Episodic Memories: How do the Hippocampus and the Entorhinal Ring Attractors Cooperate to Create Them?

Abstract

The brain is capable of registering a constellation of events, encountered only once, as an episodic memory that can last for a lifetime. As evidenced by the clinical case of the patient HM, memories preserving their episodic nature still depend on the hippocampal formation, several years after being created, while semantic memories are thought to reside in neocortical areas. The neurobiological substrate of one-time learning and life-long storing in the brain, that must exist at the cellular and circuit level, is still undiscovered. The breakthrough is delayed by the fact that studies jointly investigating the rodent hippocampus and entorhinal cortex are mostly targeted at understanding the spatial aspect of learning. Here, we present the concept of an entorhinal cortical module, termed EPISODE module, that could explain how the representations of different elements constituting episodic memories can be linked together at the stage of encoding. The new model that we propose here reconciles the structural and functional observations made in the entorhinal cortex and explains how the downstream hippocampal processing organizes the representations into meaningful sequences.

Keywords: entorhinal cortex; episodic memory; grid cells; hippocampus; phase precession; plasticity; theta phase skipping; theta sequences.

Figure 1

Rodent grid cell activity cannot be explained by a simple 1-dimensional ring attractor. The figure shows the firing fields of two grid cells (“A” and “B”) from the same hypothetical 1-dimensional ring attractor, having equal grid scales but different spatial phases. Forward and backward movement along a linear track could easily be interpreted: when the animal starts its journey from the center of the field of the neuron “A”, the bump can travel the half of the distance of the whole attractor either clockwise or anticlockwise (routes 1 and 2) and reach the same neuron “B”. However, trajectories along 3 and 4 cannot be reconciled with this 1-dimensional model: 3 would guide back the bump to the same neuron (“A”) while 4, a slightly longer route would guide the bump only halfway around the ring attractor to reach the neuron “B”.

Figure 2

Proposed ring attractor in the superficial layers of the entorhinal cortex. Five different cell types are shown, their names are abbreviated according to the nomenclature in the MEC: (i) Stellate or fan neurons (ST), both named SODeR cells in our generalized model, a fraction of which are grid cells in rodents; (ii) Putative pyramidal cells from layer III (CGN and CGS), named PiPhA cells in our generalized model, capable of displacing the activity center in the ring of ST cells. A fraction of these neurons shows conjunctive grid (“CG”) cell activity in rodents, the north (“N”) and south (“S”) subtypes move the bump of activity in opposite directions; and (iii) Superficial entorhinal neurons (HDN and HDS), both named ErReD cells in our generalized model, that provide continuous input to the CGN and CGS cells when an event is ongoing. Medial entorhinal head direction (“HD”) cells are typical neurons falling into this category. For the sake of clarity several features are simplified in the figure: (i) the physical shape of the attractor is not supposed to be circular, the ring only displays its topology; (ii) the two antagonistic set of PiPhA cells (CGN and CGS in the depicted example) do not have to be located on the opposite sides of the SODeR cells located in layer II, actually their most likely position is below them (layer III) for both PiPhA cell type; (iii) the bump of activity is not restricted to one single SODeR cell but a group of adjacent SODeR cells is supposed to be active at the same time; and (iv) the number of cells in the attractor is arbitrary in the figure and is chosen solely to illustrate the logic of wiring.

Figure 3

Creating higher dimensional units from multiple 1-dimensional ring attractors. (A) Medial entorhinal 1-dimensional ring attractors could collaborate to form a 2-dimensional network that accounts for rodent grid cell activity. Each 1-dimensional ring attractor sends information to the layer V of the medial entorhinal cortex to neurons such as Ctip2 positive pyramidal cells that relay this information to a neighboring 1-dimensional ring attractor. The structure could encircle the calbindin positive pyramidal patches, however, this kind of arrangement is not required for the model to be functional, the wrap-around mechanism for the extra dimension is still assured. (B) Medial entorhinal 1-dimensional ring attractors could collaborate to form a 3-dimensional network that can better subserve the hexagonal symmetry of the firing fields that rodent grid cells display. The communication between the 1-dimensional units is similarly possible via layer V, however, the way the wrap-around connectivity is implemented is less straightforward in this case.

Figure 4

Converging and diverging routes of the information from the entorhinal cortex to the granule cells of the DG. Reelin-positive neurons (functionally SODeR cells in our model) provide excitatory input to the granule cells of the DG. Newborn granule cells are particularly receptive within a defined temporal window of their maturation to form connections with incoming fibers that are active. Therefore the elements represented by the 1-dimensional ring attractors can be associated together, and the convergence assures that granule cell activity will be highly specific to the combination of the elements encountered. Of note, the elements themselves can be of any complexity given the different levels of integration in the entorhinal cortex. It is reasonable to assume that early on elements like simple objects can be learned by associating their unique features based on this mechanism. Later such representations could be semanticized and more complex episodic memories based on already learned objects can be created the same way. The figure illustrates this latter stage of learning and shows how objects in a room (door, pendant, window, and shelf) can be associated together to be part of episodic memory. The results from Nakashiba et al. (2012) show that old granule cells can mediate pattern completion to some extent, this would be explained by the fact that a slightly reduced number of active elements (such as three out of the four elements represented in the figure) is still capable of activating a given granule cell that established contacts with a full set of elements (four in the figure) during its young and receptive period.

Figure 5

The proposed mechanism by which the theta phase code generated by the lateral entorhinal (LEC) 1-dimensional ring attractors is passed on to the granule cells of the DG. The 1-dimensional ring attractors are naturally capable of generating a signal that shows theta phase precession at the level of their individual neurons (Navratilova et al., 2012) and phase precession originating in the ECII is passed on to the hippocampus (Schlesiger et al., 2015). The representation of relevant events (such as cues, odors, sounds, places, and complex events in rodent experiments) shows phase precession, and a reasonable assumption is that a lateral entorhinal 1-dimensional ring attractor consisting of many SODeR cells represents an event and generates the event-specific phase precessing signal. As a consequence, the target granule cell must be informed of the current theta phase equally during the anticipation, the onset and the termination of the event. The most likely and simplest network wiring compatible with this requirement is an excitatory connection from each fan cell of a given 1-dimensional ring attractor to a single granule cell. The figure shows that the theta phase of the currently active SODeR cell (fan cell in red) determines the theta phase at which the target granule cell fires. We would like to emphasize that such wiring is neither required nor plausible for rodent grid cells despite their similar organization into 1-dimensional entorhinal ring attractors.

Figure 6

Processing of the phase precessing signals by the recurrent CA3 network. (A) Signals that are subsequent along a theta cycle but arrive within a defined window of time can be made coincident by the recurrent collateral CA3 network. The event represented by the red CA3 neuron (and the remaining unshown neurons of the corresponding population) is previous to the event represented by the gray CA3 neuron on the left (and the remaining unshown neurons of the corresponding population) as evidenced by its earlier theta phase. However, the time it takes for the action potentials to travel along the CA3 recurrent collateral axons can make the red and the gray representations coincident at the level of CA3-CA3 synapses (red dot). (B) The recurrent CA3 network is especially well-positioned to perform pattern completion: the event represented by the gray CA3 neurons can complete the representation of the event the red CA3 neurons code for. The neuron depicted with the red arrow in the perisomatic region is not activated by the external cues but it can still be activated by the population of the gray neurons via the synapses potentiated by previous experience (encircled gray dots). Best pattern completion is assumed to be achieved in the CA3 on one hand for events already organized in a sequence and on the other hand by an event that is previous to the event with incomplete representation.