User manual | Paired spiking robustly shapes spontaneous in vitro

Paired spiking robustly shapes spontaneous
activity in neural networks in vitro
Aurel Vasile Martiniuc1*, Victor Bocoş-Binţinţan2, Rouhollah Habibey3,
Asiyeh Golabchi3, Alois Knoll1, Axel Blau3
1
Computer Science Department VI, Technical University Munich, Boltzmannstraße 3,
85748 Garching, Germany, http://www.in.tum.de/
2
Dept. of Environmental Science & Engineering, Babeş-Bolyai University, 30 Fantanele Street,
400294 Cluj-Napoca, Romania, http://enviro.ubbcluj.ro/
3
Dept. of Neuroscience and Brain Technologies, Italian Institute of Technology, via Morego 30,
16163 Genoa, Italy, http://www.iit.it
*Correspondence: [email protected]
Keywords: Cultured neurons, Paired spiking activity, Long-term recording,
Bursting activity, Neural information content per spike.
Abstract
In vivo, neurons establish functional connections and preserve information along their
synaptic pathways from one information processing stage to the next in a very
efficient manner. Paired spiking (PS) enhancement plays a key role by acting as a
temporal filter that deletes less informative spikes. We analyzed the spontaneous
neural activity evolution in a hippocampal and a cortical network over several weeks
exploring whether the same PS coding mechanism appears in neuronal cultures as
well. We show that self-organized neural in vitro networks not only develop
characteristic bursting activity, but feature robust in vivo-like PS activity. PS activity
formed spatiotemporal patterns that started at early days in vitro (DIVs) and lasted
until the end of the recording sessions. Initially random-like and sparse PS patterns
became robust after three weeks in vitro (WIVs). They were characterized by a high
number of occurrences and short inter-paired spike intervals (IPSIs). Spatially, the
degree of complexity increased by recruiting new neighboring sites in PS as a culture
matured. Moreover, PS activity participated in establishing functional connectivity
between different sites within the developing network. Employing transfer entropy
(TE) as an information transfer measure, we show that PS activity is robustly involved
in establishing effective connectivities. Spiking activity at both individual sites and
network level robustly followed each PS within a short time interval. PS may thus be
considered a spiking predictor. These findings suggest that PS activity is preserved in
spontaneously active in vitro networks as part of a robust coding mechanism as
encountered in vivo. We suggest that, presumably in lack of any external sensory
stimuli, PS may act as an internal surrogate stimulus to drive neural activity at
different developmental stages.
1
Martiniuc et al.
1.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
Introduction
Both, in vivo and vitro, stimuli may trigger
bursting (Krahe 2004; Akerberg 2011) and PS
activity. The early visual system is a prominent
example. PS activity in retinal ganglion cells is
driving suprathreshold responses at postsynaptic
targets in the lateral geniculate nucleus (Usrey
et al., 1998; Sincich et al., 2007; Weyand,
2007). PS enhancement contributes to
preserving the information of a visual stimulus
from one processing stage to the next (Rathbun
et al., 2010; Sincich et al., 2009; Uglesich et al.,
2009). It has been shown that the second spike
in a pair evoked a postsynaptic potential with
maximum efficacy for inter-spike intervals
(ISIs) in the range of 2 - 5 ms. Efficacy rapidly
decreased to zero for ISIs larger than 40 ms
(Usrey et al., 1998; Sincich et al., 2007).
Both in vivo and in vitro, synchronous
correlated activity known as bursting is one of
the information processing mechanisms that
shape network interconnectivity, both at single
cell and network level (van Pelt et al., 2004;
Wagenaar et al., 2005; McCabe et al., 2006;
Wagenaar et al., 2006; Rolston et al., 2007;
Mazzoni et al., 2007; Sun et al., 2010).
Bursting not only occurs in brain slices with
partially intact interconnectivity (Blankenship
and Feller 2010; Rolston et al., 2007), but is
also found in neural cultures derived from
dissociated brain tissue where it becomes
predominant as cultures mature (Wagenaar et
al., 2005; Wagenaar et al., 2006). Bursting
activity varies with culture age (Nadasdy 2000;
van Pelt et al., 2004), and other factors, i.e.
culture density (Wagenaar et al., 2006).
However, little is known on the evolution and
role of PS activity in neural cultures derived
from dissociated brain tissue, on its relationship
to bursting activity and on its participation in
the organization of functional and effective
network connectivity. To address these
questions, we defined activity consisting of two
spikes being separated by an interval of up to
5 ms followed by an inter-paired-spike interval
(IPSI) larger than 40 ms as PS activity (Methods
2.3). We then analyzed 58 streams of
continuously
extracellularly
recorded
spontaneous neural activity in random networks
for PS occurrence and for the spatio-temporal
evolution of PS activity patterns over several
weeks. In this context, we wondered whether
any PS-induced effect was locally confined or
led to changes on network level. We finally
investigated the robustness of PS activity and its
independence in driving spontaneous neural
activity, thereby affecting functional and
effective connectivity.
Different spatio-temporally recurring patterns
occur in both stimulus-induced (Ferrández et
al., 2013) and spontaneous activity. They are
usually dynamic over time (i.e. the spatial
location of active sites may change), thereby
having different, yet characteristic spatiotemporal shapes (Shahaf and Marom, 2001; van
Pelt et al., 2005; Sun et al., 2010; DeMarse et
al., 2001; Pasquale et al., 2008; Pasquale et al.,
2010; Ruaro et al., 2005; Nadasdy 2000;
Nomura et al., 2009).
Without any external stimulus, cultured neurons
show significant changes in their spontaneous
neural activity at different stages toward
maturity.
Moreover,
network
activity
fluctuations at later stages may be a
consequence of repetitive internal stimuli that
revive prior network activity and are thought to
alter network connectivity to compensate for the
lack of external stimuli. Such self-organized
events based on spontaneous neural activity
were previously reported at different culture
ages (Rolston et al., 2007; Pasquale et al., 2010;
Sun et al., 2010).
2
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
2.
Material and methods
2.1
Continuous 59-channel MEA
electrophysiology and spike train
assembly
2.2
Local and network firing and burst
rates
Firstly, we quantified the local firing rates
(LFRs) at individual sites as the number of
recorded spikes divided by TTrial for each local
spike train (LST) (Suppl. Fig. 1(B)). At network
level, we pooled all spikes from all 57 and 58
sites, respectively, for each trial into a single
network spike train (NST) by sorting them in a
time-ascending order. The NST represented the
MEA-wide activity for each trial. The network
firing rate (NFR) was then quantified as the
total number of spikes in an NST divided by
TTrial (Suppl. Fig. 1(C)).
The data for this analysis was provided by a
recently developed cell culture perfusion system
that allowed us to continuously track both
network activity and morphology on the lab
bench at ambient CO2 levels under rather well
controlled environmental conditions (i.e. pH,
temperature and osmolality). Technological and
procedural details will appear in a dedicated
article (Saalfrank et al., submitted). With this
setup, the activity evolution in a hippocampal
and a cortical network on MEAs was
continuously recorded over 30 and 53 days in
vitro (DIV), respectively. These datasets were
analyzed for PS activity. To reduce data file
size, only upward (positive) and downward
(negative) spike cutouts from 57 (cortical) and
58 (hippocampal) out of 59 recording electrodes
were stored in 5 min packets. They consisted of
5 ms pre-spike and 5 ms post-spike fragments
after first threshold crossing at ± 5.5 SD with
respect to peak-to-peak noise (Suppl. Fig. 1(A)).
Only timestamps from downward thresholdcrossings were extracted using NeuroExplorer
(Nex
Technologies).
After
removing
simultaneous timestamps that occurred on all
channels due to electrical or handling artefacts,
subsequent 5 min datasets comprised of ≤ 58
timestamp streams were bundled in 12 hour
timestamp packets for further analysis in Matlab
(MathWorks). In the following, these half-day
packets will be called trials of duration TTrial ≤
12 h. On some days, trials encompassed less
than 12 h due to temporary interruptions for
system reconfiguration, maintenance work or
power failure. Our recording sessions consisted
of 65 trials (32.5 DIVs) for the hippocampal
culture and 106 trials (53 DIVs) for the cortical
culture.
To further investigate activity dynamics, we
used the burst rate (BR), a well-known
parameter for characterizing synchronous
network activity. We scanned all LSTs at the 57
(cortical) and 58 (hippocampal) individual sites
for each trial and defined bursting activity as
events with more than 10 subsequent spikes
being individually separated by an ISI of less
than 100 ms, followed by an interburst interval
(IBI) larger than 200 ms (Wagenaar et al., 2005)
(Suppl. Fig. 1(C)). The local burst rate (LBR) at
individual sites was calculated by dividing the
number of bursts by TTrial. Equally, the network
burst rate (NBR) was obtained by scanning the
NST for bursts using above mentioned criterion
and dividing the number of bursts by TTrial.
2.3
Paired spiking activity
Based on the previously described findings by
Ursey (Usrey et al., 1998) and Sincich (Sincich
et al., 2007), we investigated the occurrence and
effect of paired spiking on neural activity in
network cultures. As sketched in Suppl. Fig.
1(D), we defined PSs at individual sites as the
neural activity consisting of two spikes recorded
from the same electrode separated by an interval
of up to 5 ms, followed by an IPSI larger than
40 ms (in order to assure that a second spike in a
PS does not influence a first spike in a second
PS for two consecutive PSs).
3
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
Equally, we scanned all the NSTs using the
above mentioned algorithm to quantify the PS
activity at network level for each trial. In this
case, the two spikes separated by an interval of
up to 5 ms and followed by an IPSI larger than
40 ms did not necessarily have to be recorded
from the same electrode.
To generate above mentioned Poisson-like spike
trains, we used Matlab user-written routines.
The Poisson distribution P represents the
probability that a homogenous Poisson process
generates n spikes in a period of trial duration
TTrial:
P ( n) 
To describe the PS activity dynamics for each
trial, we calculated the number of active sites
with PSs (NASPS) as being the number of sites
with at least two PS repetitions during TTrial.
Equally, we calculated the NASB for bursts as
the number of sites with more than one burst per
trial.
(rTTrial ) n
exp(  rTTrial )
n!
Eq. 1a
where r is the spike count rate defined as the
total number of spikes divided by TTrial for each
LST and NST.
Timestamps were generated by the following
interspike interval formula:
To check if PS activity forms robust
spatiotemporal patterns, we firstly calculated the
IPSI as being the difference between two
consecutive PSs at both network level and
individual sites. From the IPSI histograms for
each trial we extracted the highest number of
IPSI repetitions and the most frequently
encountered IPSI value (IPSImfo) at individual
sites.
1
t i 1  t i   ln( rand )
r
Eq. 1b
where rand is a random number uniformly
distributed over the open interval (0 : 1); ti
represent the spike timestamps for i = 1, 2,..., n
spikes (Martiniuc and Knoll, 2012).
2.4
Moreover, to confirm that PS activity was
neither governed strictly by firing rates (and
thus represented an intrinsic neural response
property) nor, at network level, by chance as a
procedural result of projecting spiking activity
from individual sites onto a single NST
timeline, we generated Poisson-like network
spike trains for comparison. In these, the firing
probability was distributed according to a
homogenous Poisson process without refractory
period. If PS activity were strictly governed by
firing rate, Poisson-like spike trains with the
same firing rate as the recorded NSTs would
give a similar PS distribution. Additionally, we
shuffled the NSTs 100 times for each trial and
quantified PS activity to check the degree of
randomness of NST PS activity. NST spike
times were randomly rearranged with the
randperm (Matlab, MathWorks) function. 100
repetitions were chosen to warrant statistical
significance at acceptable computational costs.
Post-stimulus time histogram and timevarying firing rates at network level
To investigate the hypothesis that PS activity
might replace external stimuli sources, we
considered each PS a stimulus-resembling event
for the network. Thus, for this particular
investigation at network level, we considered PS
onset (first spike) as the beginning of a stimulus
(t = 0 s) with a duration of TPS = 2 seconds,
which is close to the shortest IPSI duration
found for each trial.
To calculate the post-stimulus time histograms
(PSTHs), the timestamps of PS-elicited spikes
during a TPS were aligned relative to t = 0 s for
each period TPS. n reflects the number of PS
stimuli at network level in a trial. nmin was
found to be 200. We divided the stimulus period
TPS into N bins of duration Δt = 5 ms and
counted the number of spikes ki from all n
sequences that fall into bin i. After averaging for
the n stimulus repetitions and dividing by bin
4
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
duration Δt, we obtained the time-varying firing
rates r(t) with respect to stimulus (PS) onset.
2.5
This measure of information content does not
make any assumption about the stimulus
features; it only reveals the information content
carried by individual spikes.
Information content per spike
In order to evaluate the information about the
stimulus (PS) carried by individual spikes
following a PS within TPS, we used the above
calculated time varying firing rates r(t) and
computed the entropy estimates (H) as follows
(Strong et al., 1998; Brenner et. al., 2000;
Sincich et al., 2009):
1
H
TPS
TPS
r (t )
  r  log
0
2
r (t )
dt
r
2.6
The highly variable spontaneous spiking activity
of cultured neurons features robust patterns (i.e.
bursting activity), which might participate in the
establishment of functional connections
between different sites within the culture. To
investigate whether PS activity plays a role in
shaping the dynamic interconnectivity map at
different developmental stages, we adopted a
variant of a cross-correlation algorithm initially
introduced as conditional firing probability
(CFP; le Feber et al., 2007). This method was
widely used in the investigation of activity
relationships between different electrodes to
reveal the formation and strength evolution of
functional connectivity within in vitro networks
(Zullo L et al., 2012; Chiappalone et al., 2007;
Garofalo et al., 2009). Here, we used CFP to
reveal any PS-related cross-correlations
between different sites within the network. That
is, at each electrode i (i = 1 : 57 or 58,
respectively) considered as the reference, we
selected the second spike in each PS as a
reference with new relative time ti = 0. We then
calculated the CFP as the probability of spike
occurrences at any of the other 56 or 57
recording electrodes j (j = 1: 56 or 57,
respectively) within the time interval TCFP
[ti : ti + 500 ms] divided by the total number of
second reference spikes of a PS at reference
electrode i over the entire trial duration of
TTrial = 12 h. The spikes found at electrode j
during TCFP were aligned relative to each ti and
binned with a bin size of t = 1 ms. If any of the
resulting 57*57or 58*58 CFP(i,j) distribution
curves showed a clear peak, we considered
electrode j being correlated to the PS activity on
reference electrode i. The peak amplitude was a
measure of correlation strength. Its timestamp
reflected the PS-related synchronization delay
between the two neurons.
Eq. 2
where TPS = 2 s represents the above mentioned
duration of a stimulus and <r> the average firing
rate; also in this case, the bin size was
t = 5 ms.
For each particular trial, we obtained the
average estimate for the information content per
spike by averaging the estimated entropy by the
number of stimulus repetitions n:
H  
1 n
 Hi
n i 1
Conditional firing probability
Eq. 3
To account for limited dataset size and to
correct the resulting bias, the information
content per spike was estimated as a function of
bin size t. We performed a linear fit to these
data to extract the intercept corresponding to the
limit when t approaches zero. We used the
shuffle verification method to check for the
robustness of information content per spike.
Briefly, we randomly rearranged each NST as
described in section 2.3 and computed estimates
of the information content per spike as
mentioned above (Eq. 2). For each of the
shuffled NSTs, we repeated this procedure100
times and obtained a standard deviation of
estimated information content that was smaller
than one standard deviation of the fitting
intercept obtained from above mentioned linear
fit.
5
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
Additionally, two boundary conditions were
chosen as restrictive validity criteria: a CFP(i,j)
was rejected if the width at 80% of the peak
value was shorter than 5 ms (five bin sizes; to
avoid false correlations caused by outliers) and
for synchronization delays larger than 250 ms
(to avoid curves that decreased to zero beyond
the 500 ms window).
2.7
al., 2011) derived from the original definition
given by Schreiber (Schreiber, 2000) as follows:
Eq. 4
A complete description of the TE toolbox
algorithm can be found in Ito et al., 2011.
Briefly, p describes a probability, At depicts
whether at time t a spike at unit A was recorded
(and thus At = 1) or not (At = 0). Similarly, Bt
and Bt+1 describe the status of unit B at times t
and t+1. Conditional probabilities of observing
the particular status of units A and B are marked
by vertical bars while the sum is over all
possible combinations Bt+1, Btk and Atl, where
parameters k and l express the number of time
bins in the past that allow us to take the time
delay and the message length into account when
calculating TE. For biophysical reasonability,
we chose k = 1:30 ms and l =1:250 ms.
Transfer entropy
Above mentioned concepts (i.e. information
content per spike and CFP) quantify statistical
dependences of observed variables (i.e. recorded
spike trains), thereby describing functional
connectivity maps which do not allow us to
investigate the direction of information flow
(i.e. the causality) between the recorded units.
By definition (Wienner, 1956), an effective
connectivity between two neurons exists when
knowledge about the past of one neuron predicts
the future activity of its counterpart better than
the prediction based on the past activity of the
receiver (neuron) alone. This effective
connectivity is quantified by an information–
theoretic measure called transfer entropy (TE)
that was introduced by Schreiber (Schreiber,
2000). TE is an asymmetric measure of
interactions between two coupled neurons
which reveals effective connectivities and
indicates the direction of information flow
between recorded units. It permits to predict the
spiking activity of a post-synaptic neuron by
taking past spiking activity of its pre-synaptic
partner into account. In our work, TE is positive
and thus the information is directed from a
sender unit A to a receiver unit B (i.e. there is an
effective connectivity from unit A to unit B)
only when the information about the spiking
activity recorded at unit A improves the
prediction of the spiking activity in the future of
unit B better than any prediction derived from
past spiking activity recorded at the unit B
alone. Thus, to identify and assess effective
connectivity within the neural network from
recorded spiking activity, we used a recently
introduced toolbox for calculating the TE (Ito et
In this general framework, we exclusively
considered PS activity at unit A (the sender)
while unit B (the receiver) encompassed the
entire recorded spiking activity. In this way, we
could estimate whether PS activity at unit A was
involved in information transfer toward unit B.
Furthermore, we exemplarily chose the eight
closest recording units as depicted in Figure 6A
to check if PS was involved in information
transfer and thus in establishing effective
connectivity between these eight closest
neighbors within the network. Thus, each of the
closest eight units was scanned for PS activity
and considered as the sender with respect to the
entire spiking activity of the remaining seven
closest neighbors. This resulted in a TE map,
which depicts the PS information transfer
dynamics of each of the selected senders (eight
units A) toward the selected receivers (seven
units B). For computational reasons, we split
each trial duration T into 30 minutes subsets of
recorded data.
We applied the same algorithm at network level
to investigate the effect of information transfer
6
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
from the above mentioned eight units A toward
the rest of the network. In this case, we
calculated the PS-related TE for each of the
eight selected channels with respect to the entire
NST as the only receiver unit B.
mean values in Figure 1A. This increasing
spontaneous spiking activity in maturing
cultures is in accordance with previously
reported results (van Pelt et al., 2005).
Significant changes have also been found at
individual sites where the number of active sites
during the same developmental periods
increased as the culture grew toward maturity.
In contrast, we found fluctuating neural activity
periods containing or terminating with high
network firing rates (Figure 1A, red bars) for the
more mature cortical culture. We could
distinguish six time periods marked by an
increasing period followed by a decreasing
trend. The individual periods lasted from
24 DIV to 37 DIV (C2-1) with a mean of 15±8
spikes/s, from 38 DIV to 45 DIV (C2-2) with a
mean of 25±11 spikes/s, from 46 DIV to
50 DIV (C2-3) with a mean of 8.5±7 spikes/s,
from 51 DIV to 65 DIV (C2-4) with a mean of
7±3.4 spikes/s, from 66 DIV to 71 DIV (C2-5)
with a mean of 6±2 spikes/s and from 72 DIV to
77 DIV with a mean of 5.5±1.8 spikes/s for the
last period (C2-6).
Further on, we calculated the differences
between the resulting TEs:
Eq. 5
When ΔTE is positive, the information transfer
is directed from A to B; in the opposite case, the
information flows from B to A.
3.
Results
3.1
Evolution of firing and burst rates
through different developmental stages
Taking advantage of the uninterrupted
extracellular recording technology for cultured
neurons based on 59-channel microelectrode
arrays (MEA), we analyzed the day to day
evolution of spontaneous neural activity at both
individual sites and network level. We
extracellularly recorded activity from two
different networks cultured under similar
conditions. Quasi-continuous datasets from
7 DIV (first extracellularly recorded spikes
emerged from the 5 µV noise floor and crossed
the -5.5 SD of the peak-to-peak noise spike
detection threshold) to 39 DIV for the
hippocampal culture and from 24 DIV to
77 DIV for the cortical culture were analyzed.
Firing rates are usually used to reveal
characteristic communication mechanisms that
are different for spontaneous and induced
activity, respectively. In contrast, bursting
activity plays a role in filtering spontaneous
neural activity (van Pelt et al., 2004; Wagenaar
et al., 2005). In our recordings, spiking activity
tended to induce bursts of synchronized activity
at different developmental stages.
Similarly to the network firing rate (NFR,
Figure 1A, blue trace), the hippocampal culture
showed a significantly (p < 0.05) increasing
trend in bursting activity at network level (see
Methods 2.2, Figure 1B, blue bars) over the
three periods. A mean of 0.018 bursts/s during
the first period increased to a mean of 0.22
bursts/s for the last period (Figure 1B, yellow
dots). This developmental trend has already
been reported in other network studies (van Pelt
et al., 2004). However, while the number of
active bursting sites (NASB) increased from the
first period to the second, it returned close to the
For the hippocampal culture, the evolution of
spiking activity at network level could be
clearly divided into three periods (Figure 1A –
blue bars) as follows: 7 DIV to 14 DIV as the
first period (C1-1), 15 DIV to 26 DIV as the
second period (C1-2) and 27 DIV to 39 DIV as
the last period (C1-3). An obvious finding was a
significantly increasing network firing rate
(NFR – Methods 2.2) (p < 0.05, t-test) from one
period to the next, starting with a mean firing
rate of 1.2 spikes/s in the first period to 12.7
spikes/s in the last as indicated by the yellow
7
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
value of the first period at the end of the
recording. It dropped sharply during a power
blackout (temperature dropped and stayed at
room temperature for several hours), from
which it recovered slowly to its previous value
(see Figure 1D and corresponding yellow circles
depicting means). In the second period, these
extremely high NASB are explained by massive
neural avalanches that take place within the
network and recruit neurons at most sites (ca.
82% of the 58 recording electrodes) for a short
time. Figure 1C exemplarily shows such a
network avalanche that arose at 18 DIV.
In the more mature cortical culture, we found
less bursting activity at network level than in the
hippocampal culture at earlier developmental
stages, but with a high NASB. That is, a larger
number of neurons contributed to the network
bursting, but with a lower number of bursts/s,
which did not lead to a comparable increase in
the NBR. In addition, the NBR, NASB and NFR
of the cortical culture oscillated within each of
the six periods, as indicated in Figure 1A, B and
D (red bars).
Figure 1 Evolution of network firing rate (NFR) (A) and network burst rate (B) over time for the first hippocampal (blue)
and second cortical (red) culture. Three (hippocampal, blue rectangular delimiters, C1-1 to C1-3) and six (cortical, red
rectangular delimiters, C2-1 to C2-6) recording periods were distinguished by significant changes in their NFRs. For each
period, the mean NFRs and their SDs are displayed as circles with error bars. (C) Example of a network avalanche at 18
DIV.(D) Number of active sites (NASB) with respect to bursting activity in the hippocampal (blue) and cortical (red)
culture.
3.2
activity as reported before (van Pelt et al.,
2005). While spike pairing was very low at
early DIVs, it consistently increased after 3-4
weeks in vitro (WIV). This trend was robust
not only at individual sites (Figure 2C, blue
bars), but also at network level (Figure 2D,
blue bars). The mean of PS occurrences at
individual sites significantly (p < 0.05)
Evolution of paired spiking activity at
individual sites and at network level
Activity patterns consisting of PSs separated
by ISIs of up to 5 ms were rarely encountered
in the young hippocampal culture. Instead,
random isolated spiking rather than
synchronized rapid firing dominated neural
8
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
increased about thirtyfold from 193 for the
first period to 5741 for the last period (Figure
2C, yellow circles indicating mean highest
number of PS at individual sites). At network
level, the mean number of PSs in the first
period was 486 and increased tenfold to 4911
in the last period (Figure 2D, yellow circles).
In contrast, PS activity in the more mature
cortical culture did not grow monotonically,
but fluctuated rather synchronously with both
the firing and burst rates at network level
(Figure 2C, D, red bars; Figure 1A, B).
Figure 2 Highest PS activity (black circles) and highest bursting activity (orange circles) in the hippocampal (A) and
cortical (B) culture may not necessarily be recorded from the same electrode in subsequent trials. (C) Number of
occurrences of the most frequently, locally occurring PS L in the hippocampal (C1, blue bars) and cortical culture (C2, red
bars) at the individual recording sites denoted on the y-axes in (A) and (B). (D) Total number of PSN occurrences at
network level for the hippocampal (C1, blue bars) and the cortical culture (C2, red bars). Circles and error bars in (C) and
(D) display mean PS values and their SDs (yellow: hippocampal; green: cortical).
average values exactly at those DIVs with
strong activity avalanches (as exemplified in
Figure 1C at 18 DIV). This suggests that such
extremely high neuronal activity at individual
sites (as revealed by the NASPS in Figure 4A,
blue bars) accounts for the elevated PSN
activity at network level, while firing rates
kept a uniformly increasing trend at those
DIVs (Figure 1A). After three WIVs, PS
activity dramatically increased (Figure 2C,
individual sites; Figure 2D, network level, blue
bars), while the duration of the most frequently
occurring IPSIsmfo decreased and robustly
settled at 2-3 s until the end of the recording
session (Figure 3A, individual sites; Figure
3B, network level, blue bars). Interestingly, the
duration of the IPSIsmfo at network level
Interestingly, the IPSI (see Methods2.3), a
parameter which quantifies the temporal gap
between two consecutive PS, exemplarily
suggests that PS activity becomes robust as the
culture ages. For the hippocampal culture, the
duration of the most frequently occurring
(mfo) IPSIsmfo at both individual sites (≤ 90 s,
index ‘L’) and network level (≤ 48 s, index
‘N’) was very long and fluctuated highly with
a low number of repetitions in the first
three WIVs (Figure 3A, individual sites;
Figure 3B, network level, blue bars). The nonuniform temporal distribution of PS activity
during this developmental period denotes that
PS was not yet robust. Remarkably, PSN
activity at network level shows higher than
9
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
(Figure 3B) stabilized earlier than at individual
sites (Figure 3A). Over the same period, the
number of the IPSIsmfo increased consistently
both at individual sites and network level
(Figure 3C and D, blue bars). These three
trends (increasing overall number of PS,
decreasing duration of IPSImfo, increasing
number of the most frequently occurring
IPSImfo) suggest that PS activity develops
homogeneously and consistently throughout
the network to result in robust PS activity
patterns with an increasing number of
occurrences with rather constant ISIs and
IPSIs, especially after three WIVs.
For the more mature cortical culture, we
observed the same inverse correlation between
duration and number of the IPSIsmfo (see
Figure 3A and D, red bars), however,
oscillating over time. There were recurring
periods with large IPSI values and low
numbers of IPSI repetitions followed by
periods with lower IPSI values but high
repetition frequencies.
Figure 3 Duration of the most frequently occurring (mfo) IPSI mfo at individual sites (A) and at network level (B) for the
hippocampal culture (C1, blue bars) and the cortical culture (C2, red bars) and its evolution from one period to the next
(C1-1 to C1-3 and C2-1 to C2-6, respectively). Number of the IPSI mfo at individual sites (C) and at network level (D)
(with the same color and period coding as in (A) and (B)).
3.3
could be found in less than 40% of the bursts
(see Figure 4B, blue bars). The same trend was
observed for the cortical culture where, except
for seven trials (3.18% of total trials), the
percentage of bursts that contained PS
remained below 50% (see Figure 4B, red
bars).
PS activity versus burst activity
In order to check if PS activity is one of the
driving forces for the self-organization of
functional
network
connectivity,
we
investigated the relationship between bursting
and PS activity. For the hippocampal culture,
on average only 11% of bursts contained PS at
early DIVs. Their number slightly increased to
16% in the last recording period. Except for
four trials (6.25% of total trials), PS activity
10
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
PS and bursting activity coincided in 42 trials
(38.18 % of total trials) for the cortical culture,
41 of them occurred after 44 DIV. In both
cases, the network location with dominant PS
and bursting activity could change over the
days. Over the course of the entire recording,
highest PS activity was detected on 12
different electrodes (21%) for the hippocampal
culture and on 11 electrodes (19%) for the
cortical culture. Highest bursting activity could
be associated with just five electrodes (8.3%)
in the hippocampal culture (Figure 2A) while
it occurred on 12 electrodes (20%) in the
cortical culture (Figure 2B).
Importantly, in almost 92% of the recording
trials PS activity preceded bursting activity
(Figure 4C, blue bars) in the younger
hippocampal culture and in 84% of the
recording trials for the more mature cortical
culture (Figure 4C, red bars). Additionally, the
average temporal delay between a PS and a
burst mostly remained below 50 ms. This
suggests that PS presumably initiated bursting
activity. Interestingly, this PS-burst coupling
occurred on the same electrode of the cortical
culture in 34 instances, suggesting robustness
of PS-dominant sites. This electrode also
recorded highest PS activity in almost 50% of
the trials.
Figure 4 (A) NAS for PS activity for the hippocampal
(blue) and the cortical (red) culture. (B) Percentage of
bursts that contained PS at individual site level. (C)
Average temporal delay in ms with which a burst
followed a PS at individual sites.
We further investigated the stability of the
spatio-temporal distribution of PS patterns.
The middle insets in Figure 5 (i1 - i6)
exemplarily show color-coded PS spiking
activity maps at individual sites in the
hippocampal culture for six trials at different
developmental stages. Robust spatial patterns
of PS activity were found over the entire
recording period; two examples are
highlighted by blue circles.
For both the hippocampal and cortical
network, PS and bursting activity were present
at most sites (Figure 5A). However, in most
cases, highest PS activity was recorded from
different electrodes than those that recorded
the highest bursting activity, as pointed out by
the black circles (PS) and orange circles
(burst) in Figure 2A (hippocampal) and Figure
2B (cortical). For the hippocampal culture,
highest PS and bursting activity were spatially
collocated in only six trials (9.3% of total
trials). In contrast, the electrodes with highest
A total of 12 sites with stable PS patterns
could be identified in more than 50% of the 60
trials. Among these sites, eight lasted longer
than 73% of the total recording period. They
formed robust, long-lasting patterns that
presumably recruited new neighboring sites in
11
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
different trials. Furthermore, seven of these
sites also formed robust burst patterns as
marked by yellow circles in Figure 5A.
Figure 5B exemplarily shows PS activity
patterns that were formed during the first
period (yellow circles) and lasted until the end
of the recording session, patterns that were
newly formed during the second period and
lasted until the end (black circles) and new
neighboring sites that emerged only during the
third recording period (white circles). Thus, 10
sites formed patterns that lasted for more than
four trials during the first period, 22 sites for
the second period and 34 sites for the last
period. That is, for each new period, up to 12
neighboring sites were recruited in generating
PS activity, thereby increasing the degree of
PS pattern complexity as the network entered
later developmental stages.
3.4
Information content per spike and
CFP analysis
Next we asked to what degree spontaneous in
vitro PS activity preserves its role encountered
in vivo and thus participates in the formation
of functional connectivity and in information
processing within the cultured neural network
at different developmental stages. With this
motivation in mind, we considered PS activity
as an internal stimulus and thus calculated the
PS-related CFP (see Methods 2.6) and PSrelated information content (see Methods 2.5)
in the hippocampal network both at individual
sites and at network level.
Figure 5 (A) Spatial PS and burst pattern distribution
with respect to the 8 x 8 electrode matrix for the
hippocampal culture. Yellow circles mark the seven
electrodes from which both PS and bursting activity
could be recorded according to the criterion described in
Methods 2.2 and 2.3. One trial represents half a DIV
(Methods 2.1). Marked electrodes recorded activity in
the majority of the trials, though not necessarily
consecutively. (B) Evolution of PS pattern complexity
in the hippocampal culture from one period to the next:
the NAS with PS patterns increased during
development. In most cases, new PS emerged on
electrodes adjacent to those with previous PS activity.
For the second, more mature cortical culture, 54 sites
(93%) participated in PS activity that lasted for at least
two trials, while in more than 50% of the trials the
number of sites decreased to four. This suggests that at
later developmental stages the role of dominant sites
gains importance.
Because 12 sites presented robust, long-lasting
PS activity for more than 50% of the total
trials (Figure 5A), we exemplarily calculated
the PS-related correlation between eight
closest neighbors out of these 12 sites as
indicated in Figure 6A (red ellipses).
12
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
We used CFP to construct the interconnectivity
maps
for
both
the
interconnectivity strength (Figure 6B) and the
temporal delay (Figure 6C). For each trial and
site i (i = 1:8), we quantified the PS-related
CFP (i,j) for all possible pairs (j = 1:7).
Interestingly, we found no connectivity during
the first 11 DIV, coinciding with the period in
which PS activity was not robust yet (i.e. with
large IPSIs and few repetitions). In contrast,
from 17 DIV onward the number of
connections significantly increased (p < 0.01)
and remained high until 27 DIV. This
corresponds to the second period, where PS
activity gained robustness. The highest number
of connections was found in this period
(Figure 6D), which decreased thereafter.
Moreover, the temporal delay of the
correlations between formed pairs increased
until 18 DIV with a mean of up to 45.2 ms
(±20 ms) and consistently decreased thereafter
with a mean of 16.2 ms (±10 ms) (Figure 6E
and F). Connectivity strength stayed rather
constant over several DIVs with almost
identical means of around 0.07 (±0.03) for all
three recording periods (Figure 6E).
information content per spike with PS, it does
not make any statement on how the content is
actually carried by the PS and on whether the
PS is the only information carrying
mechanism.
The three observations, i) the decreasing trend
in temporal delay between PS-induced
correlated activity, ii) an almost constant
interconnectivity
strength
at
later
developmental stages and iii) the formation of
robust spatiotemporal PS activity patterns as
the culture matured may indicate that PS
activity participates in the development and
stabilization of functional connections at
individual sites. To check for the robustness of
the PS-related connectivity map, we
constructed artificial Poisson-like spike trains
according to Eq. 1 (Methods 2.3) for these
eight electrodes (Figure 6A, red circles) for all
trials.
Next we investigated whether the PS-related
connectivity trend at individual sites is also
found at network level for the different
developmental stages. Firstly, we checked the
robustness of each NST by asking whether PS
activity at network level is a “by chance”
result of mapping all spikes from individual
sites onto a single timeline. We therefore
shuffled all of the NSTs repeatedly for 100
times and quantified PS activity for each
individual case. We then investigated whether
the artificial spike trains that mimic the
recorded spike trains (i.e. artificial spike trains
have the same firing rates as the recorded
ones) develop similar connectivity maps as the
real spike trains. Robustly, we found no PSrelated connectivity between the constructed
spike trains for any of the trials in the artificial
spike trains.
Next we looked at the PS-related information
content per spike (Eq. 2, Methods 2.5) for each
selected channel pair (i,j) of the eight
interconnected sites. The resulting information
content map is presented in Figure 8(A).
Information content per spike was highest
during the second period, thereby correlating
with the highest numbers of connections
between these eight most active channels
(Figure 8B compared with Figure 6B and D).
This trend is also reflected by the mean values
(Figure 8B). The initial increase in information
content per spike from 0.8±0.2 bits/spike
during the first period to 2.2±0.3 bits/spike in
the second period is followed by a decrease to
1.2±0.6 bits/spike during the last period. While
this statistical measure associates the
13
Figure 6 (A) Spatial arrangement of the eight closest, most PS-active sites in the hippocampal culture with respect to the
8x8 MEA matrix layout. The first number in a pair refers to the column, the second to the row. The insets under (A)
exemplarily show PS-related CFP curves of reference channel 31 vs. channel 42 at three different DIVs (14, 19 and 28).
The flat CFP curve framed by a red box illustrates the lack of PS-related connectivity for artificial spike trains that mimic
channels 31 and 42. Connectivity evolution over time expressed in strength (B) and temporal response delay (C) between
the exemplarily selected eight most active PS sites. Each pixel column represents one of the seven recording sites
connected to the respective reference channel indicated on the x-axis and pointed out in (A). Sorting order is column, then
row. The red vertical bars delimit the eight reference channel permutations. (D) Evolution of the number of connections
between the selected channels. (E) Evolution of the average connectivity strength for the selected channels and their
means for each period (yellow circles). (F) Evolution of the average connectivity time delays for the selected channels
and their means for each period (yellow circles).
connectivity remained fairly constant, the time
delay decreased during the last period (28
DIV). In contrast, the red box inset
exemplarily shows no connectivity between
the two artificial spike trains that mimic the
Insets of Figure 6A exemplarily show the
connectivity between PS activity of channel 31
and spiking activity at channel 42 for three
different DIVs. As mentioned before, while for
the recorded spike trains the strength of the
14
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
same two electrode recordings over the same
period.
statistically significant (p<0.001) different PS
activity as if it was strictly governed by firing
rates (Figure 7C, green bars). These Poissonlike spike trains lack a spike history (i.e.
without refractory period). Thus, very large
NST firing rates and an exponential ISI
distribution (Eq. 1b) favor short ISIs (i.e. up to
five ms), which leads to an unrealistically high
number of PS occurrences.
We found the PS activity for each trial to be
almost zero (Figure 7C, blue bars represent PS
in recorded NSTs; inset with red bars represent
PS in shuffled NSTs). To mimic the recorded
NSTs, we further constructed artificial
Poisson-like spike trains with similar firing
rates as the NSTs (Eq.1, Methods 2.3). Also in
this case, such artificial NSTs showed
Figure 7 (A) Evolution of PS-related CFP strength for all hippocampal NSTs. (B) Evolution of PS related CFP time delay
for all NSTs. (C) The number of PS for artificial spike trains at network level (green bars) and the number of PS at
network level for recorded NSTs (blue bars). The inset shows a zoom onto the number of PSs for shuffled NSTs (red
trace). The insets R1 – R6 (left side column) display examples of PS-related CFPs at network level for different DIVs
(11, 12, 14, 22, and 28). The insets A1 – A6 (right side column) display examples of PS-related CFPs at network level for
artificial spike trains for the same DIVs (11, 12, 14, 22, and 28).
15
spike in a PS (with a decreasing time delay)
for all NSTs, we checked whether PS activity
may also be involved in information
processing at network level. We calculated the
PS-related information content per spike at
network level for each NST considering PS
activity as an internal stimulus (Eq. 2,
Methods 2.5). Indeed, we found that in both
cultures PS activity is involved in information
processing at network level as well (Figure
8C). Moreover, in the hippocampal culture, the
trend found at local sites was preserved at
network level. For the first period, we found a
mean of 3.4 bits/spike (SD = 1.3 bits/spike),
which increased during the second period to
4.9 bits/spike (SD = 0.4 bits/spike) and
decreased in the last period to 3.6 bits/spike
(SD = 1.2 bits/spike). For the cortical culture,
the information content increased and
decreased for different periods from 4.4
bits/spike (SD = 1.8 bits/spike) up to a value of
6.5 bits/spike (SD = 0.8 bits/spike).
Next, we checked for PS-correlated activity at
network level by calculating the CFP
(Methods 2.6) for every NST, this time with
respect to developmental evolution of
network-wide, PS-induced activity instead of
local connectivity. As before, the second spike
in a spike pair (that matched the PS criterion
of ≤5 ms) of an NST served as the reference
for calculating the CFP with respect to the
following 500 ms of NST activity. This autocorrelation-like analysis provided information
on the strength and time delay of PS-related
spiking activity for individual NSTs. If
repeated for all NSTs, the evolution of PScorrelated activity can be plotted for all trials
(Figure 7A, B). Remarkably, we found that
PS-related spiking activity started at 11 DIV
for all of the trials. Its time delay increased
until 18 DIV and significantly decreased
during the last recording period, which is
strikingly similar to the previously observed
trend at individual sites as reported above.
3.5
The mean time delay (Figure 7B, yellow
circles) increased during the second recording
period to 74.5±30 ms and decreased in the
third period to 12.4±10 ms while the mean
correlation strength (Figure 7A, yellow circles)
increased during the second period to
0.28±0.09 and decreased to 0.14±0.09 in the
last recording period.
PS activity participates in controlling
the direction of information flow
within the coupled neuronal units
So far we have seen that PS activity is
involved
in
establishing
functional
connectivity and carries information both at
local side and at network level. We then asked
whether PS is also controlling the direction of
information flow within the cultured network.
A TE analysis (Methods 2.7) may reveal how
presynaptic PS activity predicts activity at its
postsynaptic target or even at network level.
We exemplarily choose the eight most closest
recorded units mentioned above (as depicted in
Figure 6) and calculated ΔTE (Eq. 5) for PS
activity for each of the channels with respect
to the entire spiking activity of the remaining
seven channels for different time lags and
message lengths (Ito et al., 2011). We found
that PS activity is robustly involved in
information transfer between cultured neurons.
Furthermore, despite significantly larger PS
activity found in artificial spike trains that
were supposed to mimic NSTs, we found no
correlated spiking activity in any of the trials
(insets A1 – A6 to the right side of Figure 7).
In contrast, insets R1 – R6 on the left side of
Figure 7 exemplarily show the CFP of PSrelated spiking activity for six of the NSTs
indicating a decreasing time delay of the PScorrelated spiking activity.
Finally, after noticing that PS activity is
involved in developing functional connections
at individual sites and that spiking activity at
network level is correlated with each second
16
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
connectivity between selected recording units
at an early developmental stage (i.e. the first
17 DIVs) with a mean ΔTE = 0.2·10-2 bits/s
and SD = 0.18·10-2 bits/s. Additionally, for
some channels, PS activity did not establish
any effective connectivity during this period.
Only three out of eight selected channels
showed effective connectivities with their
postsynaptic partners. Those disappeared and
reappeared in different recording periods
indicating that PS activity was not robust yet
and consequently could not reliably predict or
cause spiking activity at their targets. This
situation changed dramatically from 17 DIV
onward. ΔTE significantly increased (p<0.001)
up to 29 DIV indicating that PS activity
established robust and effective connectivities.
Only on channel 51 PS activity established
less effective connectivities during this
developmental stage. For this developmental
period we found a mean ΔTE = 1.13·10-2 bits/s
with a SD = 0.14·10-2 bits/s. From 30 DIV
onward until the end of the recording period,
ΔTE decreased with a mean of 0.4·10-2 bits/s
and a SD = 0.13·10-2 bits/s, while only four
(50%) channels with PS activity had
established effective connectivity with large
ΔTEs. Remarkably, PS activity could predict
98.58% of the spiking activity at their targets.
Only in 1.42% of the cases ΔTE took on
negative values (Figure 9A, dark blue pixels),
which shows that PS activity at individual sites
could be predicted reversely from the spiking
activity of their postsynaptic partners.
Figure 8 (A) Matrix of PS-related information content
expressed in bits/spike between the exemplarily selected
eight most active PS sites in the hippocampal culture
(Figure 6). Each pixel column represents one of the
seven recording sites being connected to the respective
reference channel indicated on the x-axis. Sorting order
is column, then row. The red vertical bars delimit the
eight reference channel permutations. (B) Evolution of
the PS-related mean information content in bits per
spike (blue bars) for the selected channels and its
respective average for a given period (yellow circles).
(C) Evolution of information content for PS-related
activity in both cultures at network level for each
period. Blue bars represent the hippocampal culture
with mean and standard deviations in yellow; red bars
represent the cortical culture with mean and standard
deviation in green.
Further on we asked if the local effect is
preserved at network level. In this case, the PS
activity at the eight selected channels
represented now the senders and the NSTs
were considered the receivers. As Figure 9B
shows, the trend that was observed at local
sites was also found at network level. Again,
until 17 DIV, spiking activity of NSTs could
be poorly predicted by PS activity only (mean
ΔTE = 0.4·10-2 bits/s with SD = 0.3·10-2
bits/s). Only two channels (i.e. 31 and 42)
could be identified in causing spiking activity
Figure 9A shows the constructed ΔTE map,
which indicates a very low effective
17
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
in the NSTs. From 17 DIV to 29 DIV, the
mean ΔTE significantly increased to 1.5·10-2
bits/s with a SD = 0.2·10-2 bits/s suggesting
that PS activity at local sites increasingly
drove network activity. However, from 30
DIV onward, PS activity increased only
slightly (mean ΔTE = 1.7·10-2 bits/s with a SD
= 0.3·10-2 bits/s), mostly due to the same trend
as observed at local sites. That is, the same
four channels (50% out of the eight selected
channels) presented larger ΔTE values and
thus strengthened their influence on the
spiking activity of each NST. Moreover, at
network level, ΔTE never took on negative
values, suggesting that PS activity at selected
local sites always predicted NST activity and
not vice versa.
4.
Discussion
Neurons from dissociated brain tissue are
capable
of
self-organizing
their
interconnectivity in cell culture. They become
active even in the absence of any external
sensory stimuli (Feller et al., 1999). Besides
random spiking and concerted bursting, neural
networks use various modalities and activity
patterns to both transfer information and form,
as well as maintain, functional connections
(Sun et al., 2010). Such spontaneous, often
synchronized neural activity increases in firing
rate and in the number of active sites from one
developmental stage to the next, as observed in
vivo (Nadasdy 2000; Chiu et al., 2001; Weliky
et al., 1999), as well as in vitro (van Pelt et al.,
2004, Wagenaar et al., 2006; Rolston et al.,
2007; Pasquale et al., 2010). Our uninterrupted long-term recording study,
spanning over several weeks, confirmed this
trend in two different neural in vitro networks,
in a young hippocampal culture and for the
first two periods in a more mature cortical
culture. At later developmental stages, the
activity and number of active sites slightly
decreased. In addition, bursting frequency at
network level did not increase anymore.
Instead, it distributed spatially by involving
more active sites. In contrast to previously
reported snapshot activity recordings, our
almost continuous recordings revealed large
activity variations at particular DIVs. We
therefore asked what network-inherent coding
mechanisms shape and drive network activity.
Interestingly, we found that predominantly
robust PS activity rather than bursts drove
neural activity in the investigated cultures. PS
not only developed stable spatiotemporal
patterns, but also participated in shaping the
interconnectivity map. Previous reports on its
role in vivo and in vitro suggest that PS
activity may act as a temporal filter and be part
of a mechanism involved in information
processing at different hierarchical information
processing stages (Krahe 2004; Akerberg
Figure 9 (A) Matrix of PS-related ΔTE (Methods 2.7)
for eight recording sites. Vertical red bars delimit the
sender channel that connects with the remaining seven
receiver channels. Individual pixels represent the PSrelated ΔTE averaged over a 30 minutes period and over
the seven receiver channels. Each pixel row covers one
DIV. (B) Matrix of PS-related ΔTE for eight senders
(delimited by red bars) and the entire network as the
receiver. Each pixel represents the ΔTE averaged over a
30 minutes period, each row covers one DIV.
18
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
2011; Rathbun et al., 2010; Sincich et al.,
2009; Uglesich et al., 2009).
of several tens of milliseconds. This suggests
that PS may act as an internal surrogate
stimulus, which triggers synchronized neural
activity and avalanches (i.e. exemplified at 18
DIV, Figure 1C).
In this study, the following main findings in
support of these statements emerged from the
analysis of the continuously recorded datasets:
7. Its relation to other types of activity
furthermore suggests that PSs act as networkintrinsic stimulus sources also at network
level. From 11 DIV onward, network spiking
activity was strongly correlated to the second
spike of a PS with decreasing temporal delay
as cultures matured.
1. The network firing rate, network burst rate
and number of active sites increased as a
hippocampal culture grew toward maturity and
slightly decreased when the respective cultures
started to decay.
2. At later developmental stages, spontaneous
neural activity in a cortical culture oscillated
periodically over several days. While activity
could be very high at some DIVs, on average it
evolved rather constantly.
8. Besides temporal patterns consisting of
short IPSIs and frequent occurrence, PS
activity formed increasingly complex spatial
patterns during development by recruiting
neighboring sites. Some of them fluctuated
temporally; others lasted until the end of the
recording session. A similar, yet activity typeunspecific increase in pattern complexity has
been reported before (Rolston et al., 2007; Sun
et al., 2010).
3. PS activity became robust after three WIVs
when IPSIs settled down to 2-3 s and the
number of PS occurrences significantly
increased at both individual sites and at
network level. This may signal the passing of a
critical maturation stage in spontaneously
active in vitro networks.
These findings combined with the above
mentioned results suggest that PS activity
evolves in distinct spatiotemporal patterns
within a non-stimulated, spontaneously active
network. PS presumably initiates synchronized
bursting activity that may be responsible for
forming particular functional connections both
in vivo and in vitro (Rolston et al., 2007; Sun
et al., 2010; Chiappalone et al., 2012;
Blankenship et al., 2010; Mazzoni et al., 2007;
Pimashkin et al., 2011).
4. In both cultures, highest PS activity could
change its spatial location from one DIV to the
next, which in most cases did not coincide
with the location of the highest bursting
activity. Furthermore, highest PS activity
involved a larger number of neurons than the
highest bursting activity, which stayed
spatially confined to a few dominant
electrodes throughout the entire recording
period.
9. CFP and information content per spike have
been proven to reveal the statistical
dependency between coupled neurons
(Maccione et al., 2012; le Feber et al., 2007).
Our CFP analysis revealed that PS activity is
likely involved in the establishment and
shaping of functional connections between
different individual sites within the network.
While we found few or no PS-related
correlations between different sites at early
stages when PS activity was not yet robust, the
5. Although PS activity was found both
outside and inside of bursting activity, it is an
independent type of neural response, typically
spatially separated from bursting activity. For
the majority of the trials, the percentage of
bursts which contained PS activity remained
below 50%.
6. In most trials, PS activity preceded bursting
activity at the same recording site with a lead
19
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
number of connections increased as the culture
matured. They were characterized by rather
constant average strengths and decreasing time
delays for subsequent developmental stages.
Interestingly, the PS-related information
content per spike at individual sites was
highest for the same period that the number of
connections was found to be highest. These
findings suggest a consolidation of PScorrelated activity at individual sites over time.
besides forming spatio-temporal patterns and
being involved in the formation of functional
connectivity within the cultured neurons, also
predicts the directionality of a connection at
both local sites and network level (Figure 9A).
TE analysis also strongly supported the trends
revealed by information content per spike and
CFP analysis. Presumably, PS activity was
expansively involved in driving neuronal
communication between 17 DIV and 29 DIV
to arrive at a relative stability after 29 DIV
where the PS neurons played a key role in
driving neural activity in the already mature
network.
10. As the elevated PS-related information
content per spike indicated, PS activity was
involved in information transmission at
individual sites and at network level. This
finding is in concordance with previous studies
on the role of different spiking patterns in
spontaneous neural activity (Wagenaar et al.,
2006; Pasquale et al., 2010; Sun et al., 2010;
Rolston et al., 2007; Nadasdy 2000). It
suggests that the network presumably uses
only a fraction of the total number of spikes to
transmit most of the information. As
mentioned earlier, this concept of sparse
coding was also found in the early visual
system. It improves the overall coding
efficiency by a mechanism that deletes the less
informative spikes from one stage to the next
(i.e. from the retina to the lateral geniculate
nucleus (LGN)) while preserving relevant
information with a lower number of spikes
(Sincich et al., 2009; Uglesich et al., 2009).
The construction of NSTs by simply collecting
the entire spiking activity and arranging the
time stamps in an ascending temporal order
may seem problematic at first glance. This
way of constructing NSTs does not seem to
have any biological relevance. However, using
artificial spike trains and shuffling methods,
we could show that PS in NSTs does not occur
by chance. While artificial spike trains have
similar firing rates as the recorded trains, no
correlated activity between neurons (other than
by pure random) is expected due to the lack of
connectivity between individual neurons. The
absence of PS-related connectivity for
artificial spike trains suggests that PS activity
is not a random and strictly firing ratedependent neural phenomenon. Instead, it
seems to be an intrinsic mechanism of cultured
neurons in support of shaping neural
interconnectivity. The local effect of PS
activity seems to be preserved in NSTs as
well. Figure 9B revealed a similar trend in the
prediction of spiking activity at network level
by PS. CFP analysis revealed a strong
correlation between spiking activity at network
level and each second spike in a PS. A
similarity between CFP and information
content per spike shapes strengthened the
hypothesis that PS is involved in carrying
information at network level as well. In
summary, PS seems to play a key role in
shaping the local and network-wide input-
11. The transfer entropy (Schreiber, 2000), an
asymmetric information theoretic measure
recently introduced to neuroscience, allows to
estimate the direction of information flow
within a network of locally coupled neurons
(Ito et al., 2011 – for spiking cortical network;
Gourevitch and Eggermont, 2006 – for
auditory cortical neurons; Garofalo et al.,
2009) or even between different brain areas
(Battaglia et al., 2012 – inter-areal brain
circuits; Buehlmann and Deco, 2010; Lindner
et al., 2011 – directed interactions from the
retina to the tectum; Lungarella and Sporns,
2006 – sensorimotor networks). Here we used
TE to independently confirm that PS activity,
20
Martiniuc et al.
output relationship
networks.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
in
cultured
neural
in support of this research are highly
appreciated.
The two main hypotheses stated in this study
have to be tested further in future experimental
work. Firstly, in lack of any external stimulus,
does PS indeed act as an internal surrogate
stimulus that is capable of shaping neural
activity by driving the input – output
relationship of the spiking activity at network
level? If this assumption turned out to be true,
controlled PS-like electrical stimulation
(similar to Zullo et al., 2012) instead of single
pulses or tetanic stimuli could more reliably
drive a predictable neural output, i.e. in a
closed-loop stimulation paradigm (Rolston et
al., 2010; Ruaro et al., 2005; Novellino et al.,
2007). PS-like stimulation may find possible
application in neurally-controlled artefacts
(robotics, neuroprosthetics).
References
Akerberg, O.A. and Chacron, M.J. (2011). In vivo
conditions influence the coding of stimulus features
by bursts of action potentials. J.Comput. Neurosci.
31(2), 369-83.
Banker, G. and Goslin, K. (1998). Culturing Nerve
Cells. Cambridge: The MIT Press.
Battaglia, D., Witt, A., Wolf, F. and Geisel, T. (2012).
Dynamic effective connectivity of inter-areal brain
circuits. PLoS Comput. Biol. 8: 3.
doi:10.1371/journal.pcbi.1002438.
Bienenstock, E.L., Cooper, L.N., and Munro, P.W.
(1982). Theory for the development of
neuronselectivity: orientation specificity and
binocular interaction in visual cortex. J Neurosci. 2,
32–48.
Blankenship, A.G. and Feller, M.B. (2010).
Mechanisms underlying spontaneous patterned
activity in developing neural circuits. Nat. Rev.
Neurosci. 11, 18–29.
Blau, A., Neumann, T., Ziegler, C., and Benfenati, F.
(2009). Replica-molded poly(dimethylsiloxane)
culture vessel lids attenuate osmotic drift in longterm cell culturing. J.Biosci. 34(1), 59-69.
Brenner, N., Strong, S.P., Koberle, R., Bialek, W. and
de Ruyter van Steveninck, R.R. (2000). Synergy in a
neural code. Neural Comput. 12, 1531–1552.
Brewer, G.J., Boehler, M.D., Ide, A.N. and Wheeler,
B.C. (2009). Chronic electrical stimulation of
cultured hippocampal networks increases
spontaneous spike rates. J. Neurosci. Methods 184,
104-109.
Buehlmann, A. and Deco, G. (2010). Optimal
information transfer in the cortex through
synchronization. PLoS Comput.Biol. 6(9):e1000934,
doi: 10.1371/journal.pcbi.1000934.
Chiappalone, M., Vato, A., Berdondini, L., KoudelkaHep, M., and Martinoia, S. (2007). Network
dynamics and synchronous activity in cultured
cortical neurons. International Journal of Neural
Systems 17, 87–103.
Chiu, C. Y. and Weliky, M. (2001). Spontaneous
activity in developing ferret visual cortex in vivo. J.
Neurosci., 21(22), 8906-8914.
DeMarse, T.B., Wagenaar, D.A., Blau, A. and Potter,
S.M. (2001). The neurally controlled animat:
Biological brains acting with simulated bodies.
Autonomous Robots 11, 305–310.
Feller, M. B. (1999). Spontaneous correlated activity in
developing neural circuits. Neuron 22, 653–656.
Secondly, because PS activity cannot only be
found downstream of a stimulus both in vitro
and in intact brain architectures (Usrey et al.,
1998; Sincich et al., 2009), but also in
spontaneously firing neural in vitro networks,
it may represent an ubiquitous information
processing response property of neurons of
different species and organization. However,
little is known on how PS depends on the
neural cell type. A prominent example is the
early visual system, where the PS contribution
to information processing varies for different
retinal ganglion cell types and the number of
synaptic connections they are engaged in
(Martiniuc et al., in preparation). A combined
electrophysiology and imaging in vitro study
on how PS depends on cell type and cell
morphology will shed more light on this
question.
Acknowledgements
We thank Francesca Succol and Marina Nanni
for their expert advice and assistance in the
cell culture preparation. IIT intramural funds
21
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
Ferrández, J.M., Lorente, V., de la Paz, F. and
Fernández, E. (2012). Training biological neural
cultures: Towards Hebbian Learning.
Neurocomputing 114, 3-8.
Garofalo, M., Nieus, T., Massobrio, P. and Martinoia, S.
(2009). Evaluation of the performance of
information theory-based methods and crosscorrelation to estimate the functional connectivity in
cortical networks. PLoS ONE 4(8):e6482. doi:
10.1371/journal.pone.0006482.
Graham, D.J., and Field, D.J. (2007). Statistical
regularities of art images and natural scenes: spectra,
sparseness and nonlinearities. Spatial Vision 21,
149-164.
Gourevitch, B., Eggermont, J.J. (2007). Evaluating
information transfer between auditory cortical
neurons. J Neurophysiol. 97, 2533–2543.
Krahe, R. and Gabbiani, F. (2004). Burst firing in
sensory systems. Nat. Rev. Neurosci., 5(1), 13-23.
le Feber, J., Rutten, W.L.C., Stegenga, J., Wolters, P.S.
and Ramakers, G.J.A.(2007). Conditional firing
probabilities in cultured neuronal networks: a stable
underlying structure in widely varying spontaneous
activity patterns. Journal of Neural Engineering 4,
54–67.
Lindner, M., Vicente, R., Priesemann, V. and Wibral,
M. (2011). TRENTOOL: a Matlab open source
toolbox to analyze information flow in time series
data with transfer entropy. BMC Neurosci. 12, 119.
Lungarella, M. and Sporns, O. (2006). Mapping
information flow in sensorimotor networks. PLoS
Comput.Biol. 2(10):e144. doi:
10.1371/journal.pcbi.0020144.
Maccione, A., Garofalo, M ., Nieus, T., Tedesco, M.,
Berdondini, L. and Martinoia, S. (2012). Multiscale
functional connectivity estimation on low density
neuronal cultures recorded by high density CMOS
Micro Electrode Arrays. Journal of Neuroscience
Methods 207(2).
doi:10.1016/j.jneumeth.2012.04.002.
Martiniuc, A.V. and Knoll, A. (2012). Interspike
Interval Based Filtering of Directional Selective
Retinal Ganglion Cells Spike Trains.
Comput.Intell.Neurosci. doi:10.1155/2012/918030.
Martiniuc, A.V., Zeck, G., Stürzl, W. and Knoll, A.
(2011). Sharpening of directional selectivity from
neural output of rabbit retina. J. Comput. Neurosci.
30(2), 409-26.
Mazzoni, A., Broccard, F.D., Garcia-Perez, E.,
Bonifazi, P., Ruaro, M.E. and Torre, V. (2007). On
the dynamics of the spontaneous activity in neuronal
networks. PLoS ONE 2(5),e439. doi:
10.1371/journal.pone.0000439.
McCabe, A.K., Chisholm, S.L., Picken-Bahrey, H.L.
and Moody, W.J. (2006). The self-regulating nature
of spontaneous synchronized activity in developing
mouse cortical neurones. J. Physiol. 577, 155–167.
Nadasdy, Z. (2000). Spike sequences and their
consequences. J. Physiol.94, 505-524.
Nomura, M., Ito, D., Tamate, H., Gohara, K. and
Aoyagi, T. (2009). Estimation of Functional
Connectivity that Causes Burst-like Population
Activities. Forma 24 (1), 11-16.
Novellino, A., D'Angelo, P., Cozzi, L., Chiappalone,
M., Sanguineti, V. and Martinoia S. (2007).
Connecting neurons to a mobile robot: an in vitro
bidirectional neural interface. Comput. Intell.
Neurosci. 12725, doi: 10.1155/2007/12725.
Olhausen, B.A. and Field, D.J. (2004). Sparse coding of
sensory inputs. Curr.Opin.Neurobiol. 14, 481–487.
Pasquale, V., Massobrio, P., Bologna, L.L.,
Chiappalone, M. and Martinoia, S. (2008). Selforganization and neuronal avalanches in networks of
dissociated cortical neurons. Neuroscience 153,
1354–1369.
Pasquale, V., Martinoia, S. and Chiappalone, M. (2010).
A self-adapting approach for the detection of bursts
and network bursts in neuronal cultures. Journal of
Computational Neuroscience 29(1-2), 213-229.
Pimashkin, A., Kastalskiy, I., Simonov, A., Koryagina,
E., Mukhina, I. and Kazantsev, V. (2011). Spiking
signatures of spontaneous activity bursts in
hippocampal cultures. Front. Comput. Neurosci.
5:46. doi: 10.3389/fncom.2011.00046.
Rathbun, D.L, Warland, D.K., and Usrey, W.M. (2010).
Spike timing and information transmission at
retinogeniculate synapse. The Journal of
Neuroscience 30(41), 13558-13566.
Rolston, J.D., Wagenaar, D.A. and Potter, S.M. (2007).
Precisely timed spatiotemporal patterns of neural
activity in dissociated cortical cultures.
Neuroscience 148, 294–303.
Rolston, J., Gross, R. and Potter, S. (2010). Closedloop, open-source electrophysiology. Front.
Neurosci. 4:31. doi: 10.3389/fnins.2010.00031.
Ruaro, M.E., Bonifazi, P. and Torre, V. (2005). Toward
the neurocomputer : Image Processing and pattern
recognition with neuronal cultures. IEEE Trans.
Biomed. Engineering 52(3), 371-383.
Schreiber, T. (2000). Measuring information transfer.
Phys Rev Lett85, 461–464.
Shahaf, G. and Marom, S. (2001). Learning in Networks
of Cortical Neurons. The Journal of Neuroscience 21
(22), 8782–8788.
Sincich, L.C., Adams, D.L., Economides, J.R. and
Horton, J.C. (2007). Transmission of spike trains at
retinogeniculate synapse. The Journal of
Neuroscience 27(10), 2683–2692.
22
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
Sincich, L.C., Horton, J.C. and Sharpee, T.O. (2009).
Preserving Information in Neural Transmission. The
Journal of Neuroscience 29(19), 6207– 6216.
Sjöström, P.J., Turrigiano, G.G., and Nelson, S.B.
(2003). Neocortical LTD via coincident activation of
presynaptic NMDA and cannabinoid receptors.
Neuron39, 641–654.
Strong, S.P., Koberle, R., de Ruyter, van Steveninck,
R.R. and Bialek, W.(1998). Entropy and information
in neural spike trains. Phys. Rev. Lett. 80, 197–200.
Sun, J.J., Kilb, W. and Luhmann, H.J. (2010). Selforganization of repetitive spike patterns in
developing neuronal networks in vitro. Eur. J.
Neurosci. 32(8), 1289-99.
Uglesich, R., Casti, A., Hayot, F. and Kaplan, E. (2009).
Stimulus size dependence of information transfer
from retina to thalamus. Front SystNeurosci. 3:10.
doi: 10.3389/neuro. 06.010.2009.
Usrey, W.M., Reppas, J.B. and Reid, R.C. (1998).
Paired-spike interactions and synaptic efficacy of
retinal inputs to the thalamus. Nature 395(6700),
384-387.
van Pelt, J., Corner, M.A., Wolters, P.S., Rutten, W.L.
and Ramakers, G.J. (2004). Long term stability and
developmental changes in spontaneous network
burst firing patterns in dissociated rat cerebral cortex
cell cultures on multielectrode arrays. Neurosci.Lett.
361, 86–89.
van Pelt, J., Vajda, I., Wolters, P.S., Corner, M.A. and
Ramakers, G.J. (2005). Dynamics and plasticity in
developing neuronal networks in vitro. Prog.Brain
Res. 147, 173–188.
Vicente R, Wibral M, Lindner M, and Pipa G.
(2011).Transfer entropy — a model-free measure of
effective connectivity for the neurosciences. J.
Comput.Neurosci 30: 45–67. doi : 10.1007/s 10827
– 010 – 0262-3.
Wagenaar, D.A., Madhavan, R., Pine, J. and Potter,
S.M. (2005). Controlling bursting in cortical cultures
with closed-loop multi-electrode stimulation. J.
Neurosci. 25, 680-8.
Wagenaar, D.A., Pine, J. and Potter, S.M. (2006).
Searching for plasticity indissociated cortical
cultures on multi electrode arrays. J. Neg.Res.
Biomed. 5:16. doi:10.1186/1477-5751-5-16.
Wagenaar, D.A., Nadasdy, Z., Potter, S.M. (2006).
Persistent dynamic attractors in activity patterns of
cultured neuronal networks. Phys. Rev. E Stat.
Nonlin. Soft. Matter. Phys. 73 (5 Pt 1): 051907.
Weyand, T.G. (2007). Retinogeniculate Transmission in
Wakefulness. Journal of Neurophysiology 98, 769–
785.
Wibral, M., Pampu, N., Priesemann, V., Siebenhuehner,
F., Seiwert, H., Lindner, M., Lizier, J.T. and
Vicente, R. (2013). Measuring information-transfer
delays. PLoS ONE 8(2):e55809. doi:
10.1371/journal.pone.0055809.
Wiener, N. (1956). “The theory of prediction”, in:
Modern Mathematics for Engineers, ed. E.F.
Beckenbach (New York:McGraw-Hill), 165-190.
Weliky, M., Katz, L.C. (1999). Correlational structure
of spontaneous neuronal activity in the developing
lateral geniculate nucleus in vivo. Science
285(5427), 599-604.
Zullo, L., Chiappalone, M., Martinoia, S. and Benfenati,
F. (2012). A “spike-based” grammar underlies
directional modification in network connectivity:
effect on bursting activity and implications for biohybrids systems. PLoS ONE 7(11):e49299. doi:
10.1371/journal.pone.0049299.
23
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
Supplementary Information
Paired spiking robustly shapes spontaneous
activity in neural networks in vitro
Aurel Vasile Martiniuc1*, Victor Bocoş-Binţinţan2, Rouhollah Habibey3,
Asiyeh Golabchi3, Alois Knoll1, Axel Blau3
1
Computer Science Department VI, Technical University Munich, Boltzmannstraße 3,
85748 Garching, Germany, http://www.in.tum.de/
2
Dept. of Environmental Science & Engineering, Babeş-Bolyai University, 30 Fantanele Street,
400294 Cluj-Napoca, Romania, http://enviro.ubbcluj.ro/
3
Dept. of Neuroscience and Brain Technologies, Italian Institute of Technology, via Morego 30,
16163 Genoa, Italy, http://www.iit.it
*Correspondence: [email protected]
1.
Pictorial definition of spike train and burst parameters
Suppl. Fig. 1 Graphical depiction of spike train and burst parameters. (A) Spike cutout: only upward (positive) and
downward (negative) spike cutouts from 57 (cortical) and 58 (hippocampal) out of 59 recording electrodes were stored in
5 min packets. They consisted of 5 ms pre-spike and 5 ms post-spike fragments after first threshold crossing at ± 5.5 SD
with respect to peak-to-peak noise. Only timestamps from downward threshold-crossings were extracted using
24
Martiniuc et al.
Paired spiking robustly shapes spontaneous activity in neural networks in vitro
NeuroExplorer (Nex Technologies). After removing simultaneous timestamps that occurred on all channels due to
electrical or handling artefacts, subsequent 5 min datasets comprised of ≤ 58 timestamp streams were bundled in 12 hour
timestamp packets for further analysis in Matlab (MathWorks). These half-day packets were called trials of duration
TTrial ≤ 12 h. (B) We quantified the local firing rates (LFRs) at individual sites as the number of recorded spikes divided
by TTrial for each local spike train (LST). (C) At network level, we pooled all spikes from all 57 and 58 sites,
respectively, for each trial into a single network spike train (NST) by sorting them in a time-ascending order. The NST
represented the MEA-wide activity for each trial. The network firing rate (NFR) was then quantified as the total number
of spikes in an NST divided by T Trial. To further investigate activity dynamics, we used the burst rate (BR) for
characterizing synchronous network activity. We scanned all LSTs at the 57 (cortical) and 58 (hippocampal) individual
sites for each trial and defined bursting activity as events with more than 10 subsequent spikes individually separated by
an ISI of less than 100 ms, followed by an interburst interval (IBI) larger than 200 ms (Wagenaar et al., 2005). The local
burst rate (LBR) at individual sites was calculated by dividing the number of bursts by T Trial. Equally, the network burst
rate (NBR) was obtained by scanning the NST for bursts using above mentioned criterion and dividing the number of
bursts by TTrial. (D) Definition of a paired spike. Two subsequent spikes with inter-spike intervals (ISIs) between 2 and
5 ms were considered paired spikes (PS). Only PS with inter-paired-spike intervals (IPSIs) over 40 ms were counted.
2.
MEA layout
Suppl. Fig. 2 MEA electrode layout and channel association for the 8 x 8 MEA with the hippocampal (A) culture and
for the 6 x 10 MEA with the cortical (B) culture. Grey squares indicate the relative position of the grounded counter
electrode 15 (column, row). Blue squares with red numbers indicate switched-off channels.
3.
References
Wagenaar, D.A., Madhavan, R., Pine, J. and Potter, S.M. (2005). Controlling bursting in cortical cultures with closedloop multi-electrode stimulation. J. Neurosci. 25, 680-8.
25

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project