Available via license: CC BY 4.0
Content may be subject to copyright.
Charlesworth et al.
RESEARCH
Quantitative differences in developmental profiles
of spontaneous activity in cortical and
hippocampal cultures
Paul Charlesworth1,2†, Ellese Cotterill3†, Andrew Morton1,4, Seth G. N. Grant5and Stephen J. Eglen3*
*Correspondence: sje30@cam.ac.uk
3Cambridge Computational
Biology Institute, University of
Cambridge, Wilberforce Road,
CB3 0WA Cambridge, UK
Full list of author information is
available at the end of the article
†Equal contributor
Abstract
Background: Neural circuits can spontaneously generate complex spatiotemporal
firing patterns during development. This spontaneous activity is thought to help
guide development of the nervous system. In this study, we had two aims. First,
to characterise the changes in spontaneous activity in cultures of developing
networks of either hippocampal or cortical neurons dissociated from mouse.
Second, to assess whether there are any functional differences in the patterns of
activity in hippocampal and cortical networks.
Results: We used multielectrode arrays to record the development of spontaneous
activity in cultured networks of either hippocampal or cortical neurons every two
or three days for the first month after plating. Within a few days of culturing,
networks exhibited spontaneous activity. This activity strengthened and then
stabilised typically around 21 days in vitro. We quantified the activity patterns in
hippocampal and cortical networks using eleven features. Three out of eleven
features showed striking differences in activity between hippocampal and cortical
networks. 1: Interburst intervals are less variable in spike trains from hippocampal
cultures. 2: Hippocampal networks have higher correlations. 3: Hippocampal
networks generate more robust theta bursting patterns. Machine learning
techniques confirmed that these differences in patterning are sufficient to reliably
classify recordings at any given age as either hippocampal or cortical networks.
Conclusions: Although cultured networks of hippocampal and cortical networks
both generate spontaneous activity that changes over time, at any given time we
can reliably detect differences in the activity patterns. We anticipate that this
quantitative framework could have applications in many areas, including
neurotoxicity testing and for characterising phenotype of different mutant mice.
All code and data relating to this report are freely available for others to use.
Keywords: multielectrode arrays; spontaneous activity; cortex; hippocampus;
principal components analysis; support vector machines; classification trees
Introduction
During development, many parts of the nervous system generate patterns of spon-
taneous activity. These patterns of activity are thought to be instructive in the as-
sembly of neural connectivity, for example by driving activity-dependent mech-
anisms [1]. To date, most recordings of spontaneous activity have been in vitro,
although recent in vivo studies using calcium imaging also report the presence of
patterned spontaneous activity [2, 3]. In vitro recordings are typically made with
multielectrode arrays (MEAs) which contain at least 60 electrodes. These record-
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 2 of 15
ings allow us to assess activity at a range of levels from the single-unit to the net-
work. Beyond its relevance for understanding how activity might guide develop-
ment of the nervous system, spontaneous activity recordings have also been used
as an assay for network performance in applied settings, like neurotoxicity screen-
ing [4].
In recent years there has been significant interest in measuring the developmen-
tal patterns of spontaneous activity in networks cultured from neurons in control
and experimental conditions [4, 5, 6, 7, 8]. Although many properties of sponta-
neous activity have been reported, we do not yet have a systematic sense of how
these features change across development, or which features of neural activity are
useful at describing the observed patterns of activity.
To address both these questions, we have cultured two types of network on
MEAs and recorded their activity every 2–3 days up to around one month post-
plating of neurons onto the array. In the first type of network, we cultured hip-
pocampal neurons taken from embryonic mice. The second type of network was
created using exactly the same protocol except with neurons dissected from cortex.
Recordings of spontaneous activity from both types of network were quantified
using eleven different features at the level of individual electrodes, pairs of elec-
trodes, or the entire array. We found that hippocampal networks tend to generate
more regular bursting activity, including theta bursts, and more correlated activity
than the corresponding cortical networks at the same age.
Results
Development of spontaneous activity
Within seven days of culturing neurons on MEAs, spontaneous activity can be
reliably recorded (Fig. 1) from both hippocampal and cortical networks. As devel-
opment progresses, we find that the firing rate increases, and that the frequency
of bursting increases. To quantify these differences, we have used a range of mea-
sures (Fig. 2) to assess the activity at a single-electrode level, pairwise, and at the
level of the entire network. All of these measures are defined in the methods.
Overall firing rates During development, there are slight, statistically-significant
differences in firing rates, with median firing rates being slightly higher for hip-
pocampal networks, but overall there are no key differences at maturity (Fig. 3A).
Bursting properties Neurons typically fire in bursts, and are thought to be a reli-
able unit of neuronal information for functions such as coincidence detection and
synaptic modification [9]. We find that bursts emerge around 7 DIV (days in vitro)
and strengthen until about 14 DIV after which the bursting properties tend to sta-
bilise. Among the bursting properties that we have measured, two factors seem
to differentiate hippocampal and cortical networks. First, there is a higher fraction
of spikes occurring within bursts for hippocampal networks (Fig. 3E), although
the difference is no longer significant by 28 DIV. Second, the spike trains from hip-
pocampal networks seem to be more regular, as indicated by the lower coefficient
of variation for interburst intervals (CV of IBI) (Fig. 3F). The other burst-based
measures that we calculated, namely within-burst firing rate (Fig. 3B), burst rate
(Fig. 3C) and duration (Fig. 3D) show weaker differences between the two types
of network.
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 3 of 15
Network activity The previous measures analysed spiking data independently on
each electrode. As a first approximation to assessing network activity, we used
the concept of “network spikes” [10] to define the degree to which activity is co-
ordinated across the entire array. At any time twe count the number of active
electrodes; when this count exceeds a threshold, we say that a network spike has
occurred. We measured three properties of these network spikes: their rate (per
minute), their duration and their peak amplitude. Hippocampal networks tend to
have more network spikes than cortical networks (Fig. 3G) and the network spikes
involve more electrodes across development (Fig. 3H). The hippocampal network
spikes tend to last slightly longer, although this is not consistent across develop-
ment (Fig. 3I). Overall, this suggests that network activity tends to be stronger and
more coordinated in hippocampal than cortical networks.
Pairwise correlations As a further method to detect coincident activity on elec-
trodes, we computed correlation coefficients for all possible pairs of electrodes on
the array. For any pair of spike trains, we computed the spike time tiling coeffi-
cient, as this measure is particularly well-suited for relatively sparse spike trains
[11]. For N(typically 59) electrodes on the array we compute N(N−1)/2 correla-
tion coefficients (i.e. ignoring autocorrelations) and plot them as a function of the
distance separating the two electrodes (Fig. 2B). This technique has been used in
studies of spontaneous activity in developing retina, and often reveals that corre-
lations are distance-dependent, typically following a decaying-exponential profile
[12]. However, we found that there is little, if any, distance-dependence upon the
correlation coefficients (Fig. 2B), similar to that reported before [13]. We therefore
decided to compute the average of all pairwise correlations.
Across all developmental ages, we find that the mean correlation is higher in
hippocampal than in cortical networks (Fig. 3J). From 7 DIV to 14 DIV we see that
the mean correlation becomes reliably stronger; after 14DIV the correlations tend
to stabilise.
Presence of theta bursting The theta rhythm is a prominent 4–10Hz oscillation mea-
sured in the hippocampus, and is thought to be involved in a range of neural func-
tions [14]. We decided to examine whether our networks exhibited such oscilla-
tions by checking for peaks in the log interspike interval histogram in the range
0.1–0.25 s. Figure 2B shows an example of one electrode (recording from a 14DIV
hippocampal network) that exhibited theta bursting. Our approach was to mea-
sure the fraction of electrodes exhibiting theta bursting. Perhaps the most striking
feature that discriminates hippocampal from cortical networks is the presence of
theta bursting in hippocampal networks. Although only about 10% of electrodes
in hippocampal networks are classified as theta bursting at 7 DIV, it is after 11 DIV
that theta bursting is found on 50–75% of electrodes (Fig. 3K). By contrast, most
electrodes in cortical networks do not detect theta bursting, except at 25 DIV.
Discrimination of hippocampal and cortical networks
Each of the eleven features documented in Figure 3 shows that there are signifi-
cant differences between hippocampal and cortical networks. However, given that
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 4 of 15
the distributions of values can overlap and yet still be statistically significant (e.g.
firing rate at 21 DIV; Fig. 3A), we cannot use individual features to reliably dis-
criminate between the two types of network. We therefore used machine learning
techniques to address two related questions:
1 Out of the eleven features, which are the most important for discriminating
hippocampal versus cortical networks?
2 Given a recording of a network at a given age, is it possible to predict the
identity (hippocampal or cortical) of the network?
We therefore translated each fifteen minute recording into an eleven-dimensional
vector, with element iof the vector storing the value of feature imeasured from the
recording. This vector is described below as a feature vector of the recording.
Principal Components Analysis If there is a consistent difference in the properties of
hippocampal and cortical recordings, we would hope that the corresponding fea-
ture vectors cluster into two distinct regions. However, as these feature vectors are
eleven-dimensional, we must first reduce their dimensionality to visualise them.
Many such dimensionality-reduction techniques are available; we chose to use the
best-known method, principal components analysis. Figure 4 shows the projection
of the feature vectors at three different ages down into two-dimensional space. At
7 DIV there is significant overlap between the hippocampal and cortical record-
ings, which might suggest that is hard to discriminate between the two types of
recordings; however at 14 and 21 DIV the recordings from the same cell type clus-
ter and there is significant separation of the hippocampal and cortical recordings.
The graphs underneath each scatter plot show the cumulative percentage of vari-
ance accounted for by the components. In each case, the first two principal com-
ponents account for at least 60% of the variance.
Classification of recordings The principal components analysis suggests that, espe-
cially at the latter ages, the feature vectors contain sufficient signal to discriminate
between hippocampal and cortical networks. However, given the overlap between
clusters, we next used classification techniques to quantify the degree to which the
two classes of recording can be separated. We used two classification methods, de-
tailed below. In both methods, 2/3 of the feature vectors at a given age are used to
train a classifier to discriminate between the two types of recording. The remain-
ing 1/3 of the feature vectors are then used as a test set to evaluate how well the
classifier performs on data unseen during training.
We first used classification trees [15]. We built ten classifiers, one per age stud-
ied, to test whether the recordings could be grouped into cortical or hippocampal
networks. We found that for any given age, the prediction accuracy of the trees
was high — usually over 75% correct, depending on the age of the recording. (Per-
formance would be 50% if there were no information to distinguish the two types
of recording.) Once these trees had been built, we were able to interrogate them to
find out which features were dominant in driving the classification of the network.
At different ages, unsurprisingly, different features were dominant, but an overall
trend clearly emerged when we averaged across developmental ages. Table 1 lists
the features in decreasing order of their importance, along with their relative score
(column 2).
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 5 of 15
Out of the eleven features, three stood out. The most important was CV of IBI;
this is a measure of firing regularity, which tends to be higher in hippocampal
networks. Second, theta bursting is a key indicator (once it emerges at DIV 11) of
hippocampal networks. Third, mean correlation is one of the top three features
roughly half the time, and again is higher in hippocampal networks.
We chose classification trees as our first classification method primarily because
of their simplicity (there are no free parameters) and ability to easily assess which
features are driving the classification. Although its classification performance was
good, we compared its performance against another common machine learning
technique, namely Support Vector Machines (SVMs). We found that the SVM clas-
sifiers tended to result in slightly higher classification accuracy than the classifi-
cation trees; e.g. when all 11 features were used, performance was 75–97% across
ages (bottom row of Table 1).
Finally, given that our classifier trees provide us with a natural ordering of the
importance of features, we asked how performance varied as we reduced the num-
ber of features that each recording is represented by. We found that performance
remained high even as the number of features was gradually reduced (moving up
through the rows of Table 1). It is clear however that multiple features are required
for good classification; when only the single-most important feature is used (top
row of Table 1), performance was only just above chance at some ages. However,
with only three or four features, we obtained good performance across all ages.
In conclusion, the results from the classifiers tell us that three features of network
activity (CV of IBI, theta bursting, and mean correlation) are strong predictors of
whether a recording is from a hippocampal or a cortical network.
Discussion
We have found that cultured networks of either hippocampal or cortical neurons
generate spontaneous activity. These patterns of activity change during develop-
ment and already by 7 DIV significant differences in their activity patterns begin
to emerge. We have developed a quantitative framework for examining these ac-
tivity patterns. By calculating eleven features of activity patterns, we can represent
each recording of spontaneous activity as a point in (eleven-dimensional) feature
space. When we examine recordings from any one given developmental age, we
find that recordings from the same neuronal type (cortical or hippocampal) cluster
in this feature space such that we can reliably discriminate between hippocampal
and cortical networks.
Furthermore, out of the eleven features, we find that three are critically impor-
tant for this classification in feature space. First, the CV of IBI was most impor-
tant on average in driving the classification. Hippocampal spike trains tend to
fire in bursts that are more regularly spaced than spike trains from cortical neu-
rons. Second, after about 11 DIV most electrodes in hippocampal recordings detect
theta-bursting, compared to a minority in cortical recordings. Third, the mean cor-
relation between pairs of electrodes tends to be higher in hippocampal networks.
These three measures are all relatively simple, and measure activity on either a
single-electrode level or from pairs of electrodes. By contrast, although significant
differences were found in the network spike measures, most importantly, the peak
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 6 of 15
of network spikes, (Fig. 3F), these measures were not deemed to be critical in clas-
sification.
We have deliberately chosen simple features to characterise spiking activity to
see if they suffice to discriminate between cortical and hippocampal networks. It is
entirely likely that other more complicated measures of activity, particularly at the
network-level, may also reveal clear differences between these two types of net-
work [7]. For example, network connectivity measures have been used to explore
differences in spontaneous activity patterns between mature hippocampal and cor-
tical networks in a range of frequency bands [16]. However, the simple measures
we have chosen here suffice to reliably differentiate the two types of network. Like-
wise, even higher classification performance may be possible with more complex
machine learning techniques. However, our primary interest was to see whether
in principle the feature space can be reliably separated with standard approaches
[15]. Similar machine learning methods are not yet routinely used in analysing
spontaneous activity, although see [17] for a recent example showing how single-
cell activity could be classified as either in vivo or in vitro. Finally, with the advent
of a new generation of higher density MEAs containing up 4096 electrodes [18], it
is likely that there are much richer patterns of activity than we describe here.
We believe that our framework lends itself nicely to many applications, for exam-
ple in neurotoxicity testing where spontaneous activity from a network is recorded
whilst it is exposed to a particular compound [4]. By building up a representative
feature space of recordings from compounds known to be either toxic or safe, our
approach can be used to predict the toxicity of novel compounds. This idea builds
upon earlier work where mean profiles of activity in each condition were used as
simple classifiers [19]. More recently, SVMs were used for toxicity prediction [20].
We imagine that our approach could also be used to detect the impact of particu-
lar genetic mutations, given earlier work suggesting that there may be significant
differences in network activity [21].
Conclusions
We report key differences in the developmental spontaneous activity patterns of
cultured networks of hippocampal and cortical neurons. We have proposed a
quantitative framework for evaluating these patterns. Our database of recordings
and computer programs are all freely available for others to build upon. Future
work in this area could be to dissect the cellular or network mechanisms driv-
ing the differences in network activity. For example, the differences between the
cortical and hippocampus cultures could reflect molecular differences in cells or
synapse or cellular differences in the populations of cells. Alternatively, differ-
ences in functional connectivity might partially account for these results [7]. Dis-
secting these differences will require a detailed understanding on the diversity of
cell types defined by single cell transcriptomes in these brain regions, which is still
lacking.
Materials and Methods
Primary neuronal culture
Primary cultures of dissociated hippocampal and cortical neurons were prepared
from embryonic day (E) 17–18 mice. Hippocampi / cortices were dissected from
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 7 of 15
E17.5 mouse embryos (2–4, pooled) and transferred to papain (10 units/mL Wor-
thington, Lakewood, NJ) for 22 min at 37 ◦C. Cells were manually dispersed in
Dulbecco’s Modified Eagle’s Medium containing 10% v/v foetal bovine serum and
centrifuged twice at 400 g for 3.5 min. The final pellet was resuspended in Neu-
robasal/B27 supplemented with 0.5 mM Gln (Invitrogen), and dissociated cells
(2 ×105per dish) were seeded in the centre of poly-D-lysine / laminin coated
multielectrode arrays (60MEA200/30-Ti, Multi Channel Systems, Reutlingen, Ger-
many) containing 600 µl full Neurobasal medium. Zero-evaporation lids [22] were
fitted and the MEAs housed in tissue culture incubators maintained humidified
at 37 ◦C and 5% CO2/ 95% air. Twenty-four hours post-plating, sample MEAs
were placed on an inverted microscope with heated stage (Axiovert 200; Zeiss)
and photographed through a 32x phase objective at five different fields of view.
These images were then analysed with CellProfiler [23] to quantify the neuronal
density over the electrode array, giving an average value of 1500 cells/mm2.
At 3–4 days in vitro, cultures were fed by replacing 200µl medium with pre-
warmed fresh full Neurobasal medium. Cultures were subsequently fed using the
same method after each recording, equating to a one-third medium change twice
per week.
All procedures were performed in accordance with the United Kingdom An-
imals (Scientific Procedures) Act 1986. The mouse line used in this study was
C57BL/6-Tyrc-Brd (C57; albino C57BL/6).
MEA recording
Multielectrode arrays and all data acquisition hardware and software were from
MultiChannel Systems (Reutlingen, Germany). Pairs of MEAs were interfaced
with duplex 60 channel amplifiers and 15 minute recordings of spontaneous ac-
tion potentials were made twice per week during the four weeks following plat-
ing. MEAs were heated and kept under a light flow of 5% CO2/ 95% air during
recordings. Signals were digitised with a 128-channel analogue/digital converter
card at a rate of 25 kHz and filtered (100 Hz High pass) to remove low frequency
events and baseline fluctuations. Action potentials were detected by crossing of
threshold set to a fixed level of -20 µV, which typically approximated to 6–8 stan-
dard deviations from the baseline noise level. Record samples (1ms pre- and 2 ms
post-crossing of threshold) confirmed the characteristic action potential waveform.
Application of tetrodotoxin (1 µM) totally abolished spiking activity, confirming
the absence of false positive event detection using these methods. Spikes were
not sorted to distinguish signals generated by individual neurons, so represent
multiunit activity. Action potential timestamps were extracted using batch scripts
written for NeuroExplorer [24] and analysed using software developed in the R
statistical programming environment to compute parameters that quantitatively
describe network activity. In total, 214 recordings were taken from 32 arrays of cul-
tured cortical neurons, and 329 recordings from 61 arrays of cultured hippocampal
neurons.
Data analysis
To summarise a fifteen minute recording of network activity, we computed the fol-
lowing features. As all recordings detected activity from multiple electrodes, we
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 8 of 15
calculated summary scalar values (termed the “array value” below) by summaris-
ing the information from multiple electrodes. In this way, each recording was then
represented as an eleven-dimensional vector.
1Firing rate. The mean firing rate of each electrode was calculated. The array
value was the median of all electrode firing rates.
2Within-burst firing rate. Bursts were detected independently on each elec-
trode using our implementation of the Max Interval method from Neuroex-
plorer [24]. The parameters for burst detection were: maximum beginning
ISI 0.1 s; maximum end ISI — 0.25 s; minimum interburst interval — 0.8 s;
minimum burst duration — 0.05 s; minimum number of spikes in a burst —
6. For each electrode we calculated the mean of the firing rate during each
burst. The array value was the median of the within-burst firing rates, ignor-
ing electrodes where no bursts were detected.
3Burst rate. For each electrode we calculated the number of bursts per minute.
The array value was as per feature 2.
4Burst duration. The electrode value was the mean duration of bursts on that
electrode. The array value was as per feature 2.
5Fraction of spikes in bursts. The electrode value was the total number of
spikes classified as belonging to a burst divided by the total number of spikes
on the electrode. The array value was as per feature 2.
6CV of IBI. The electrode value was the coefficient of variation (s.d. divided
by mean) of the interburst intervals. The array value was as per feature 2.
7Rate of network spikes. Network spikes were defined as the array-wide av-
erage population activity [10]. It is defined as dividing time into small bins
(here 3 ms) and counting the number of electrodes that generated at least one
action potential in that bin. A network spike is then defined as the period of
time when more than a threshold (here n=10) electrodes are simultaneously
active. The array value was the number of network spikes per minute.
8Network spike peak. During each network spike we found the maximum
number of active electrodes. The array value was the median of the values
from each network spike in a recording.
9Network spike duration. The duration of each network spike was the time
(in seconds) that the count of active electrodes exceeded the threshold value.
The array value was as per feature 8.
10 Mean correlation. Given two different spike trains from the recording, we
calculated the correlation between them using the spike time tiling coefficient
[11] with the coincidence window of ∆t=5 ms. (We also tried ∆t=50 ms
and 0.5 ms, but results were qualitatively similar.) The array value was the
mean of all distinct pairs of electrodes.
11 Fraction of electrodes exhibiting theta bursting. For each electrode, the log
interspike interval histogram was calculated and smoothed with the default
kernel density estimation routine in R. A spike train was classed as showing
theta bursting if a peak was present in the 4–10 Hz band of the histogram. The
array value was the fraction of electrodes on the array that were classified as
theta bursting.
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 9 of 15
The Mann-Whitney test was used to test whether the median array values at
any given developmental age differed between the hippocampal and cortical net-
works. The p values were then corrected for multiple comparisons using the false
discovery rate method [25].
Clustering and classification
Standard principal components analysis was performed (with variance normal-
isation for each feature) for all feature vectors of any given age. Two standard
machine learning classifiers were tested: classification trees with boosting (“ran-
dom forests”) and support vector machines (SVMs) using radial kernel functions
with γ=1/11 [15]. For each age, we built binary classifiers to predict the region
(CTX/HPC) based upon the eleven features measured from each recording. For
both classifiers, we used 2/3 of the recordings as training data, with the remaining
1/3 used as test data. Performance is reported as mean percentage of correct clas-
sifications, averaged over 500 repeats using different splits of the data into train-
ing and test sets. The classification tree approach allows us to assess the relative
importance of features by measuring the degree to which they decrease the Gini
index [15, p319]. These values were normalised to the value of the top-performing
feature.
Data and code availability
Statistical analysis was performed in the R programming environment using the
SJEMEA package [26]. Data files containing the spike times from all recordings
analysed here were stored in the HDF5 format using the framework created for
spontaneous activity in retinal recordings [27]. The only addition to the framework
was a new metadata item /meta/region containing either the phrase “CTX” or
“HPC” depending on the network type. All data files and analysis code relating to
this paper are freely available at http://github.com/sje30/g2chvc. This includes
all the material required to regenerate the figures and Table in this article.
Abbreviations
CTX cortex
DIV days in vitro
HPC hippocampus
MEA multielectrode array
SVM support vector machine
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
PC, AM and SGNG conceived and designed the project. PC and AM performed all the experiments. EC and SJE
provided analysis tools. PC, EC and SJE analysed the data. SJE drafted the manuscript; PC, EC, SGNG and SJE
edited the manuscript. All authors read and approved the final version of the manuscript.
Acknowledgements
Thanks to Diana Hall, Johannes Hjorth and Ole Paulsen for comments on this work. PC and AM were supported by
the Wellcome Trust Genes to Cognition programme. PC received additional support from the BBSRC
(BB/H008608/1). EC was supported by a Wellcome Trust PhD studentship and Cambridge Biomedical Research
Centre studentship. SJE was supported by an and EPSRC grant (EP/E002331/1).
Author details
1Genes to Cognition Programme, Wellcome Trust Sanger Institute, Genome Campus, CB10 1SA Hinxton, UK.
2Current address: Department of Physiology, Development and Neuroscience, Physiological Laboratory, Downing
Street, CB2 3EG Cambridge, UK. 3Cambridge Computational Biology Institute, University of Cambridge,
Wilberforce Road, CB3 0WA Cambridge, UK. 4Current address: Centre for Integrative Physiology, University of
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 10 of 15
Edinburgh, School of Biomedical Sciences, EH8 9XD Edinburgh, UK. 5Centre for Clinical Brain Sciences and
Centre for Neuroregeneration, Chancellors Building, Edinburgh University, 49 Little France Crescent, EH16 4SB
Edinburgh, UK.
References
1. Blankenship, A.G., Feller, M.B.: Mechanisms underlying spontaneous patterned activity in developing neural
circuits. Nat. Rev. Neurosci. 11, 18–29 (2010). doi:10.1038/nrn2759
2. Siegel, F., Heimel, J.A., Peters, J., Lohmann, C.: Peripheral and central inputs shape network dynamics in the
developing visual cortex in vivo. Curr. Biol. 22, 253–258 (2012). doi:10.1016/j.cub.2011.12.026
3. Ackman, J.B., Burbridge, T.J., Crair, M.C.: Retinal waves coordinate patterned activity throughout the
developing visual system. Nature 490, 219–225 (2012). doi:10.1038/nature11529
4. Johnstone, A.F.M., Gross, G.W., Weiss, D.G., Schroeder, O.H.-U., Gramowski, A., Shafer, T.J.: Microelectrode
arrays: a physiologically based neurotoxicity testing platform for the 21st century. Neurotoxicology 31, 331–350
(2010). doi:10.1016/j.neuro.2010.04.001
5. Wagenaar, D.A., Pine, J., Potter, S.M.: An extremely rich repertoire of bursting patterns during the
development of cortical cultures. BMC Neurosci. 7, 11 (2006). doi:10.1186/1471-2202-7-11
6. Rolston, J.D., Wagenaar, D.A., Potter, S.M.: Precisely timed spatiotemporal patterns of neural activity in
dissociated cortical cultures. Neuroscience 148, 294–303 (2007). doi:10.1016/j.neuroscience.2007.05.025
7. Soriano, J., Rodr´ıguez Mart´ınez, M., Tlusty, T., Moses, E.: Development of input connections in neural
cultures. Proc. Natl. Acad. Sci. U. S. A. 105, 13758–13763 (2008). doi:10.1073/pnas.0707492105
8. Gullo, F., Manfredi, I., Lecchi, M., Casari, G., Wanke, E., Becchetti, A.: Multi-electrode array study of neuronal
cultures expressing nicotinic β2-V287L subunits, linked to autosomal dominant nocturnal frontal lobe epilepsy.
an in vitro model of spontaneous epilepsy. Front. Neural Circuits 8, 87 (2014). doi:10.3389/fncir.2014.00087
9. Lisman, J.E.: Bursts as a unit of neural information: making unreliable synapses reliable. Trends Neurosci. 20,
38–43 (1997). doi:10.1016/S0166-2236(96)10070-9
10. Eytan, D., Marom, S.: Dynamics and effective topology underlying synchronization in networks of cortical
neurons. J. Neurosci. 26, 8465–8476 (2006). doi:10.1523/JNEUROSCI.1627-06.2006
11. Cutts, C.S., Eglen, S.J.: Detecting pairwise correlations in spike trains: an objective comparison of methods and
application to the study of retinal waves. J. Neurosci. (2014). doi:10.1101/006635. (in press)
12. Wong, R.O., Meister, M., Shatz, C.J.: Transient period of correlated bursting activity during development of
the mammalian retina. Neuron 11, 923–938 (1993). doi:10.1016/0896-6273(93)90122-8
13. Godfrey, K.B., Eglen, S.J.: Theoretical models of spontaneous activity generation and propagation in the
developing retina. Mol. Biosyst. 5, 1527–1535 (2009). doi:10.1039/b907213f
14. Buzs´aki, G.: Theta oscillations in the hippocampus. Neuron 33, 325–340 (2002)
15. James, G., Witten, D., Hastie, T., Tibshirani, R.: An Introduction to Statistical Learning: with Applications in
R. Springer, New York (2014)
16. Ito, S., Yeh, F.-C., Hiolski, E., Rydygier, P., Gunning, D.E., Hottowy, P., Timme, N., Litke, A.M., Beggs, J.M.:
Large-scale, high-resolution multielectrode-array recording depicts functional network differences of cortical and
hippocampal cultures. PLoS One 9, 105324 (2014). doi:10.1371/journal.pone.0105324
17. Okamoto, K., Ishikawa, T., Abe, R., Ishikawa, D., Kobayashi, C., Mizunuma, M., Norimoto, H., Matsuki, N.,
Ikegaya, Y.: Ex vivo cultured neuronal networks emit in vivo-like spontaneous activity. J. Physiol. Sci. (2014).
doi:10.1007/s12576-014-0337-4
18. Maccione, A., Hennig, M.H., Gandolfo, M., Muthmann, O., van Coppenhagen, J., Eglen, S.J., Berdondini, L.,
Sernagor, E.: Following the ontogeny of retinal waves: pan-retinal recordings of population dynamics in the
neonatal mouse. J. Physiol. 592, 1545–1563 (2013). doi:10.1113/jphysiol.2013.262840
19. Gramowski, A., J¨ugelt, K., Weiss, D.G., Gross, G.W.: Substance identification by quantitative characterization
of oscillatory activity in murine spinal cord networks on microelectrode arrays. Eur. J. Neurosci. 19, 2815–2825
(2004). doi:10.1111/j.0953-816X.2004.03373.x
20. Mack, C.M., Lin, B.J., Turner, J.D., Johnstone, A.F.M., Burgoon, L.D., Shafer, T.J.: Burst and principal
components analyses of MEA data for 16 chemicals describe at least three effects classes. Neurotoxicology 40,
75–85 (2014). doi:10.1016/j.neuro.2013.11.008
21. MacLaren, E.J., Charlesworth, P., Coba, M.P., Grant, S.G.N.: Knockdown of mental disorder susceptibility
genes disrupts neuronal network physiology in vitro. Mol. Cell. Neurosci. 47, 93–99 (2011).
doi:10.1016/j.mcn.2010.12.014
22. Potter, S.M., DeMarse, T.B.: A new approach to neural cell culture for long-term studies. J. Neurosci. Methods
110, 17–24 (2001)
23. Carpenter, A.E., Jones, T.R., Lamprecht, M.R., Clarke, C., Kang, I.H., Friman, O., Guertin, D.A., Chang, J.H.,
Lindquist, R.A., Moffat, J., Golland, P., Sabatini, D.M.: CellProfiler: image analysis software for identifying and
quantifying cell phenotypes. Genome Biol. 7, 100 (2006). doi:10.1186/gb-2006-7-10-r100
24. Nex Technologies: NeuroExplorer Manual. (2012).
http://www.neuroexplorer.com/downloads/NeuroExplorerManual.pdf
25. Benjamini, Y., Hochberg, Y.: Controlling the false discovery rate: a practical and powerful approach to multiple
testing. J. R. Stat. Soc. Series B Stat. Methodol. 57, 289–300 (1995). doi:10.2307/2346101
26. R Core Team: R: A Language and Environment for Statistical Computing. Vienna, Austria (2014)
27. Eglen, S.J., Weeks, M., Jessop, M., Simonotto, J., Jackson, T., Sernagor, E.: A data repository and analysis
framework for spontaneous neural activity recordings in developing retina. Gigascience 3, 3 (2014).
doi:10.1186/2047-217X-3-3
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 11 of 15
DIV 7
HPC
DIV 14 DIV 21
CTX
Figure 1 Examples of spontaneous activity in developing hippocampal (HPC; top row) and
cortical (CTX; bottom row) cultures. Each column represents one day in vitro (DIV). Within each
raster plot, one row represents the spike train from one electrode; six (out of typically 59)
electrodes are shown. Scale bar for all rasters is 10s.
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 12 of 15
time (s)
#active units
−0.3 0.0 0.3
0
10
20
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●●
●
●●●
●●
●●●
●
●
●●
●
●●●
●●● ●●●
●
●
●●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●●
●
●●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●●●
●
●
●●
●
●
●●
●
●
●
●
●
●●●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●●●
●
●
●
●●
●●●
●
●●
●●
●
●
●
●●●
●●
●
●
●
●
●
●●●
●
●
●
●●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●●●
●
●
●●
●
●
●●
●●
●
●●●
●
●●●
●
●
●
●●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●●
●
●
●●
●
●●●
●
●
●●●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●●
●
●●●
●
●
●
●●
●●●
●
●●
●●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●●●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●●●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
intercell distance (µm)
correlation
0 800 1600
0.0
0.5
1.0
frequency (Hz)
density
0.0
0.1
0.2
0.3
1000 100 10 1 0.1 0.01
●
●
●
C
B
A
Figure 2 Examples of features calculated for each recording. The hippocampal recording from
14 DIV in Figure 1 was used as an example for this figure. A: mean network spike. B: pairwise
correlation calculated using the spike time tiling coefficient. As there is weak dependence on
distance, we take the mean (grey solid line). C: detection of theta bursting on an electrode with a
firing rate close to the median activity on the array.
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 13 of 15
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
(4 outliers)
0.0
2.5
5.0
7.5
10.0
7
** 10
** 11
* 14
* 17 18
* 21
* 24 25 28
DIV
firing rate (Hz)
●
●
●
●
●
●
●
●
●
●
●
●
●
(9 outliers)
0
25
50
75
100
125
7
** 10
** 11 14 17 18
** 21 24 25 28
DIV
w/in burst firing rate (Hz)
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
0
5
10
15
20
7
** 10
** 11 14
* 17
* 18 21 24 25 28
DIV
burst rate (/min)
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
0
1
2
3
4
5
7
** 10
** 11 14
** 17
* 18
** 21
* 24 25 28
DIV
burst duration (s)
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0
25
50
75
100
7
** 10
** 11
* 14
** 17
** 18
** 21
** 24
* 25 28
DIV
% spikes in burst
●●
●
●● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
(6 outliers)
0.0
0.5
1.0
1.5
7 10
** 11
** 14
** 17
** 18 21
* 24
** 25 28
*
DIV
CV of IBI
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
(3 outliers)
0
50
100
150
7
** 10
** 11
** 14
* 17 18
** 21
** 24 25 28
DIV
network spike rate (/min)
●●●
●
●
●
●●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
0
10
20
30
40
50
7
** 10
** 11
** 14
** 17
** 18
** 21
** 24
** 25
** 28
**
DIV
network spike peak
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●
0.00
0.01
0.02
0.03
0.04
0.05
7
** 10
** 11 14 17
** 18 21
** 24
* 25 28
*
DIV
network spike duration (s)
●
●
●
●
●
●
●
●
●
0.0
0.2
0.4
0.6
0.8
7
** 10
** 11
** 14
** 17
** 18
** 21
** 24
** 25
** 28
**
DIV
mean correlation
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0
25
50
75
100
7 10 11
** 14
** 17
** 18
** 21
** 24
** 25
* 28
**
DIV
% theta bursting
region ctx hpc
A B C
D E F
G H I
J K L
ctx
hpc
7
32
57
10
19
17
11
13
34
14
32
49
17
19
17
18
13
34
21
32
55
24
19
17
25
13
33
28
22
16
#arrays per age
Figure 3 Characterisation of spontaneous activity in hippocampal and cortical networks. Panels
A–K shows the values of one feature (named on the y axis) as a function of age. Boxplots show
the median and interquartile range, with whiskers extending out the most extreme values within
1.5 times the interquartile range. Individual points outside this range are regarded as outliers and
drawn as points; in a few cases these outliers are not drawn to keep the y-axis within a meaningful
range. Underneath each age, stars denote significant difference of median values for cortical and
hippocampal networks at either 0.05 (*) or 0.01 (**) level (Mann-Whitney test, with p values
corrected for multiple comparisons with false discovery rate method). L: number of arrays
analysed at each age.
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 14 of 15
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●●●
●●●
●
●●
●●●●●
●●
●●
●
●
●
●●
●●
●
DIV 7
●
●
CTX
HPC
PC2
●
●
●
●●●●●●●●
cumulative fraction of variance
1 11
0.0
0.5
1.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
DIV 14
PC1
●
●
●
●●●●●●●●
principal component
1 11
0.0
0.5
1.0
●●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●●
●●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
DIV 21
●
●
●
●
●●●●●●●
1 11
0.0
0.5
1.0
Figure 4 Principal components analysis of hippocampal and cortical feature vectors. Each column
represents principal components analysis of the 11-dimensional feature vectors of all recordings at
a given age (days in vitro). In the scatter plot, each point represents one recording projected down
into the two dimensions that account for maximal variance and is coloured according to its cell
type. Each graph shows the cumulative fraction of variance accounted for by the principal
components.
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
Charlesworth et al. Page 15 of 15
Table 1 Classifier performance at discriminating cortical from hippocampal cultures. Features are
listed in decreasing order of importance (score; column 2) normalised to the top score. The following
numbers in each row i=1 . . . 11 are the mean percentage of correct classifications at each age using
the top ifeatures.
Percentage correct at given age
Feature Score 7 10 11 14 17 18 21 24 25 28
CV of IBI 1.00 64 85 88 89 84 71 61 72 73 60
theta burst 0.84 64 82 96 88 91 89 84 86 74 63
mean correlation 0.50 67 89 96 91 86 94 89 84 75 75
burst duration 0.33 72 86 97 91 92 90 89 82 77 83
burst rate 0.32 78 84 97 92 90 95 89 91 74 78
% of spikes in bursts 0.25 82 86 92 94 92 94 89 87 77 79
firing rate 0.21 83 84 94 94 90 94 92 89 74 78
NS peak 0.19 83 86 91 95 91 96 91 88 76 79
NS duration 0.19 83 89 90 92 89 93 93 86 72 78
w/in burst firing rate 0.16 83 89 96 93 92 96 93 87 77 79
NS rate 0.11 83 88 92 92 90 97 93 85 76 75
.CC-BY 4.0 International licensea
certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under
The copyright holder for this preprint (which was notthis version posted October 1, 2014. ; https://doi.org/10.1101/009845doi: bioRxiv preprint
