CNS, JC talks

Activity-dependent myelination: A glial

mechanism of oscillatory self-
organization in large-scale brain
networks
Rabiya Noori, Daniel Park, John D. Griffiths, Sonya Bells, Paul W.
Frankland, Donald Mabbott and Jeremie Lefebvre
PNAS June 16, 2020 117 (24) 13227-13237
https://doi.org/10.1073/pnas.1916646117

Presented by Gustavo Patow, ViRVIG–UdG & CNS

JC@CNS - 11/05/2021
Very short overview
• Shifts the focus from neural to glial contribution to
brain synchronization
• Adaptive, activity-dependent changes in conduction
velocity
• Such plasticity endows white matter with self-
organizing properties
• This mechanism stabilizes oscillatory neural activity
across a wide range of connectivity gain and
frequency bands, making phase-locked states more
resilient to damage JC@CNS - 11/05/2021 2
So, why this paper?
• Fascinating paper!
• This mechanism must play a very important role in disease and
lesion
• Stroke
• Alzheimer (?)
• Should be taken into account in any long-term implementation

• Warning: No validation at all!

JC@CNS - 11/05/2021 3
Overview: activity-dependent plastic
myelination

JC@CNS - 11/05/2021 4
Quick note on methods
• Network of Kuramoto oscillators
• Kuramoto order parameter (r) to measure syncronicity

JC@CNS - 11/05/2021 5
 Connecvity weights ωij (au)
FN

Macaque Parcellation
• CoCoMac (Kötter and Wanke)
• 96 regions = 80 cortical + 16
subcortical
• Mapped from macaque to lij (mm)
Axonal tract lengths
human neuroanatomy using a
regional map (RM) parcellation
scheme
• Taken from TVB…

JC@CNS - 11/05/2021 6
Spontaneous transition to phase-locked
synchrony
A- The network was
initially set with
• Uniform conduction
velocity
• Randomized phases
B- After 10h, the
conduction velocity
changed and adjusted so
that phase-locked
synchrony emerges in the
system
7
Methods: The model
Network of N coupled nonlinear and delayed Kuramoto oscillators

Where Wij is the synaptic connectivity matrix

gain
JC@CNS - 11/05/2021
Binarized! ! 8
Kuramoto order parameter
• Values are between 0 (fully incoherent phases) and 1 (full
synchronization)
• fluctuates in time and scales with the degree of phase locking

• Also, the mean phase across the network

JC@CNS - 11/05/2021 9
Net effective conduction delay over a
given axonal connection

total axonal tract length of the axonal path A between

populations i and j

JC@CNS - 11/05/2021 10
Myelination rule
Phenomenological phase-dependent myelination rule that
governs the evolution of the conduction velocity cij
Activity-dependent myelination term

Inertial Drift

@Code:
M ensures that conduction velocity remains
Cij[Cij<min] = 3
positive and bounded between 3 and 100 m/s
Cij[Cij>max] = 100
JC@CNS - 11/05/2021 11
Inertial Drift

net effect of a combination of metabolic and/or

biophysical forces preventing and/or slowing myelin
formation

JC@CNS - 11/05/2021 12
Activity-dependent myelination term

Phase difference

JC@CNS - 11/05/2021 13
Spontaneous transition to
phase-locked synchrony
C- Evolution of randomly selected
plastic conduction velocities
• Increase and a plateau once phase-
locking is achieved
D- Associated time delays as a function
of time
• Showing how fast axonal connections
formed over the duration of the
simulations
JC@CNS - 11/05/2021 14
Distribution of conduction velocities after
myelination

Negative correlation:
longer connections are
statistically slower

15
Distribution of conduction delays after
myelination
Two main components:
• Delays that are linearly
related to tract length
(i.e., constant velocity)
• Delays that result from
myelination
high degeneracy:
≠ lengths and velocities
→ similar delays

16
Order parameter as a function of the synaptic
gain (g) of connections and frequency bands
(δ,α ,β, γ) for increasing the myelination rate (ϵ)

JC@CNS - 11/05/2021 17
Damage
• We considered γ as a disconnection probability and modified
elements of the matrix W according to the following rule

• p = [0,1] is a uniform random deviate

• Collectively, the application of this rule sets a total of γN
connection weights to zero
JC@CNS - 11/05/2021 18
Network resilience to sudden and diffuse
decrease in connectivity
• Random and spatially
diffuse injury scaled by
the insult index γ
• 5 h of network
evolution
• Compared the
synchronization before
and after injury, after
network conduction
velocities have adapted
JC@CNS - 11/05/2021 19
Network resilience to sudden and diffuse
decrease in connectivity
• Injury of 80% of the
connections
• Adaptive myelination
allows the system to
maintain a high degree of
phase-locked synchrony
• High r → oscillatory
resilience

JC@CNS - 11/05/2021 20
Network resilience to sudden and diffuse
decrease in connectivity
• The dynamics of the
conduction velocities for
preserved connections are
altered
• A change in overall
conduction velocity
statistics takes place

JC@CNS - 11/05/2021 21
Network resilience to sudden and diffuse
decrease in connectivity
• The myelination rate
compensates for an
increased degree of injury
to preserve oscillatory
synchrony
• The more significant the
connectivity insult, the
more myelination is
required to keep the order
parameter high
• and hence, keep the
remaining networks
synchronized
JC@CNS - 11/05/2021 22
Bidirectional Myelin Remodeling
Retraction and stabilization
establish and maintain network-
wide phase synchronization:
• If the phase offset between two
connected oscillators is positive
(red), the phase precession
causes a retraction of myelin,
and hence the conduction
velocity decreases
• If the phase offset between two
connected oscillators is negative
(blue), the phase lag causes a
stabilization of myelin, and
hence the conduction velocity
increases

JC@CNS - 11/05/2021 23
Bidirectional Myelin Remodeling
• Activity-dependent myelination rule to reflect bidirectional
phase-dependent changes (i.e., increase and decrease) in
conduction velocity that occur at different rates

Retraction/Stabilization ratio
JC@CNS - 11/05/2021 24
Bidirectional Myelin Remodeling
• Kuramoto order param (r) as a func of:
• The R/S ratio
• Increasing synaptic gain (g)
• Connections and various frequency bands (δ, α, β, γ)
• Measured across independent trials (10 @Code)
• For R/S = 0.5, there is more myelin being
formed than being lost overall
• Results in more phase locking across the
network W.R.T. R/S = 1.0
• For R/S = 1.0, less phase locking in higher-
frequency bands for all synaptic gains
• However, the phase locking still takes place
over a wide range of synaptic coupling
JC@CNS - 11/05/2021 25
About the code
• C files (.cpp files)
• Straightforward to read
• No parallelization (nested for loops, 1 thread)
• Needs a bit of refactoring… ;-)

JC@CNS - 11/05/2021 26
Implementation problem – acc Gus(P)
!
• WM plasticity occurs on a much slower timescale compared with
the dynamics of its oscillators
• WM Plasticity: hours
• Neural activity: milliseconds
• They do a “time rescaling” in which the time constant αc = 1 (see
code), to match the dynamics of the neural oscillators
• To validate this approach, they analyzed the asymptotic dynamics
of the myelination rule αc-1 → 0
• Assumed (observations) that individual phase differences Δij follow,

JC@CNS - 11/05/2021 27
Now… why not…?
!
• In my “previous life”, we also had very large timescale differences
• E.g., rain, weather and climate
• E.g., IA, animation and rendering
• Solution: to decouple the respective equations!

JC@CNS - 11/05/2021 28
Now… why not…?
!
• Proposal: Solve for myelin cij evolution disregarding oscillations
• At each step, compute the full neuronal dynamics as if myelin was static
• Advantages:
• dt for myelin: 1 to 10 minutes
• dt for neuronal dynamics: milliseconds
• No “strange” interactions!

JC@CNS - 11/05/2021 29
Limitations (I)
• The phenomenological myelination rule
considered, despite time rescaling, occurs at a
very fast pace compared with experimental
observations
• Indeed, the changes in conduction velocity in the
model occur much faster than the myelin
formation reported in animals, which stabilizes
after a few hours

Remember: NO validation!
JC@CNS - 11/05/2021
! 30
Limitations (II)
• The myelination model remains highly
simplified
• The different physiological constraints that
dictate activity-dependent myelination in
the brain are diverse, spatially distributed,
and correlated
• The physiological mechanisms involved in
activity-dependent myelination in the brain
remain poorly understood

Remember: myelination
! rule only depends on Δij!
JC@CNS - 11/05/2021 31
Limitations (III)
• The use of phase oscillators to model local neural dynamics,
while well supported in the literature, remains limited
• The local synchronization features, and how they are impacted by
afferent inputs, are likely to impact the myelination process and hence,
greatly enhance the richness of the dynamics, as well as the complexity
of the analysis

Remember:
! Kuramoto!

JC@CNS - 11/05/2021 32
Extra!

• Analyzed the equilibria and stability of this system, and applied

the results to two-oscillator and large-dimensional networks
• The mathematical and numerical analysis demonstrates that
plastic delays act as a stabilizing mechanism promoting the
network’s ability to maintain synchronous activity
• Also show that global synchronization is more resilient to
perturbations and injury towards network architecture

JC@CNS - 11/05/2021 33
Extra! Take home message
[Lefebvre2021, personal communication]
• State-dependent delays reinforce the synchronization
properties of networks of oscillators, notably by creating
multi-stability
• i.e., multiple synchronization frequencies - as opposed to a
single one in the fixed delay case
• In simple terms: creating degenerate solutions to the
synchronization problem

JC@CNS - 11/05/2021 34
CNS, JC talks

Activity-dependent myelination: A glial

mechanism of oscillatory self-
organization in large-scale brain
networks
Rabiya Noori, Daniel Park, John D. Griffiths, Sonya Bells, Paul W.
Frankland, Donald Mabbott and Jeremie Lefebvre
PNAS June 16, 2020 117 (24) 13227-13237
https://doi.org/10.1073/pnas.1916646117

Thanks!!!
JC@CNS - 11/05/2021
Model Parameters

JC@CNS - 11/05/2021 36