You are on page 1of 12

HIPPOCAMPUS 00:112 (2014)

Intrinsic Rescaling of Granule Cells Restores Pattern Separation Ability


of a Dentate Gyrus Network Model During Epileptic Hyperexcitability
Man Yi Yim,1 Alexander Hanuschkin,2 and Jakob Wolfart3*

ABSTRACT: The dentate gyrus (DG) is thought to enable efficient hippocampal memory acquisition via pattern separation. With patterns
defined as spatiotemporally distributed action potential sequences, the
principal DG output neurons (granule cells, GCs), presumably sparsen
and separate similar input patterns from the perforant path (PP). In
electrophysiological experiments, we have demonstrated that during
temporal lobe epilepsy (TLE), GCs downscale their excitability by transcriptional upregulation of leak channels. Here we studied whether
this cell type-specific intrinsic plasticity is in a position to homeostatically adjust DG network function. We modified an established
conductance-based computer model of the DG network such that it
realizes a spatiotemporal pattern separation task, and quantified its performance with and without the experimentally constrained leaky GC
phenotype. Two proposed TLE seizure mechanisms were implemented
in various degrees and combinations: recurrent GC excitation via mossy
fiber sprouting and increased PP input. While increasing PP strength
degraded pattern separation only gradually, already the slight elevation
of sprouting drastically (non-linearly) impaired pattern separation. In
most tested hyperexcitable networks, leaky GCs ameliorated pattern
separation. However, in some sprouting situations with all-or-none seizure behavior, pattern separation was disabled with and without leaky
GCs. In the mild sprouting (and PP increase) region of non-linear
impairment, leaky GCs were particularly effective in restoring pattern
separation performance. These results are compatible with the hypothesis that the experimentally observed intrinsic rescaling of GCs serves to
C 2014 Wiley
maintain the physiological function of the DG network. V
Periodicals, Inc.
KEY WORDS:
hippocampus; computational model; epilepsy; ion
channel homeostasis; electrophysiology

INTRODUCTION
The dentate gyrus (DG) is involved in hippocampal memory acquisition and is important for the discrimination of similar sensory input
patterns (Gilbert et al., 2001; Leutgeb et al., 2007; McHugh et al.,

Department of Mathematics, University of Hong Kong, Hong Kong;


Institute of Neuroinformatics, University of Zurich and ETH Zurich,
Zurich, Switzerland; 3 Oscar Langendorff Institute of Physiology, University of Rostock, Rostock, Germany.
Grant sponsor: Deutsche Forschungsgemeinschaft (DFG); Grant number:
SFB780/C2, WO1563/1-1.
*Correspondence to: Dr. Jakob Wolfart, Oscar Langendorff Institute of
Physiology, University of Rostock, Gertrudenstr. 9, 18057 Rostock, Germany. E-mail: jakob.wolfart@uni-rostock.de
Received 16 May 2014; Revised 29 September 2014; Accepted for
publication 29 September 2014.
DOI 10.1002/hipo.22373
Published online 00 Month 2014 in Wiley Online Library
(wileyonlinelibrary.com).
2

C 2014 WILEY PERIODICALS, INC.


V

2007; Moser et al., 2008). This behavioral discrimination is consistent with theoretical models predicting a
role of the DG in computational orthogonalization of
overlapping activated neuronal assemblies, an operation termed pattern separation (Treves and Rolls,
1992; OReilly and McClelland, 1994; Santoro,
2013). The principal neurons of the DG, the granule
cells (GCs), are thought to contribute to DG function
via their low firing probability and connectivity (Jung
and McNaughton, 1993; Chawla et al., 2005; Leutgeb
et al., 2007; Moser et al., 2008). While connectivity
has received a lot of attention in DG network modeling (Santhakumar et al., 2005; Dyhrfjeld-Johnsen
et al., 2007; Morgan and Soltesz, 2008; de Almeida
et al., 2009; Myers and Scharfman, 2009, 2011;
Schneider et al., 2012), the impact of cell typespecific, intrinsic properties (such as membrane conductivity) on network performance has generally
received less attention. There are however two studies
on the influence of mossy cells (MCs) and interneurons on seizure spread in the DG (Howard et al.,
2007; Yu et al., 2013). The pattern separation ability
of previous DG computer models has been investigated via spatial binary patterns (Myers and Scharfman, 2009, 2011) but not via spatiotemporal spike
coding. Yet, spike timing is an important aspect of
the GC pattern transformation and neuronal information encoding in general (Lisman et al., 2005; Leutgeb et al., 2007; de Almeida et al., 2009; Buzsaki,
2010; Panzeri et al., 2010; Eichenbaum, 2013; Rangel
et al., 2013; Yim et al., 2014).
Temporal lobe epilepsy (TLE) is a common form
of epilepsy characterized by focal seizures in the
entorhinal-hippocampal loop and partial memory
impairments (Spencer, 2002; Bonilha et al., 2007).
Discussed seizure mechanisms include elevated output
from the entorhinal cortex to the DG via perforant
path (PP) axons (Spencer and Spencer, 1994; Kobayashi et al., 2003) and the backsprouting of mossy fiber
(MF) axons of GCs resulting in recurrent excitation
of GCs (Tauck and Nadler, 1985; Buckmaster, 2012).
The latter inspired modeling studies to demonstrate
that MF sprouting indeed produces seizure-like runaway excitation in DG networks (Santhakumar et al.,
2005; Morgan and Soltesz, 2008; Schneider et al.,
2012).
In surgically resected hippocampi of patients with
TLE, we discovered that GCs of the sclerotic focus

YIM ET AL

possess abnormal electrophysiological properties (Stegen et al.,


2009, 2012). Using a TLE animal model, we revealed that
upregulation of leak ion channel protein is responsible for
effective downscaling of intrinsic GC excitability in seizureaffected hippocampi (Young et al., 2009; Stegen et al., 2012).
Interestingly, GCs are not simply silenced; they maintain their
resting potential via combined regulation of depolarizing and
hyperpolarizing leak conductances (Young et al., 2009; Stegen
et al., 2012). It appears as if the additive/subtractive scaling of
the inputoutput curve of GCs via shunting channels (Wolfart
et al., 2005) serves to homeostatically maintain DG function
under epileptic conditions (Meier et al., in press).
The question addressed here is whether the experimentally
observed changes of GC properties are in a position to regulate
DG network performance with respect to its proposed pattern
separation function. To answer this question, we used an established conductance-based computer model of the DG network
(Santhakumar et al., 2005) and modified it such that it could
perform a spatiotemporal pattern separation task. The network
was rendered epileptic by MF sprouting and PP hyperactivity
which produced seizure-like activity and impaired pattern separation. When we implemented the leaky GCs (Young et al.,
2009), the pattern separation ability of the DG was ameliorated in almost all cases. Thus, the intrinsic plasticity of GCs is
able to not only enhance cell survival during pathological
hyperexcitation (Kirchheim et al., 2013; Meier et al., 2014)
but also maintain DG network function.

MATERIALS AND METHODS


Dentate Gyrus Network Model
Simulations were performed using the software NEURON
7.1 for Linux (http://www.neuron.yale.edu) and a HewlettPackard 8100 Elite personal computer with an Intel Core i7
860/2.8 GHz processor and 8 GB random access memory for
initial simulations. Bulk simulations and analyses were conducted using the high performance computing facilities offered
by the Information Technology Services of the University of
Hong Kong. Integration time step was 100 ls. We employed
the conductance-based network model of the DG (Santhakumar et al., 2005) as the basis for our work. Here we describe
mainly the changes we made to this model; all other parameters were left as in the original model, the detailed description
and justification of which can be found in the work by Santhakumar et al. (2005) and references therein, as well as in our
Appendix. Briefly, the DG model consisted of the following
cell types (see simplified overview cartoon in Fig. 1A): 500
excitatory GCs, 6 inhibitory basket cells (BCs), 15 excitatory
MCs, and 6 inhibitory hilar PP-associated (HIPP) cells (HCs).
The intrinsic properties of these neurons contained a variety of
voltage-dependent sodium, potassium, and calcium channels as
well as leak conductances. Of these properties we modified
only those of GCs. Specifically, we added the classic inward
Hippocampus

FIGURE 1.
DG network model performing pattern separation. A:
Control connectivity (CTRL DG) with PP input, output via 500 GCs
(control properties, CTRL GCs), BCs, HIPP cells (HCs), and MCs.
Vertical tick lines represent APs. B: Summation of active inputs to
GC8 (upper panel) during input pattern 1 (IP1) evoked one AP in
GC8 (lower panel). Inset: voltage responses to 2 pA steps (scale bars:
1 mV, 100 ms). C,D: The IP3 (C) led to a sparse output pattern (OP)
30 (D, all GCs and BCs are displayed; HCs and MCs were inactive).
E: Four of 13 IPs with increasing spatiotemporal difference and
respective OPs (y-scales contain all PPs and GCs). Right panel, similarity scores between IPs vs. those of OPs, fitted by a shifted power
law. Data below the dashed line signify pattern separation, e.g. at 0.60
input similarity, the CTRL DG achieved an output similarity of 0.29.

INTRINSIC RESCALING AND NEURONAL NETWORK PATTERN SEPARATION


rectifier potassium (Kir subtype Kir2) conductance and the
tonic GABAA chloride conductance as additional leak conductances, since we found these upregulated in GCs of the
seizure-experienced hippocampus (Young et al., 2009). As
reversal potential for the GABAA leak (ECl ) we employed
270 mV (Pathak et al., 2007). The leak conductances were
scaled to approximate the experimentally determined input
resistance (Rin) of normal GCs (Young et al., 2009) for the
control GC condition (CTRL GCs). In addition, we simulated
a leaky GCs condition in which GCs possessed fourfold
increased Kir and GABAA conductances such that Rin was
about 60%, similar to the mean values in epilepsy-adapted
GCs (Young et al., 2009).

Network Connectivity and Epileptic Dentate


Gyrus
We used the topographic DG model version of Santhakumar
et al. (2005) and realized the network connectivity and the IPs
with identical initial conditions (seeds). Connectivity seeds
were only changed for specifically mentioned controls. To
obtain sparse GC activation as required for pattern separation
and as observed in vivo (Jung and McNaughton, 1993), we
reduced the strength of PP input synapses to GCs (and BCs)
to 10% of the original model (i.e., 2 nS which is our 100%).
For the same reason and because of the known influence of
BCs on GCs (Lawrence and McBain, 2003), we increased the
GC-BC-GC inhibitory feedback loop threefold. In addition to
this control DG network (CTRL DG), several forms of epileptic hyperexcitability were simulated. Because backsprouting of
MF is a proposed TLE seizure mechanism and was already
present in the original model (Tauck and Nadler, 1985; Santhakumar et al., 2005; Morgan and Soltesz, 2008; Buckmaster,
2012), we first rendered the network epileptic via the addition
of recurrent connections of GCs. Also in the TLE animal
model of which we transferred the employed GC properties,
MF sprouting is present (Suzuki et al., 1997). Santhakumar
et al. (2005) defined 100 sprouted synapses (per receiving GC)
as 100% and examined the 050% range. Because the effect
on pattern separation saturated in our study, it was sufficient to
examine the sprouting regimes of 10, 15, 20, and 30% (Spr10,
Spr15, Spr20, Spr30 DG, respectively). In addition, based on
experimental findings that TLE is associated with increased
entorhinal input to the DG (Spencer, 2002; Kobayashi et al.,
2003; Bonilha et al., 2007), we increased PP synaptic strengths
up to 160% (PP160) of our CTRL DG condition. See Appendix for further details.

Pattern Separation Task and Analysis


To design a spatiotemporal pattern separation task for the
DG, we implemented patterns consisting of an activated cell
population and taking into account the relative spike timing.
Different patterns were presented via the PP input to the network and the similarity of the different input patterns (IPs)
and GC output patterns (OPs) was evaluated. Each IP consisted of 6 PP inputs to a subset of cells in a time window of

30 ms. Input pattern 1 (IP1) involved PP axons PP1 to PP6,


IP2 involved axons PP2 to PP7, and so on. A total of 13 IPs
were presented in turn to the network with decreasing overlap
from IP1 to IP7 based on the work by Myers and Scharfman
(2009). Each PP input line could transmit three excitatory
events evoking excitatory postsynaptic potentials (EPSPs) in
one time window, and IPs were independent of each other.
Action potential (AP) output events and PP input events in
GCs (3 ms before respective EPSPs to account for synaptic
delay) are displayed as vertical bars in figures.
The pattern separation ability was quantified as the amount
of similarity between two OPs compared to the similarity
between the two respective IPs. We chose a time window of
200 ms for comparison, corresponding roughly to the physiologically relevant theta frequency interval (Rangel et al., 2013).
Given two patterns, we first lowpass-filtered their constituent
spike trains with a triangular kernel of base length 20 ms, close
to the time window in which neurons integrate synaptic inputs
(Lisman et al., 2005; Buzsaki, 2010). Then we computed the
Pearson correlation coefficient of the filtered signals between
two simulations with different IPs for each PP input line (and
each GC output). The similarity score was defined as the average correlation coefficient over all PPs (and GCs) in the model.
Thus, two independent spike trains will give rise to a similarity
score of zero, whereas two identical spike trains will give a similarity score of one. Note that the scores calculated with this
method are corrected for rate dependence through baseline
subtraction and normalization and that for Poisson processes
the measure is strictly independent from rate (de la Rocha
et al., 2007). Similarity scores were plotted as output (y-axis)
vs. input (x-axis) and the resulting separation curve was fitted
with a shifted power law of the form y5a112a3x b . The
values of a and b were determined by the method of least
squares. Statistical differences between the similarity score distributions were compared via least square nonlinear regression
fits and extra sum-of-square F test using Prism (GraphPad, La
Jolla, USA). The Adobe Creative Suite 6 was used for figure
preparation.

RESULTS
A Dentate Gyrus Network Model Performing a
Temporal Pattern Separation Task
First, we presented the temporally distributed patterns via
PP inputs to a DG network with control connectivity and control GC properties (Fig. 1A, CTRL DG/CTRL GCs, see
Methods). In Figure 1B, the signal integration is displayed
from the perspective of one representative GC (GC8). This
GC received inputs from various sources (Fig. 1B, upper panel)
during input pattern 1 (IP1). The GC8 summated the excitatory inputs and eventually fired an AP (Fig. 1B, lower panel);
some inhibitory inputs from BCs followed. The necessity of
EPSP summation for AP generation is consistent with the
Hippocampus

YIM ET AL

situation in real GCs (McNaughton et al., 1981). Overall, the


IPs presented to the CTRL DG led to sparse activation of
GCs. For example, IP3 (Fig. 1C) led to the DG output pattern
30 (OP30 ) with only few APs in the entire cell population
(Fig. 1D; mean percentage of activated GCs: 10.2%). Sparse
activity is consistent with in vivo recordings of GCs (Jung and
McNaughton, 1993). Four of 13 IPs and respective OPs are
displayed in Figure 1E (left panels). The difference between
the IPs increased from IP1 to IP13, i.e. their similarity score
decreased. A scatter plot of output similarity scores vs. respective input similarity scores is shown in the right panel of Figure
1E. The fitted pattern separation curve is well below the
dashed diagonal line, suggesting that under control conditions,
this model of the DG performs pattern separation. In other
words, from IPs which were rather similar (x-axis, high similarity score), the DG produced OPs which were less similar
(y-axis, lower similarity score). To compare one value for this
ability: patterns with an input similarity of 0.60 were separated
to an output similarity of 0.29. Thus, consistent with the proposed function of the DG, our network model decorrelated
received patterns in space and time.

Mossy Fiber Sprouting Leads to Breakdown of


the Pattern Separation Ability
Next, we rendered the network structure epileptic via
recurrent MF sprouting, a presumed mechanism for epileptic
hyperexcitability already present in the original model (Santhakumar et al., 2005). Initially, we implemented 30% sprouting
(Spr30 DG) and applied the IPs (as before) to the epileptic
DG with CTRL GCs (Fig. 2A). As a result, the DG network
showed seizure-like hyperexcitation. All GCs were strongly activated and forced to fire bursts of APs (Figs. 2AC). Under
these conditions, the feedforward and feedback inhibition via
BCs, and delayed activation of HCs and MCs also became
more prominent (Figs. 2A,B). Because we used a topographic
model, seizure spread was also slightly topographic (Fig. 2C).
However, this was not as apparent as when seizures spread
from synchronous activation of 100 directly neighboring GCs
(Santhakumar et al., 2005). The seizure-like activity was not
self-sustaining and not terminated by a specific cell type, similar to mild sprouting examples in the work by Santhakumar
et al. (2005). Although in the seizure context, 30% sprouting
may be considered mild, the DG already completely failed in
the pattern separation task with this level of recurrent excitation: all output similarity scores were now above the dividing
line (Fig. 2C, right panel), i.e. in the area of pattern convergence (Santoro, 2013).

With Leaky Granule Cells and SproutingRelated All-or-None Seizures, Pattern Separation
Cannot be Computed
Electrophysiological experiments have revealed that during
chronic TLE, GCs display a permanent upregulation of leak
channels and specifically, increased Kir and tonic GABAA conductances were found responsible for a reduction of Rin and
Hippocampus

FIGURE 2.
MF sprouting leads to breakdown of DG pattern
separation ability. A: The DG model of Figure 1 was rendered epileptic by 30% GCGC backsprouting (Spr30 DG, see inset).
Intrinsic GC properties were unchanged (CTRL GCs). During
IP1, GC8 experienced massive excitatory input from other GCs
(upper panel) and was forced to burst (lower panel). B: The IP3
(see Fig. 1C) now led to repeated excitation of all GCs (OP30 ).
Panel B contains all cells, active and inactive (HC, HIPP cells 22
27, light gray; MC, mossy cells 721, dark gray; BC, basket cells
16, medium gray). C: Quantification of pattern separation with
IPs as in Figure 1E (y-scales in left panels correspond to y-axis in
D). The right graph shows the devastating effect of 30% sprouting: the DG was unable to perform pattern separation: at 0.60
input similarity, the output similarity was now 0.91.

excitability (Young et al., 2009). Are these changes in the position to affect the pattern separation ability of the DG? Constrained by our experimental results, we implemented these
leak mechanisms in the epileptic DG model described above
(Figs. 3A,B, Spr30 DG/Leak GCs). Consistent with the respective experiments (Young et al., 2009; Kirchheim et al., 2013),

INTRINSIC RESCALING AND NEURONAL NETWORK PATTERN SEPARATION

the leaky GCs had a similar resting potential (277.2 mV) as


the CTRL GCs (280.1 mV) but only about 60% of the Rin
(Fig. 3A, middle panel; leaky, gray traces, 264.3 MX; CTRL,
black traces, 429.9 MX) and a right shifted currentfrequency
(IF) curve compared to CTRL GCs (Fig. 3A, right panel).
The shifted IF curve was not changed in slope, consistent
with the upscaling of a static, pure shunting conductance
(Wolfart et al., 2005). As a result of these intrinsic changes,
the excitability of GCs and thus the DG network was strongly
reduced (Figs. 3B,C). However, the pattern processing abilities
of the DG were not restored at all under these conditions
(which turned out to be an exception). Some patterns still provoked runaway excitation as with CTRL GCs (Figs. 3C,D;
e.g., Pattern 4) while with other patterns, the GC population
remained silent (Fig. 3D; e.g., Patterns 13).
In Figure 3B, it is again shown from the perspective of
GC8; while IP2-evoking EPSPs were all subthreshold, IP4 triggered spiking in GC8, although the patterns were rather similar. With IP4, GC8 received even one PP input less. The
reason for the phenomenon can be estimated from the input
line of recurrently connected GCs (Fig. 3B, upper panel); while
IP2 did not achieve any GC activation, IP4 did at least once
and this was sufficient to eventually recruit all other GCs
(including GC8). As only one seizure was initiated and evolved
slowly, its topographical spread became very obvious. The slow
activity propagation was due to reduced size of EPSPs which
now required more events to summate to AP threshold (Fig.
3B, lower panel). A similar effect on seizure spread is observable when shifting the activation curve of mossy cell sodium
channels (Howard et al., 2007). Overall, the condition of 30%
MF sprouting was not suitable for the present study because
the bistable regime made calculation of pattern separation
impossible. To exclude that this network behavior was due to a
particular realization of our model connectivity, we tested
another initial condition (seed) for the randomization (with
identical structural statistics) but also these simulations resulted
in the all-or-none seizure behavior, except that instead of 4/13
patterns now 5/13 patterns evoked seizures and the rest produced silence (data not shown). Thus, there are conditions of
recurrent GC excitation in which the DG network is unable to
function as a pattern separatorwith or without leaky GCs.

FIGURE 3.
In some conditions, leaky GCs did not prevent
sprouting-related all-or-none seizures and failure of pattern separation.
A: The epileptic DG network of 30% sprouting as in Figure 2
(Spr30 DG) was used with leaky GCs (Leak GCs, gray). Middle panel:
smaller voltage responses of leaky GCs to 2 pA steps compared to
CTRL GCs (thin black traces; scale bars, 1 mV, 100 ms). Right panel:
the current (I)AP frequency (F) relation of leaky GCs was right
shifted. B: Leaky GC8 received two IPs (upper panel, IP2 and IP4). It
was activated by IP4 (lower panel) which recruited recurrent GC excitation (CCs row). C: The IP4 produced a slowly developing seizure
(panel contains all active and inactive cells). D: Since only few IPs
produced a GC output, no similarity scores could be computed (yscales contain all PPs and GCs). Hence, no meaningful pattern separation abilities remained under these conditions of recurrent excitation.

Epileptic Hyperexcitation via Increased Perforant


Path Activity Reduces the Pattern Separation
Ability of the Dentate Gyrus Gradually
It is controversial whether MF sprouting really functionally
results in intense recurrent GCGC excitation and whether it
is (alone) causal to TLE seizures (Scharfman et al., 2003;
Buckmaster, 2012; Heng et al., 2013). On the other hand,
there is good evidence that TLE is associated with an increased
PP input from entorhinal cortex layer II neurons to GCs
(Spencer, 2002; Kobayashi et al., 2003; Bonilha et al., 2007).
Therefore, we designed a mixed epileptic condition with PPGC input strength increased to 160% of our CTRL DG condition, plus 10% sprouting (Fig. 4A, PP160Spr10 DG). Under
Hippocampus

YIM ET AL

Leaky Granule Cells can Restore the Pattern


Separation Ability Lost by Epileptic
Hyperexcitation
Next, we combined our epileptic network of Figure 4 with
the leaky GCs introduced in Figure 3A for the simulations
shown in Figure 5 (Fig. 5A, PP160Spr10 DG/Leak GCs).
Under these conditions, the activity in the GC population was
very sparse (Figs. 5AC), i.e. similar to the CTRL DG

FIGURE 4.
Epileptic hyperexcitation via increased PP activity
gradually reduced the DG pattern separation ability. A,B: With
epilepsy-related increase of the PP input strength to 160% and
additional mild (10%) MF sprouting (PP160Spr10 DG), the network activity was overall enhanced and most GCs were repeatedly
activated (B, the panel contains all active and inactive cells. C:
Under these epileptic conditions, the DG possessed only weak pattern separation abilities (left panel y-scales correspond to y-axis in
B). At 0.60 input similarity, the output similarity was 0.36.

these conditions, with CTRL GCs, the DG cell population


was strongly activated and most GCs fired repeated APs (Figs.
4B,C). As aimed for, this epileptic network now showed more
mildly degraded pattern separation which was still measurable
(Fig. 4C). IPs with similarity of 0.60 were now separated to a
similarity of 0.36, i.e. pattern separation of this epileptic network was not as effective as the CTRL DG (Fig. 4C, right
panel, p < 0.0001). Thus, with an epilepsy network scenario of
increased entorhinal input combined with very mild MF
sprouting, DG-mediated pattern separation was degraded but
not abolished.
Hippocampus

FIGURE 5.
Leaky GCs can restore the pattern separation ability lost due to epileptic hyperexcitation. A,B: This simulation was
performed with the same epileptic DG network and IPs as in Figure 4 (PP160Spr10 DG), except that leaky GCs were implemented
(see Fig. 3A). The same PP inputs evoked only sparse activity in
the GC population (A, B). As in Figure 1D, HCs and MCs
remained silent. C: Left panels show representative patterns (yscales correspond to y-axis in B). The pattern separation abilities
of the DG were restored with leaky GCs (gray line), despite the
epileptic wiring and input (thin black line shows performance
with CTRL GCs as in Fig. 4C). At an input similarity of 0.60,
output similarity was now only 0.16.

INTRINSIC RESCALING AND NEURONAL NETWORK PATTERN SEPARATION

condition (Fig. 1D). Importantly, the pattern separation ability


was now fully restored despite the epileptic wiring and input
(Fig. 5C, right panel, thick gray line). Patterns with a similarity
of 0.60 were now separated by the DG network to a similarity
of 0.16, i.e. much improved compared to the DG with unleaky CTRL GCs (Fig. 5C, right panel, thin black line).
These simulations demonstrate that epilepsy-related plasticity
observed in GCs (Young et al., 2009) can indeed restore the
pattern separation ability of the DG.

Robust but Heterogeneous Amelioration of


Pattern Separation Ability by Leaky Granule
Cells in Different Conditions of Epileptic
Hyperexcitability
The precise degree of MF sprouting and increased PP input
during TLE is not known and it is likely to be variable from
patient to patient (see discussion). Therefore, we examined the
influence of these parameters systematically. Figure 6A shows
the pattern separation curves simulated as described before
(this time with a PP input strength of 140%) for 4 degrees of
MF sprouting (030%, panels Spr030). When comparing
pattern separation via CTRL GCs (Fig. 6A, gray symbols,
black fits) with pattern separation via leaky GCs (Fig. 6A, blue
symbols and fits, colors online), two effects were noted: first, in
all cases, leaky GCs ameliorated the pattern separation ability
of the DG; secondly, around 20% sprouting, leaky GCs had a
particularly high impact. The phenomenon can also be seen in
Figure 6B; there was a sudden rise of the output similarity
score (at an input similarity score of 0.6) between 10 and 20%
sprouting, i.e. at this point pattern separation drops nonlinearly (Fig. 6B, left panel, black line). To better judge the
slope change between 10 and 20% sprouting, we additionally
simulated the 15% condition.
Again, one could ask whether this network behavior was due
to an exceptional condition of our connectivity realization.
However, this was not the case as simulations with changed
initial randomization seed produced a similar non-linear rise
(Fig. 6B, error bars are STD for n 5 5 network realizations).
In contrast to sprouting, with increasing PP input strength, the
pattern separation ability deteriorated more linearly (Fig. 6B,
right panel, n 5 5 network realizations). In both epileptic situations, leaky GCs effectively increased pattern separation (Fig.
6B, blue lines). However, since with leaky GCs (and sprouting), the sharp similarity rise occurred later (at 30%), the
effectiveness to ameliorate pattern separation became very large
around 20% sprouting (Fig. 6B, left panel, compare black and
blue lines, colors online). We further combined the sprouting
and PP hyperexcitation mechanisms to variable degrees (Fig.
6C). The results confirm the general usefulness of leaky GCs
for pattern separation in a hyperexcited DG. With CTRL
(non-leaky) GCs, already mild epileptic hyperexcitation (e.g.,
20% sprouting) disabled the DG as a pattern separator (Fig.
6C, left panel, red colors). In contrast, with leaky GCs, pattern
separation can be performed well up to the condition of 30%
sprouting and 140% PP elevation (Fig. 6C, right panel).

FIGURE 6.
The amelioration of the pattern separation via
leaky GCs was robust but heterogeneous under different conditions
of epileptic hyperexcitation. A: With more sprouting (030%,
Spr030), pattern separation deteriorated (black fits, higher output
similarity scores). Leaky GCs reduced output similarity in all cases
(blue fits). B: Comparison of output similarity at 0.6 input similarity (y-axis) with different degrees of sprouting or PP strength
(CTRL GCs, black; leaky GCs, blue). Pattern separation failed
non-linearly with sprouting (left panel). Leaky GCs shifted the
curve, hence the large effect at 20% sprouting (see also A). Right
panel, with increased PP strength, pattern separation deteriorated
only gradually. Similar results were obtained with other network
realizations (error bars are STD, n 5 5). The PP strength of left
panel (and in A) was 140%. C: Comparison of pattern separation
(color-coded, measured as above) in dependence of sprouting and
PP strength. With the combination of leaky GCs and PP80100,
no similarity score could be computed (white). In many conditions
of epileptic hyperexcitation, the DG with CTRL GCs failed to separate patterns (red colors, left panel). In contrast, with leaky GCs
pattern separation was restored under conditions of epileptic
hyperexcitation (right panel). [Color figure can be viewed in the
online issue, which is available at wileyonlinelibrary.com.]
Hippocampus

YIM ET AL

DISCUSSION
Here, we studied the impact of cell-intrinsic conductance
scaling on the function of a neuronal network, using a modified version of the DG network model from the work by Santhakumar et al. (2005). We demonstrate that pattern
separation capability emerges from our version of the network,
consistent with the proposed function of the DG (Treves and
Rolls, 1992; OReilly and McClelland, 1994; Gilbert et al.,
2001; Leutgeb et al., 2007; McHugh et al., 2007; Moser et al.,
2008; Clelland et al., 2009; Myers and Scharfman, 2009;
Sahay et al., 2011; Yassa et al., 2011; Deng et al., 2013; Santoro, 2013). The pattern separation ability was degraded by
epileptic hyperexcitation and to variable degrees restored by the
epilepsy-related adaptation of intrinsic GC properties, which
we determined experimentally before (Young et al., 2009).

The Dentate Gyrus Network Model With Leaky


Granule Cells
We added Kir channels and a tonic GABAA leak to the original model by Santhakumar et al. (2005) as we found these conductances upregulated in GCs of the intrahippocampal kainate
model of TLE (Young et al., 2009). As discussed by Young et al.
(2009), the leaky GC phenotype has not been found in studies
using the systemic pilocarpine epilepsy model, although one
exception appeared recently (Mehranfard et al., 2014). The difference between GC phenotypes of TLE models, and among
TLE patients, is likely due to dissimilar degrees of hippocampal
sclerosis (Stegen et al., 2009; Young et al., 2009). We have demonstrated the leaky GC phenotype in several studies, not only for
our TLE model but also for TLE patients and we hypothesized
that it is an adaptive cellular mechanism in response to epileptic
hyperexcitation (Stegen et al., 2009, 2012; Young et al., 2009;
Kirchheim et al., 2013). A difference between our animal model
and TLE patients is that, instead of the tonic GABAA, a cation
conductance via hyperpolarization-activated nucleotide-gated
(HCN) channels was responsible for part of the increased leak in
human GCs (Stegen et al., 2012). The effect of GABAA and
HCN currents is in both cases shunting and slight depolarization.
In our view, both mechanisms represent the same cellular strategy: to obtain downscaling of excitability while keeping resting
potential homeostasis (Stegen et al., 2012; Meier et al., in press;).
The latter would be in danger with Kir upregulation alone.
It is known that the GABAA leak of GCs is depolarizing
even for mature GCs (Farrant and Nusser, 2005; Pathak et al.,
2007; Chiang et al., 2012). However, it could be more or less
depolarizing in vivo, change during repetitive stimulations
(Pathak et al., 2007), and vary within neuronal subdomains
(Duebel et al., 2006). We have performed our simulations also
with an ECl of 260 mV and obtained qualitatively the same
results (data not shown). As long as the shunting inhibition
prevails, we expect the leak effect to be similar (Meier et al., in
press). Only with an ECl above AP threshold, this increased
chloride conductance would promote seizures and our results
Hippocampus

would turn to the opposite. We did not implement variability


of GC properties (Young et al., 2009) which would further
increase the variability of shown effects.

Modeling TLE Seizures With Mossy Fiber


Sprouting and Increased Perforant Path Input
Previous DG models altered the synaptic connectivity (Santhakumar et al., 2005; Dyhrfjeld-Johnsen et al., 2007; Morgan
and Soltesz, 2008) to explain the occurrence of hippocampal
seizure activity during TLE. Although it is possible that due to
MF sprouting, GCs can act as hubs, effectively spreading epileptic excitation (Morgan and Soltesz, 2008; Scharfman and
Pierce, 2012), it is controversial how much MF sprouting really
contributes to TLE seizures (Scharfman et al., 2003; Buckmaster, 2012; Heng et al., 2013). The precise numbers for sprouted
functional GCGC connections in TLE are unavailable and
likely vary among different epilepsy models and between
patients (Scharfman et al., 2003; Santhakumar et al., 2005;
Morgan and Soltesz, 2008; Buckmaster, 2012; Scharfman and
Pierce, 2012). In our setting, even 20% already disabled DG
pattern separation, which is why we did not employ more than
30% sprouting. In the seizure context, such sprouting conditions are considered mild (Santhakumar et al., 2005).
It is difficult to decide whether a pattern constitutes a
seizure or not; single cell resolution spike patterns of the
entire GC population during behavioral TLE seizures are not
available. Perhaps during real TLE seizures, all GCs are activated in an ordered manner as in the work by Santhakumar
et al. (2005) where the seizure was started by synchronous activation of 100 neighboring GCs. It is also possible that during
real seizures, activity spread is less ordered and spatiotemporally
distributed as in some cases of our study. Irrespective of the precise activity structure, it is likely that increased GC activity during seizures is detrimental for DG pattern separation abilities.
A mechanism that has received less attention but quite likely is
involved in hyperexcitation of the hippocampus during TLE is the
increased PP input from entorhinal cortex layer II neurons to GCs
(Spencer, 2002; Kobayashi et al., 2003; Bonilha et al., 2007). We
implemented a new epileptic condition in the network model,
combining mild sprouting and increased PP input. In doing so,
we aimed for a regime in which the disturbing influence was gradual rather than all-or-none, consistent with the fact that impairment of hippocampal memory in TLE patients is also rather
gradual than all-or-none (Cashdollar et al., 2009; Coras et al.,
2014). Obviously, other hippocampal cell types (some of which
are lost during TLE) are also important for memory acquisition.
Similar to the discussed sprouting, the precise strength of all
PP inputs is unknown (Amaral and Lavenex, 2006). For our control condition, we first lowered the baseline PP strength of the
original model in order to obtain the low spike rates of GCs
observed in vivo; their sparse firing is a presumed prerequisites for
pattern separation (Treves and Rolls, 1992; Jung and McNaughton, 1993; OReilly and McClelland, 1994; Chawla et al., 2005).
This change was also necessary to obtain physiologically more
realistic EPSP summation in time (McNaughton et al., 1981)

INTRINSIC RESCALING AND NEURONAL NETWORK PATTERN SEPARATION


and thereby more fine graded pattern processing. With the above
discussed changes, we obtained a DG network model which was
indeed able to perform pattern separation.

Computation of Pattern Separation in the


Dentate Gyrus Network Model
Previous modeling studies on pattern separation in the DG
focused on synaptic connectivity and its influence on the processing of spatial codes (Myers and Scharfman, 2009, 2011). In
spatial or binary population codes, i.e. without a time domain,
neurons can be on or off and a stationary pattern consists of
the active cell identities. In the present study, we incorporated
the time domain into the pattern separation task by taking
into account both the origin and the timing of spikes. This
modification was motivated by the fact that the intrinsic
changes we observed experimentally did affect the time window
for spike generation (Stegen et al., 2009; Young et al., 2009;
Stegen et al., 2012). Furthermore, hippocampal oscillations
in vivo suggest that spike timing is crucial for DG functioning
(OKeefe and Recce, 1993; Lisman et al., 2005; Leutgeb et al.,
2007; Rangel et al., 2013). Generally, spike timing serves as an
extra dimension of information encoding (Panzeri et al., 2010)
and input timing can significantly affect both the activity level
and the correlation structure of outputs (Yim et al., 2014).
With the above changes we could produce complex partially
overlapping patterns, the similarity of which had to be quantified to assess pattern separation performance. Our quantification via the similarity score was based on a measure which is
corrected for rate dependence (de la Rocha et al., 2007). As
discussed above, the temporal structure of GC firing during
TLE seizures is unknown but generally it is thought that
within seizures the synchronicity of neuronal activity is elevated
(Uhlhaas and Singer, 2006). As the latter often occurs with
higher firing rates of neurons, an apparent rate dependence of
the similarity score is observed despite the rate correction of
the measure. Hence, the failure of pattern separation during
seizure-like activity that we measured is not due to the pure
rate effect but instead due to a combination of network and
cell-intrinsic factors. Of these, we investigated the isolated
influence of experimentally constrained GC-intrinsic plasticity.
We found that MF sprouting had a drastic non-linear impact
on the pattern separation performance of the DG. In contrast,
the PP strength had a more gradual effect. Whether the sudden
drop in pattern separation ability occurs exactly at 20% sprouting could be due to various factors. However, a non-linear relation between sprouting and pattern separation does make sense
in view of the all-or-none DG activity with recurrent GC connections (Fig. 3). Thus, the pattern separation ability of our
DG model was catastrophically disabled by sprouting but more
gradually disturbed by increasing PP input strength.

Impact of Intrinsic Scaling in Granule Cells on


Dentate Gyrus Network Function
The intrinsic properties of neurons not only determine
whether a spike is generated but also when. Yet, the impact of

cell type-specific intrinsic (non-synaptic) plasticity was investigated in only few network modeling studies (Howard et al.,
2007; Thomas et al., 2009, 2010). Pattern separation is prominently discussed in the field of dentate neurogenesis because the
addition of immature GCs to the DG affects behavioral discrimination (Aimone et al., 2009, 2011; Clelland et al., 2009; Deng
et al., 2010; Sahay et al., 2011; Nakashiba et al., 2012; Tronel
et al., 2012; Dery et al., 2013; Rangel et al., 2013). Since immature GCs have different intrinsic properties compared to mature
GCs (Haussler et al., 2012), the role of neurogenesis in pattern
separation could be viewed in favor of our hypothesis that intrinsic properties of GCs are important for the task. However, the
polarity of how newborn GCs are involved is not straightforward:
if it is true that specifically the sparse activation of GCs supports
pattern separation (Treves and Rolls, 1992; Jung and McNaughton, 1993; OReilly and McClelland, 1994; Chawla et al., 2005),
then the highly excitable immature GCs would be expected to
degrade pattern separation. Yet, GC neurogenesis appears to support DG-dependent behavioral discrimination (Sahay et al.,
2011; Nakashiba et al., 2012; Tronel et al., 2012; Dery et al.,
2013). Some of the puzzle could be due to mechanistic differences underlying computational and behavioral pattern separation
(Santoro, 2013). Furthermore, it is possible that immature GCs
are involved in tasks slightly different from classic pattern separation, e.g. increasing memory resolution and/or time stamping
on larger time scales (Clelland et al., 2009; Aimone et al., 2009,
2011; Deng et al., 2010; Rangel et al., 2013). It will be interesting to test our hypotheses by selective modification of GCs in
vivo combined with behavioral experiments.
In summary, the present study demonstrates that the intrinsic ion channel adaptations discovered previously (Stegen et al.,
2009; Young et al., 2009) are in a good position to homeostatically assure the sparseness of GC activation and thereby
dynamically maintain DG network function under different
levels of network excitability.

Acknowledgments
Authors thank Catherine E. Myers for providing and discussing her simulation code on the separation of binary patterns.
Part of the analysis was conducted using the high performance
computing facilities offered by the Information Technology
Services, University of Hong Kong.

REFERENCES
Aimone JB, Wiles J, Gage FH. 2009. Computational influence of
adult neurogenesis on memory encoding. Neuron 61:187202.
Aimone JB, Deng W, Gage FH. 2011. Resolving new memories: A
critical look at the dentate gyrus, adult neurogenesis, and pattern
separation. Neuron 70:589596.
Amaral DG, Lavenex P. 2006. Hippocampal Neuroanatomy. In:
Andersen P, Morris RG, Amaral DG, Bliss T, OKeefe J, editors.
The Hippocampus Book. Oxford, New York: Oxford University
Press. pp 37114.
Hippocampus

10

YIM ET AL

Bonilha L, Yasuda CL, Rorden C, Li LM, Tedeschi H, de Oliveira E,


Cendes F. 2007. Does resection of the medial temporal lobe improve
the outcome of temporal lobe epilepsy surgery? Epilepsia 48:571
578.
Buckmaster PS. 2012. Mossy fiber sprouting in the dentate gyrus. In:
Noebels JL, Avoli M, Rogawski MA, Olsen RW, Delgado-Escueta
AV, editors. Jaspers Basic Mechanisms of the Epilepsies, 4th ed.
Bethesda (MD): National Center for Biotechnology Information.
Available from: http://www.ncbi.nlm.nih.gov/books/NBK98174/.
Buckmaster PS, Strowbridge BW, Kunkel DD, Schmiege DL,
Schwartzkroin PA. 1992. Mossy cell axonal projections to the dentate gyrus molecular layer in the rat hippocampal slice. Hippocampus 2:349362.
Buzsaki G. 2010. Neural syntax: cell assemblies, synapsembles, and
readers. Neuron 68:362385.
Cashdollar N, Malecki U, Rugg-Gunn FJ, Duncan JS, Lavie N, Duzel
E. 2009. Hippocampus-dependent and -independent theta-networks
of active maintenance. Proc Natl Acad Sci USA 106:2049320498.
Chawla MK, Guzowski JF, Ramirez-Amaya V, Lipa P, Hoffman KL,
Marriott LK, Worley PF, McNaughton BL, Barnes CA. 2005.
Sparse, environmentally selective expression of Arc RNA in the
upper blade of the rodent fascia dentata by brief spatial experience.
Hippocampus 15:579586.
Chiang PH, Wu PY, Kuo TW, Liu YC, Chan CF, Chien TC, Cheng
JK, Huang YY, Chiu CD, Lien CC. 2012. GABA is depolarizing
in hippocampal dentate granule cells of the adolescent and adult
rats. J Neurosci 32:6267.
Clelland CD, Choi M, Romberg C, Clemenson GD Jr, Fragniere A,
Tyers P, Jessberger S, Saksida LM, Barker RA, Gage FH, Bussey
TJ. 2009. A functional role for adult hippocampal neurogenesis in
spatial pattern separation. Science 325:210213.
Coras R, Pauli E, Li J, Schwarz M, Rossler K, Buchfelder M, Hamer
H, Stefan H, Blumcke I. 2014. Differential influence of hippocampal subfields to memory formation: insights from patients with
temporal pole epilepsy. Brain 137:19451957.
de Almeida L, Idiart M, Lisman JE. 2009. The input-output transformation of the hippocampal granule cells: from grid cells to place
fields. J Neurosci 29:75047512.
de la Rocha J, Doiron B, Shea-Brown E, Josic K, Reyes A. 2007. Correlation between neural spike trains increases with firing rate.
Nature 448:802806.
Deng W, Aimone JB, Gage FH. 2010. New neurons and new memories: how does adult hippocampal neurogenesis affect learning and
memory? Nat Rev Neurosci 11:339350.
Deng W, Mayford M, Gage FH. 2013. Selection of distinct populations of dentate granule cells in response to inputs as a mechanism
for pattern separation in mice. Elife 2:e00312.
Dery N, Pilgrim M, Gibala M, Gillen J, Wojtowicz JM, Macqueen
G, Becker S. 2013. Adult hippocampal neurogenesis reduces memory interference in humans: Opposing effects of aerobic exercise
and depression. Front Neurosci 7:66.
Duebel J, Haverkamp S, Schleich W, Feng G, Augustine GJ, Kuner T,
Euler T. 2006. Two-photon imaging reveals somatodendritic chloride gradient in retinal ON-type bipolar cells expressing the biosensor Clomeleon. Neuron 49:8194.
Dyhrfjeld-Johnsen J, Santhakumar V, Morgan RJ, Huerta R, Tsimring
L, Soltesz I. 2007. Topological determinants of epileptogenesis in
large-scale structural and functional models of the dentate gyrus
derived from experimental data. J Neurophysiol 97:15661587.
Eichenbaum H. 2013. Memory on time. Trends Cogn Sci 17:8188.
Farrant M, Nusser Z. 2005. Variations on an inhibitory theme: Phasic
and tonic activation of GABA(A) receptors. Nat Rev Neurosci 6:
215229.
Freund TF, Buzsaki G. 1996. Interneurons of the hippocampus. Hippocampus 6:347470.
Hippocampus

Gilbert PE, Kesner RP, Lee I. 2001. Dissociating hippocampal subregions: Double dissociation between dentate gyrus and CA1. Hippocampus 11:626636.
Han ZS, Buhl EH, Lorinczi Z, Somogyi P. 1993. A high degree of
spatial selectivity in the axonal and dendritic domains of physiologically identified local-circuit neurons in the dentate gyrus of the rat
hippocampus. Eur J Neurosci 5:395410.
Haussler U, Bielefeld L, Froriep UP, Wolfart J, Haas CA. 2012. Septotemporal position in the hippocampal formation determines epileptic and neurogenic activity in temporal lobe epilepsy. Cereb
Cortex 22:2636.
Heng K, Haney MM, Buckmaster PS. 2013. High-dose rapamycin
blocks mossy fiber sprouting but not seizures in a mouse model of
temporal lobe epilepsy. Epilepsia 54:15351541.
Hines ML, Carnevale NT. 1997. The Neuron simulation environment. Neural Comput 9:11791209.
Howard AL, Neu A, Morgan RJ, Echegoyen JC, Soltesz I. 2007.
Opposing modifications in intrinsic currents and synaptic inputs in
post-traumatic mossy cells: Evidence for single-cell homeostasis in
a hyperexcitable network. J Neurophysiol 97:23942409.
Jung MW, McNaughton BL. 1993. Spatial selectivity of unit activity
in the hippocampal granular layer. Hippocampus 3:165182.
Kirchheim F, Tinnes S, Haas CA, Stegen M, Wolfart J. 2013. Regulation of action potential delays via voltage-gated potassium Kv1.1
channels in dentate granule cells during hippocampal epilepsy.
Front Cell Neurosci 7:248.
Kobayashi M, Wen X, Buckmaster PS. 2003. Reduced inhibition and
increased output of layer II neurons in the medial entorhinal cortex in a model of temporal lobe epilepsy. J Neurosci 23:8471
8479.
Lawrence JJ, McBain CJ. 2003. Interneuron diversity series: Containing the detonationFeedforward inhibition in the CA3 hippocampus. Trends Neurosci 26:631640.
Leutgeb JK, Leutgeb S, Moser MB, Moser EI. 2007. Pattern separation in the dentate gyrus and CA3 of the hippocampus. Science
315:961966.
Lisman JE, Talamini LM, Raffone A. 2005. Recall of memory sequences by interaction of the dentate and CA3: a revised model of the
phase precession. Neural Netw 18:11911201.
McHugh TJ, Jones MW, Quinn JJ, Balthasar N, Coppari R, Elmquist
JK, Lowell BB, Fanselow MS, Wilson MA, Tonegawa S. 2007.
Dentate gyrus NMDA receptors mediate rapid pattern separation
in the hippocampal network. Science 317:9499.
McNaughton BL, Barnes CA, Andersen P. 1981. Synaptic efficacy and
EPSP summation in granule cells of rat fascia dentata studied in
vitro. J Neurophysiol 46:952966.
Mehranfard N, Gholamipour-Badie H, Motamedi F, Janahmadi M,
Naderi N. 2014. Occurrence of two types of granule cells with different excitability in rat dentate gyrus granule cell layer following
pilocarpine-induced status epilepticus. Annu Res Rev Biol 4:3707
3715.
Meier J, Semtner M, Winkelmann A, Wolfart J. 2014. Presynaptic
mechanisms of neuronal plasticity and their role in epilepsy. Front
Cell Neurosci 8:164.
Meier J, Semtner M, Wolfart J. Homeostasis of neuronal excitability
via synaptic and intrinsic inhibitory mechanisms. In: Boison D,
Masino SA, editors. Homeostatic Control of Brain Function.
Oxford: Oxford University Press (in press).
Morgan RJ, Soltesz I. 2008. Nonrandom connectivity of the epileptic
dentate gyrus predicts a major role for neuronal hubs in seizures.
Proc Natl Acad Sci USA 105:61796184.
Moser EI, Kropff E, Moser MB. 2008. Place cells, grid cells, and the
brains spatial representation system. Annu Rev Neurosci 31:6989.
Myers CE, Scharfman HE. 2009. A role for hilar cells in pattern separation in the dentate gyrus: A computational approach. Hippocampus 19:321337.

INTRINSIC RESCALING AND NEURONAL NETWORK PATTERN SEPARATION


Myers CE, Scharfman HE. 2011. Pattern separation in the dentate gyrus:
A role for the CA3 backprojection. Hippocampus 21:11901215.
Nakashiba T, Cushman JD, Pelkey KA, Renaudineau S, Buhl DL,
McHugh TJ, Rodriguez Barrera V, Chittajallu R, Iwamoto KS,
McBain CJ, Fanselow MS, Tonegawa S. 2012. Young dentate granule cells mediate pattern separation, whereas old granule cells facilitate pattern completion. Cell 149:188201.
OKeefe J, Recce ML. 1993. Phase relationship between hippocampal
place units and the EEG theta rhythm. Hippocampus 3:317330.
OReilly RC, McClelland JL. 1994. Hippocampal conjunctive encoding,
storage, and recall: avoiding a trade-off. Hippocampus 4:661682.
Panzeri S, Brunel N, Logothetis NK, Kayser C. 2010. Sensory neural codes
using multiplexed temporal scales. Trends Neurosci 33:111120.
Pathak HR, Weissinger F, Terunuma M, Carlson GC, Hsu FC, Moss
SJ, Coulter DA. 2007. Disrupted dentate granule cell chloride regulation enhances synaptic excitability during development of temporal lobe epilepsy. J Neurosci 27:1401214022.
Rangel LM, Quinn LK, Chiba AA, Gage FH, Aimone JB. 2013. A
hypothesis for temporal coding of young and mature granule cells.
Front Neurosci 7:75.
Sahay A, Scobie KN, Hill AS, OCarroll CM, Kheirbek MA,
Burghardt NS, Fenton AA, Dranovsky A, Hen R. 2011. Increasing
adult hippocampal neurogenesis is sufficient to improve pattern
separation. Nature 472:46670.
Santhakumar V, Aradi I, Soltesz I. 2005. Role of mossy fiber sprouting and mossy cell loss in hyperexcitability: a network model of
the dentate gyrus incorporating cell types and axonal topography.
J Neurophysiol 93:437453.
Santoro A. 2013. Reassessing pattern separation in the dentate gyrus.
Front Behav Neurosci 7:96.
Scharfman HE, Pierce JP. 2012. New insights into the role of hilar
ectopic granule cells in the dentate gyrus based on quantitative
anatomic analysis and three-dimensional reconstruction. Epilepsia
53 (Suppl 1):109115.
Scharfman HE, Sollas AL, Berger RE, Goodman JH. 2003. Electrophysiological evidence of monosynaptic excitatory transmission
between granule cells after seizure-induced mossy fiber sprouting.
J Neurophysiol 90:25362547.
Schneider CJ, Bezaire M, Soltesz I. 2012. Toward a full-scale computational model of the rat dentate gyrus. Front Neural Circuits 6:
83.
Spencer SS. 2002. Neural networks in human epilepsy: Evidence of
and implications for treatment. Epilepsia 43:219227.
Spencer SS, Spencer DD. 1994. Entorhinal-hippocampal interactions
in medial temporal lobe epilepsy. Epilepsia 35:721727.
Stegen M, Young CC, Haas CA, Zentner J, Wolfart J. 2009. Increased
leak conductance in dentate gyrus granule cells of temporal lobe
epilepsy patients with Ammons horn sclerosis. Epilepsia 50:646
653.
Stegen M, Kirchheim F, Hanuschkin A, Staszewski O, Veh RW,
Wolfart J. 2012. Adaptive intrinsic plasticity in human dentate
gyrus granule cells during temporal lobe epilepsy. Cereb Cortex
22:20872101.
Suzuki F, Makiura Y, Guilhem D, Sorensen JC, Onteniente B. 1997.
Correlated axonal sprouting and dendritic spine formation during
kainate-induced neuronal morphogenesis in the dentate gyrus of
adult mice. Exp Neurol 145:203213.
Tauck DL, Nadler JV. 1985. Evidence of functional mossy fiber
sprouting in hippocampal formation of kainic acid-treated rats.
J Neurosci 5:10161022.
Thomas EA, Reid CA, Berkovic SF, Petrou S. 2009. Prediction by
modeling that epilepsy may be caused by very small functional
changes in ion channels. Arch Neurol 66:12251232.
Thomas EA, Reid CA, Petrou S. 2010. Mossy fiber sprouting interacts
with sodium channel mutations to increase dentate gyrus excitability. Epilepsia 51:13645.

11

Treves A, Rolls ET. 1992. Computational constraints suggest the need


for two distinct input systems to the hippocampal CA3 network.
Hippocampus 2:189199.
Tronel S, Belnoue L, Grosjean N, Revest JM, Piazza PV, Koehl M,
Abrous DN. 2012. Adult-born neurons are necessary for extended
contextual discrimination. Hippocampus 22:292298.
Uhlhaas PJ, Singer W. 2006. Neural synchrony in brain disorders:
Relevance for cognitive dysfunctions and pathophysiology. Neuron
52:155168.
Wolfart J, Debay D, Le Masson G, Destexhe A, Bal T. 2005. Synaptic
background activity controls spike transfer from thalamus to cortex. Nat Neurosci 8:17601767.
Yassa MA, Lacy JW, Stark SM, Albert MS, Gallagher M, Stark CE.
2011. Pattern separation deficits associated with increased hippocampal CA3 and dentate gyrus activity in nondemented older
adults. Hippocampus 21:968979.
Yim MY, Kumar A, Aertsen A, Rotter S. 2014. Impact of correlated
inputs to neurons: modeling observations from in vivo intracellular
recordings. J Comput Neurosci 37:293304.
Young CC, Stegen M, Bernard R, Muller M, Bischofberger J, Veh
RW, Haas CA, Wolfart J. 2009. Upregulation of inward rectifier
K1 (Kir2) channels in dentate gyrus granule cells in temporal lobe
epilepsy. J Physiol 587 (Part 17):42134233.
Yu J, Proddutur A, Elgammal FS, Ito T, Santhakumar V. 2013. Status
epilepticus enhances tonic GABA currents and depolarizes GABA
reversal potential in dentate fast-spiking basket cells. J Neurophysiol
109:17461763.

APPENDIX

Cellular Properties in the Dentate Gyrus


Computer Model
Our conductance-based network model of the dentate gyrus
(DG) was developed based on the model by Santhakumar et al.
(2005). If not mentioned otherwise, all parameters were inherited from the original model and their detailed description and
justification can be found in the original study (Santhakumar
et al., 2005). We used the publicly available Neuron scripts from
ModelDB (http://senselab.med.yale.edu/ModelDB) which contain some differences to the values published by Santhakumar
et al. (2005; see Table AI). The DG network consists of 500
granule cells (GCs), 6 basket cells (BCs), 15 mossy cells (MCs),
and 6 hilar perforant path (PP)-associated cells (HIPP cells, here
HCs). We modified the ion channel conductances and dynamics
of GCs according to our experimental data from the intrahippocampal kainate injection mouse model of temporal lobe epilepsy
(TLE), and incorporated a Kir conductance and a tonic GABAA
chloride conductance into the model (Young et al., 2009). Due
to the lack of knowledge about the precise subcellular distribution, these conductances were homogeneously distributed over
all somatodendritic segments. The contribution of the tonic
GABAA chloride conductance is represented by an additional
ionic current ItonicGABAA 52g tonicGABAA V 2ECl , where ECl is
the corresponding reversal potential which was set to 270 mV
(Pathak et al., 2007; Chiang et al., 2012). The Kir current is represented by IKir 52g Kir mV 2EKir , where EKir is the reversal
potential of potassium ions and was set to 2105 mV (Young
Hippocampus

12

YIM ET AL

TABLE AI.
Parameters of the Network Model During Control Condition
From (row)/to (column)
PP
Connectivity
Weight (nS)
Delay (ms)
GC
Connectivity
Weight (nS)
Delay (ms)
BC
Connectivity
Weight (nS)
Delay (ms)
MC
Connectivity
Weight (nS)
Delay (ms)
HIPP
Connectivity
Weight (nS)
Delay (ms)

GC

BC

MC

HIPP

(%)

20 (CC)
2.0a
3

20a (CC)
1.0a
3

(%)

(sprouted)
2.0
0.8

16.7
14.1a
0.8

6.7
0.2
1.5

(%)

20
4.8a
0.85

33.3
7.6
0.8

20
1.5
1.5

(%)

40
0.3
3

16.7b
0.3
3

20b
0.5
2

(%)

32
0.5
1.6

66.7
0.5
1.6

26.7
1.5b
1

Connection Properties in the Dentate Gyrus


Network Model
50
0.5
1.5

33.3
0.2
3

CC, convergent connection (each postsynaptic cell receives the same number of
presynaptic afferents); all other connections are divergent (postsynaptic cell can
have different numbers of presynaptic afferents).
a
Values were modified from the original model.
b
Entries differed between Santhakumar et al. (2005) and their ModelDB script.
We used the values and the presynaptic target pool as published in ModelDB.

et al., 2009). The dynamics of the gating variable m is governed


by dm
dt 5 m1 2m=s, where m1 51=11exp V 2V50 =k
and s 5 1=a 3 exp 2V =V0 1 b 3 exp V =V0 . The parameters for the s function (V0 5 67.08 mV, a 5 6.1 s21, and b 5
81.8 s21) were taken from the work by Stegen et al. (2012; note
typo of b value therein). The parameter m1 was obtained by fitting a Boltzmann function to the mean I/V curve without constraining k (V50 5 298.92 mV, k 5 10.89 mV; Stegen et al.,
2012).
For the control DG (CTRL DG) condition, the additional
leak conductances (g Kir 5 1.44 3 1025 S/cm2 and g tonicGABAA 5
7:22 31026 S/cm2) were scaled together with the existing passive leak (gpas 5 1.44 3 1025S/cm2), the intracellular resistivity
(Ra 5 184 X cm) and the total surface area to obtain an input
resistance of 439.6 MX to fit the measured values of a
recorded and morphologically reconstructed GC with a surface
of 9125 mm2 (Young et al., 2009; J. Wolfart, unpublished
data). The scaling was performed by fitting model parameters
to experimental data using principal axis optimization, a technique implemented as fit_praxis in the NEURON simulator
(Hines and Carnevale, 1997). To agree with the above GC surface area, the surface of model GCs was scaled down proportionally to 1/1.13 of Santhakumars GC model. For the
epileptic condition, in which GCs become leaky (Stegen et al.,
2009; Young et al., 2009), the Kir and GABAA conductances
Hippocampus

were scaled up 4-fold (to leaky: g Kir 55:7631025 S/cm2 and


g tonicGABAA 52:8931025 S/cm2), respectively, such that the Rin
was reduced to 264.3 MX, consistent with the mean experimentally observed values in GCs of epileptic animals (Young
et al., 2009).

We employed the topographic ring version of the model by


Santhakumar et al. (2005). In this model, the connectivity is
constrained by biologically inspired axonal arborizations
domains, i.e. neighboring cells have higher connectivity than
distant ones. Within these constraints, we used identical seeds
for randomizing connections and patterns in all tested conditions as in the original model, except for specifically mentioned
control purposes. Details of connectivity for the CTRL DG are
shown in Table AI (changes to the original model are marked
by superscripts). For the postsynaptic target pool of the divergent connections (the number of presynaptic afferents can
vary), see Table 5 in the work by Santhakumar et al. (2005).
As justified in the Methods, for sparse activation of GCs in the
CTRL DG, we decreased the PP input strength to GCs and
BCs to 10% of the original values (PP-GC, from 20 to
2.0 nS; PP-BC, from 10 to 1.0 nS). These values represent
100% in our study, i.e. our 100% of the PP-GC synapse is
2.0 nS. We tested the following different PP-GC input
strengths 80% (1.6 nS, labeled PP80), 100% (2.0 nS, CTRL),
120% (2.4 nS, PP120), 140% (2.8 nS, PP140), and 160%
(3.2 nS, PP160). The corresponding PP-BC strength was coscaled accordingly. To increase the inhibitory influence of BCs
(Lawrence and McBain, 2003), we increased their connection
strength threefold (GC-BC, from 4.7 to 14.1 nS and BC-GC,
from 1.6 to 4.8 nS) and raised the connectivity from PP to
BC (only 2 of the 6 BCs received 100 PP inputs in Santhakumars model while all BCs received 100 PP inputs in our
model). Note that the real functional strengths of PP synaptic
input to GCs, BCs etc., are unknown, already for the physiological situation (Amaral and Lavenex, 2006). Similarly, it is
unclear whether some PP inputs reach hilar neurons directly
(Santhakumar et al., 2005; Myers and Scharfman, 2009),
which has been doubted (Han et al., 1993; Freund and Buzsaki, 1996; Amaral and Lavenex, 2006). If they do exist, their
influence is bound to be small and unlikely to affect our results
significantly. Similarly, because the PP-MC connection is likely
to be very weak (Buckmaster et al., 1992), we left it inactivated
as in the original script available on ModelDB. The input to
GCs is received from 100 different PP input lines from which
six are active during each input pattern.
For the epileptic DG condition we implemented first the MF
sprouting as in the work by Santhakumar et al. (2005) and
used their nomenclature of % sprouting, testing 0, 10, 15, 20,
and 30% MF sprouting. Furthermore, we increased the
PP-GC synaptic strength (see above) and combined the two
epileptic mechanisms to variable degrees. The corresponding
PP-BC strength was co-scaled accordingly (see above).

You might also like