to make STN more sensitive to its input

Dopamine in striatum will make channel selection easier to achieve

13

Model neurons: summary Leaky Integrate and Fire with

AMPA NMDA, GABA, synaptic currents Shunting inhibition at proximal dendrites and soma Spontaneous currents Rebound bursting in STN, Dopamine in striatum, STN and GP. Inter-neuronal delays

All of the above parametrised by best estimates from the literature

14

Network Based on systems level model

3 discrete channels

64 neurons per channel, per nucleus

Probabilistic connection scheme within channels (only 25% of all possible connections made)

15

Constraining phenomena 1: Low frequency oscillations in STN-GP

(Magill et al, Neuroscience,106, 2001)

Low frequency oscillations (LFOs) in STN are driven by cortical slow wave under urethane anaesthesia.

GP does not oscillate in control (normal DA) conditions. Only shows oscillation under dopamine depletion (6-OHDA lesion)

Residual LFOs (with 6-OHDA lesion) in STN and GP under cortical ablation

16

Data – STN control

17

20Hz

1s

STN unitactvity

Model - STN control

Multi-taper spectrum

Frequency (Hz)

0 1 2 3 4 5

Pow

er

0

100

Time(s)

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Bin

cou

nt

0

10

20

Spike Trig Wav av

Time(s)

-1.0 -0.5 0.0 0.5 1.0

mea

n fir

ing

rate

(H

z)

0

10

20

30

Pseudo-eeg

18

Data – GP control

19

20Hz

1s

GP unitactvity

Model - GP control (1)

Spike Trig Wav av

Time(s)

-1.0 -0.5 0.0 0.5 1.0m

ean

firin

g ra

te (

Hz)

0

10

20

Time(s)

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Bin

cou

nt

0

50

Multi-taper spectrum

Frequency (Hz)

0 10 20 30 40 50

Pow

er

0

25

50

20

20Hz

1s

GP unitactvity

Model GP control (2)

Multi-taper spectrum

Frequency (Hz)

0 25 50

Pow

er

0

100

Time(s)

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Bin

cou

nt

0

50

Spike Trig Wav av

Time(s)

-1.0 -0.5 0.0 0.5 1.0

mea

n fir

ing

rate

(H

z)

0

10

20

21

Data – STN DA-depleted

22

20Hz

1s

STN unitactvity

Model STN DA-depleted

Multi-taper spectrum

Frequency (Hz)

0 1 2 3 4 5

Po

wer

0

1000

Auto corr.

Time(s)

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Bin

cou

nt

0

250

Spike Trig Wav av

Time(s)

-1.0 -0.5 0.0 0.5 1.0

mea

n fir

ing

rate

(H

z)

0

10

20

30

23

Data – GP DA-depleted (in phase)

24

20Hz

1s

GP unitactvity

Model - GP DA-depleted (in-phase)

Multi-taper spectrum

Frequency (Hz)

0 1 2 3 4 5

Pow

er

0

100

200

2D Graph 7

Time(s)

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Bin

cou

nt

0

50Spike Trig Wav av

Time(s)

-1.0 -0.5 0.0 0.5 1.0m

ean

firin

g ra

te (

Hz)

0

10

20

25

Data – GP DA-depleted (anti-phase)

26

20Hz

1s

GP unitactvity

Model - GP DA-depleted (anti-phase)

Multi-taper spectrum

Frequency (Hz)

0 1 2 3 4 5

Pow

er

0

500

Time(s)

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Bin

cou

nt

0

50

Spike Trig Wav av

Time(s)

-1.0 -0.5 0.0 0.5 1.0m

ean

firin

g ra

te (

Hz)

0

10

20

27

Data – cortical ablation and DA-depleted

Most neurons do not show LFOs but residual LFO activity…

28

Model - no cortex (DA-depleted)

Time(s)

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Bin

cou

nt

0

50

100

Multi-taper spectrum

Frequency (Hz)

0 1 2 3 4 5

Pow

er

0

100

Multi-taper spectrum

Frequency (Hz)

0 25 50

Pow

er

0

Time(s)

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Bin

cou

nt

0

10

20

1s

1s

STN

GP

Multi-taper spectrum

Frequency (Hz)

0 1 2 3 4 5

Pow

er

0

25

50

29

LFO countsS

TN +

ctx

GP

+ct

x

STN

-ctx

GP

-ctx

STN

+ct

x

GP

+ct

x

STN

-ctx

GP

-ctx

LFO

s (%

)

0

25

50

75

100

Control Data

Control Model

DA dep. data

DA dep. model

In DA control conditions, no GP LFOs, STN driven by cortex

LFO in GP promoted by DA depletion

Residual LFO in STN & GP under cortical ablation

Neuron is LFO if significant peak in power spectrum below 1.5Hz

30

Mean firing rates

STN

+ct

x

GP

+ct

x

STN

-ctx

GP

-ctx

STN

+ct

x

GP

+ct

x

STN

-ctx

GP

-ctx

Mea

n fir

ing

rate

s (H

z)

0

10

20

Control Data

Control Model

DA dep. data

DA dep. model

31

LFO – mechanistic explanation Low frequency oscillations associated with

rebound bursting will be ‘unmasked’ at low levels of dopamine….

GP more likely to generate pre-conditioning hyperpolarisation

32

Constraining phenomena 2: gamma oscillations in STN

(Brown et al., Exp Neuro. 177, 2002)

There is gamma oscillation (40-80Hz) in alert rats

This is increased (86% mean) by systemic D2 agonist (quinpirole)

Local field potential spectrum (control)

33

Frequency

0 20 40 60 80 100

Pow

er

0

20

40

60

Frequency

0 20 40 60 80 100

Pow

er

0

20

40

60

Model simulated D2 agonist

Frequency (Hz)

0 20 40 60 80 100

No.

of

sini

fican

t pe

aks

0

20

40

60ControlDA agonist

DA=0.2control

DA=0.8‘D2 agonist’

128% power

increase →

Mean power spectra (192 neurons)

Peaks in power spectrum

34

Gamma oscillations: explanation Gamma oscillations are associated with the

natural frequency of oscillation of the GP-STN circuit

determined by circuit delays

At control levels of dopamine, the presence of some LFO masks gamma

Can’t be doing gamma during quiet phase of LFO period.

At higher levels of dopamine, gamma is unmasked

35

Selection experiments

Cortical input(Mean firing rate)

time

ch1

ch2

1 2.5

36

Selection and switching

Ctx Ch1: 20 HzCtx Ch2: 40 Hz

Time

Mea

n fir

ing

rate

SN

r

ch2ch1 ch3

ctx

time

ch1

ch2

1 2.5

Firing rate

37

DA depletion prevents selection

Ctx Ch1: 12 HzCtx Ch2: 20 Hz

Time

Mea

n fir

ing

rate

SN

r

Firing rate

ctx

time

ch1

ch2

1 2.5

ch2ch1 ch3

LFOs?

38

Effects of DA depletion overcome by highly salient action

Time

Mea

n fir

ing

rate

SN

r

Ctx Ch1: 20 HzCtx Ch2: 40 Hz

Firing rate

ctx

time

ch1

ch2

1 2.5

ch2ch1 ch3

39

DA increase results in simultaneous selection

Ctx Ch1: 20 HzCtx Ch2: 40 Hz

Time

Mea

n fir

ing

rate

SN

r

Firing rate

ctx

time

ch1

ch2

1 2.5

ch2ch1 ch3

40

Summary A spiking model of BG constrained by known

physiology is able to account for a range oscillatory phenomena

Oscillations are modulated under Dopaminergic control of STN and GP

The same model displays selection and switching properties, thereby supporting the selection hypothesis for BG function

Currently exploring computational role of LFOs Perturb BG to selection in otherwise unresolved selection

competition?

41

The adaptive behaviour research group

Peter RedgravePaul Overton

Kevin GurneyTony Prescott

Mark HumphriesBen Mitchinson

Rob StewartRic Wood

Jonathan Chambers

Tom Stafford Ψ