Intrinsic mechanisms stabilize encoding and retrieval circuits differentially in a hippocampal network model

Abstract

Acetylcholine regulates memory encoding and retrieval by inducing the hippocampus to switch between pattern separation and pattern completion modes. However, both processes can introduce significant variations in the level of network activity and potentially cause a seizure-like spread of excitation. Thus, mechanisms that keep network excitation within certain bounds are necessary to prevent such instability. We developed a biologically realistic computational model of the hippocampus to investigate potential intrinsic mechanisms that might stabilize the network dynamics during encoding and retrieval. The model was developed by matching experimental data, including neuronal behavior, synaptic current dynamics, network spatial connectivity patterns, and short-term synaptic plasticity. Furthermore, it was constrained to perform pattern completion and separation under the effects of acetylcholine. The model was then used to investigate the role of short-term synaptic depression at the recurrent synapses in CA3, and inhibition by basket cell (BC) interneurons and oriens lacunosum-moleculare (OLM) interneurons in stabilizing these processes. Results showed that when CA3 was considered in isolation, inhibition solely by BCs was not sufficient to control instability. However, both inhibition by OLM cells and short-term depression at the recurrent CA3 connections stabilized the network activity. In the larger network including the dentate gyrus, the model suggested that OLM inhibition could control the network during high cholinergic levels while depressing synapses at the recurrent CA3 connections were important during low cholinergic states. Our results demonstrate that short-term plasticity is a critical property of the network that enhances its robustness. Furthermore, simulations suggested that the low and high cholinergic states can each produce runaway excitation through unique mechanisms and different pathologies. Future studies aimed at elucidating the circuit mechanisms of epilepsy could benefit from considering the two modulatory states separately.

Keywords: Pattern separation and completion; acetylcholine; biologically realistic model; interneurons; seizures.

FIGURE 1.

In vitro current injection recordings of the cell types and their matching model cells. Current injections used in both experimental recordings and model are displayed underneath each pair of recordings and model traces. Sources for the experimental data: CA3 pyramidal cell (Brown and Randall, 2009), DG granule cell (Staley et al., 1992), OLM cell (Ali and Thomson, 1998), and basket cell (Buhl et al., 1996). The parameter values for the model cells are in Table S1 of Supporting Information.

FIGURE 2.

Network 3D structure and CA3 local circuitry. (A) Schematic of the network implemented showing the modeled regions EC, CA3, and DG with their dimensions, cell numbers, and lamellar connectivity pattern. Neurons in EC are more likely to send connections to DG and CA3 neurons in their longitudinal vicinity. Similarly, DG granule cells in the same longitudinal neighborhood are likely to project to CA3 neurons in the same lamella. Cells were compacted into three sheets of cells, in the radial dimension, representing stratum-pyramidale in CA3 and the granular layer in DG. (B) Schematic with details of CA3 internal circuitry. Excitatory connections terminate in arrows and inhibitory ones in black filled circles. (C) Gaussian connection probability functions. The longitudinal organization of EC inputs to CA3 is compared to DG inputs. Inputs from DG had a more focused pattern of connectivity (see Table S3 for parameter values). (D) Projections from MF to BCs had a wider longitudinal extent, compared to the ones from MF to CA3 pyramidal cells (pyr). (E) The probability of an interneuron connecting to a pyramidal cell depended on the distance between the two in the longitudinal and transverse planes. Note that probability for the OLM domain exceeded one to ensure that OLM cells made dense connections in their immediate neighborhood.

FIGURE 3.

Matching short-term plasticity to experimental recordings. Short-term plasticity was modeled using equations proposed by Varela et al. (1997) and parameter values were obtained by matching model to experimental recordings. Note that depending on available data, some panels display the postsynaptic cell membrane potential and others display the synaptic current. Parameter values used to reproduce data are listed in Supporting Information Table S5. (A) Mossy fiber synaptic facilitation (Toth et al., 2000; scale bars: 50 ms, 100 pA). (B) CA3 Pyramidal cell to OLM interneuron (Ali and Thomson, 1998; scale bars: 20 ms, 1 mv). (C) CA3 Pyramidal cell to BC interneuron (Ali et al., 1998; scale bars: 30 ms, 0.5 mv). (D) BC interneuron to CA3 pyramidal cell (Hefft and Jonas, 2005; scale bars: 50 ms, 100 pA). (E and F) Recurrent CA3 connections stimulated at 50 and 20 Hz, respectively (Hoskison et al., 2004). Note that these connections displayed paired pulse facilitation, a phenomenon not included in our synapse model. Therefore, responses to the first stimulus in the train appear larger than the recordings (scale bars: 20 ms, 0.5 mv in E; 50 ms, 0.5 mv in F).

FIGURE 4.

Pattern completion and separation in CA3 and DG. The network learned Pattern 1 for five trials under high levels of ACh. Subsequently, long-term plasticity was inactivated to test retrieval in response to inputs Patterns 1–11. Output spikes were recorded at EC, CA3, and DG, and correlation was calculated between the output from Pattern 1 and the output from each of the test patterns. Results were averaged over data from 10 randomly constructed networks and shaded areas indicate standard deviation. (A) Correlation of output patterns produced at EC, CA3 and DG in response to probe Patterns 1–11. The correlation between input Pattern 1 and the probe patterns at EC is shown as a reference point. Correlation values at CA3 pyramidal neurons lie well above the input correlation indicating a tendency toward pattern completion in CA3. Conversely, DG correlation values lie below input correlation levels indicating pattern separation. (B) The effects of low vs. high levels of ACh during retrieval on correlation levels in CA3.

FIGURE 5.

The effects of recurrent connections on excitation within CA3. At the beginning of each trial, a stimulus of one action potential was delivered to a number of randomly selected of EC neurons, which projected to CA3. The number of EC neurons stimulated increased with each trial. The DG region and other cell types in CA3 were disconnected, with ACh at baseline levels and no short- or long-term plasticity. (A) Spike raster plot of EC neurons in response to structured inputs for 30 trials from a representative network. Each trial had a 500-ms duration, and for each trial, an increasing number of randomly selected EC neurons received a synaptic stimulus generating one action potential. As can be seen in the plot, only one EC neuron fired on Trial 1 in the first 500 ms, two fired during Trial 2, and so on, ending in 30 neurons firing on Trial 30 in the last 500 ms of the simulation. (B) Spike raster plot of CA3 pyramidal neurons without the recurrent connections in response to the input presented at EC across 30 trials. (C) Spike raster plot of CA3 from a representative network with recurrent connections restored. Spikes are shown for the 30 trials of the experiment. (D) Membrane voltage traces from a representative neuron in the network without recurrent connections (D1) and with recurrent connections (D2). The traces show the membrane voltage response during Trials 22–25, with arrowheads marking the beginning of each trial. (E) Population firing rate histogram for CA3 without recurrent connections (dashed line) and with recurrent connections (solid line). Firing rate was calculated for each trial by averaging trial spike count from all CA3 pyramidal neurons and dividing by trial duration. (F) Ratio of active to inactive neurons in EC (dotted diagonal line), in CA3 without recurrent connections (dashed line) and in CA3 with recurrent connections (solid line). Results are averages of data from 10 networks with different random seeds, and shaded areas represent standard deviation. (G) Changes in distribution of different burst sizes with and without recurrent connections. The distribution was obtained by pooling action potential counts from each neuron in each trial. The data were averaged from running 10 initializations of the network, and presented as a percentage of the total neuron-trial count for each burst size. Inset shows a schematic with model components used in this experiment. Active components are shown in dark lines, while inactive components are shown in light gray lines. RC: recurrent connections.

FIGURE 6.

Effects of BC interneurons on the stability of CA3 pyramidal cells. The ratio of neurons activated was measured in CA3 in response to increasing EC input over 30 trials. Diagonal line in graphs denotes the rate of input increase at EC. Inhibition by BCs was the only stability mechanism in this experiment, and short-term depression at recurrent connections and inhibition by OLM cells were inactivated. (A) Ratio of active pyramidal neurons for three different levels of BC-to-pyramidal inhibition. Compared to low BC inhibition, medium inhibition shifted the activity ratio curve modestly to the right, without a substantial decrease in its maximum slope. Furthermore, BC inhibition higher than optimized for the pattern separation and completion network did not effectively lower the slope of the curve. (B) The ratio of active BC interneurons showed a linear response with increase in EC inputs, with little difference between the three levels of BC inhibition. Inset shows a schematic with model components used in this experiment. Active components are shown in dark lines, while inactive components are shown in light gray lines.

FIGURE 7.

OLM interneurons stabilize CA3 and control burst size. The network was tested over 30 trials with increasing amounts of EC inputs, denoted by the diagonal line. Short-term depression at recurrent connections was blocked and BC interneurons inactivated. (A) The ratio of active CA3 pyramidal neurons at three different levels of OLM inhibition. (B) The ratio of active OLM interneurons for low, med, and high OLM-to-pyramidal inhibition showed a nonlinear OLM response. The response of OLM cells continued to be nonlinear even with high inhibition where CA3 pyramidal cells show a controlled near linear response. (C) The percentage change in the occurrence of different burst sizes when inhibition level is lowered from high to low, for BC inhibition (dashed line) and OLM inhibition (solid line).

FIGURE 8.

Recurrent connections short-term depression stabilizes CA3 activity. (A) The effect of short-term depression (three levels) at the recurrent connections on the ratio of active cells in CA3 with increasing input from EC. Both OLM and BC were inactivated. Baseline or “med” values were obtained by matching experimental data (Hoskison et al., 2004; d1: 0.7, d2: 0.92) and then low (d1: 0.5, d2: 0.86) and high (d1: 0.9, d2: 0.98) levels were created symmetrically around the baseline values. Higher levels of short-term depression stabilized CA3 responses and the curve of ratio of active cells became closer to linear and dropped below the curve for the inputs. (STD: short-term depression). (B) The effects of experimentally matched levels of OLM inhibition or of short-term depression alone compared to the case when both mechanisms were simultaneously active.

FIGURE 9.

The encoding circuit is stabilized by OLM inhibition and the retrieval circuit is stabilized by short-term depression at the recurrent CA3 connections. Ratio of active CA3 pyramidal cells was evaluated for low and high ACh levels with either short-term depression or OLM inhibition activated at three levels. (A) The effects of three different levels of short-term depression at the CA3 recurrent connections (A1), and of OLM inhibition (A2) under low levels of ACh. (B) With high ACh, short-term depression at the CA3 recurrent connections had little effect (B1) while OLM inhibition was critical to the stability of the network (B2).

FIGURE 10.

Distinct patterns of excitation during encoding and retrieval levels of ACh. Both CA3 and DG networks received increasing inputs from EC without the stabilizing effects of short-term depression or inhibition from CA3 interneurons. The occurrence of bursts of different sizes was calculated across neurons and trials, and is presented as a percentage of all neurons and trials for each burst size. The distribution of burst sizes was considered for low, med, and high levels of ACh. (A) The occurrence of different burst lengths in CA3 under low, med, and high levels of ACh. Low levels of ACh shifted the peak in burst sizes toward longer bursts (peak shifted from five to eight action potentials). Whereas high levels of ACh leads to more distributed burst sizes with a long tail including extended bursts (>20 action potentials). (B) Same analysis as in (A) applied to DG granule cells revealed that these very long bursts were not simply transmitted from DG. High levels of cholinergic transmission in DG caused an increase in the frequency of short bursts in a limited number of granule cells. (C) OLM interneurons were activated in the encoding network (ACh=2) and the distribution of burst sizes was considered at three different levels of OLM inhibition. Higher levels of OLM inhibition reduced the occurrence of extended bursts. (D) Excitation spread in CA3 network preceded OLM inhibitory current during low ACh levels. One trial is depicted with 20 EC neurons receiving one action potential at time 0. A spike raster plot of CA3 pyramidal cells is shown overlaid with an example of OLM inhibitory current as measured from the soma of a pyramidal cell. Not all pyramidal cells fired in response to the EC stimulus. A few neurons started firing late (Arrows), indicating a di-synaptic source of input from other pyramidal cells. OLM inhibition did not arrive in time to prevent the secondary spread of excitation to these neurons.

FIGURE 11.

Summary of stabilizing mechanisms in low, med and high cholinergic states. Connections are suppressed (dotted arrows) or enhanced (thick arrows) from baseline (normal arrows) by ACh levels. Connections promoting runaway excitation (indicated by *) were the recurrent CA3 connections in low and med ACh states, and MF in high ACh states. On the other hand, mechanisms maintaining network stability (indicated by gray shaded area) were short-term depression in low and med ACh states, and OLM in med and high ACh states.