The environmental sculpting hypothesis of juvenile and adult hippocampal neurogenesis

Abstract

We propose that a major contribution of juvenile and adult hippocampal neurogenesis is to allow behavioral experience to sculpt dentate gyrus connectivity such that sensory attributes that are relevant to the animal's environment are more strongly represented. This "specialized" dentate is then able to store a larger number of discriminable memory representations. Our hypothesis builds on accumulating evidence that neurogenesis declines to low levels prior to adulthood in many species. Rather than being necessary for ongoing hippocampal function, as several current theories posit, we argue that neurogenesis has primarily a prospective function, in that it allows experience to shape hippocampal circuits and optimize them for future learning in the particular environment in which the animal lives. Using an anatomically-based simulation of the hippocampus (BACON), we demonstrate that environmental sculpting of this kind would reduce overlap among hippocampal memory representations and provide representation cells with more information about an animal's current situation; consequently, it would allow more memories to be stored and accurately recalled without significant interference. We describe several new, testable predictions generated by the sculpting hypothesis and evaluate the hypothesis with respect to existing evidence. We argue that the sculpting hypothesis provides a strong rationale for why juvenile and adult neurogenesis occurs specifically in the dentate gyrus and why it declines significantly prior to adulthood.

Keywords: Adult neurogenesis; Dentate gyrus; Hippocampus; Memory.

Fig. 1.. In all species examined, neurogenesis declines before onset of old age.

These data are replotted from Snyder (2019), which synthesized several published data sets to construct these curves. Neurogenesis rates are normalized to the level at birth. See Snyder (2019) for details on methodology.

Fig. 2.. The hippocampal model.

Computations were performed using BACON, a simplified version of Marr’s Archicortex model. Episodic and context memory acquisition depends on plasticity at EC_in–DG, EC_in–CA3, CA3–CA3, and CA3–EC_out synapses, which are highlighted in green. A small set of DG cells (of size K) and their corresponding CA3 followers will become the hippocampal representation of an event being stored. The K cells chosen for the representation will be those that are most heavily innervated by the cortical (EC) cells coding the event. To simplify computations the model assumes that each DG cell innervates one dedicated DG follower (rather than a small number), and that each CA3 cell recurrently innervates every other CA3 cell (rather than only a proportion) and directly innervates each EC_out cell (instead of doing so via CA1); the rationales for these simplifications are discussed in Krasne et al (2015). For the computations of Fig. 5 we assumed (following O’Reilly and McClelland, 1994) 200,000 EC_in and EC_out cells, 12,500 of which fully characterized an event (or context), 1 million DG cells, 4,000 EC_in inputs to each DG cell, and K=4,000. In BACON, as in the biological hippocampus, EC innervates CA3 both directly and indirectly via the DG, and both pathways appear to be involved during recall [There has been some discussion to the contrary in the literature but we think the weight of both evidence and logic say this is so—see Krasne et al, 2015 for a theoretical discussion and Bernier et al, 2017 for an empirical one]. During encoding, those DG cells that are most richly innervated by the active EC cells, together with their CA3 followers, are selected to become the new representation. Because it is the richness of innervation of the DG cells themselves that is the basis for this selection, they have richer input from the relevant EC cells than do their CA3 followers. Consequently, during recall the indirect pathway provides richer information about the situation/context than does the direct one, and optimal performance is obtained when the indirect pathway is made considerably stronger than the direct one. Because using the indirect pathway alone impairs performance only slightly (see Krasne et al, 2015, Fig. 8), we have used only the indirect pathway to simplify the simulations for Figs. 4 and 5.

Figure 3.. Caricature of sculpting effects in BACON.

Here we show how the sculpting made possible by juvenile and adult neurogenesis might influence dentate innervation. Sculpting leads to enriched innervation of DG by cortical cells that will actually be used to code events. Note that the notion that some EC cells go unused is a simplification made to facilitate the simulations of Figs. 4 and 5. We would expect the reality to be that most EC cells are used in any environment but that sculpting causes late-born cells to become innervated by combinations of EC cells that commonly fire together in a given environment

Fig. 4.. Effects of sculpting on context representations in BACON’s DG and CA3.

Graphs display the results of a simulation in which two similar contexts, A and B, were encoded. For simplicity, artificially small sets of EC and DG cells were used. There were 500 EC cells in all, and a given context (or event) was composed of just 6 of them, drawn randomly from a subset of 15 environmentally-relevant EC cells. Each DG cell received input from 10 EC cells. In the non-sculpted case the inputs were chosen randomly from the full set of 500 EC cells, whereas in the sculpted case they were chosen from the subset of 15 environmentally-relevant cells. Representations of each event were composed of the 10 DG cells (and their CA3 followers) that were most richly innervated by the event’s EC cells. After A and B were encoded, the virtual subject was tested with half of event A’s ECin cells active. Top Panel, Excitation of each of 100 DG neurons during encoding in A and B. For each context, the 10 most excited cells became incorporated into the representation. The minimal excitation level (threshold) for incorporation is shown by the broken line; the threshold for A was slightly below that for B, but so close that a single line was drawn. Representation cells are indicated by triangles. Sculpting results in considerably less overlap between the representations of A and B (Rep A and Rep B) and richer innervation by the events’ ECin cells. Bottom Panel, Excitation of CA3 neurons during recall in which a subset of event A’s ECin cells were active. Broken line represents the threshold for action potential firing. Prior to completion of a representation by the CA3 recurrent collateral network, there are substantially more Rep A than Rep B cells active in the sculpted than in the non-sculpted case. Therefore, completion of Rep A rather than B is much better ensured.

Fig. 5. Effects of sculpting in BACON simulations with rat-like parameters.

As the number of events encoded (M) increases, the number of event attributes needed for correct recall increases greatly in the random (non-sculpted) network but increases only slightly in the sculpted network. Calculations were performed as described under Fig. 2. It was assumed that 30% of an event’s 12,500 attributes were encoded and that attribute overlap of one event with another was 90%. Events were constructed by drawings at random from a set consisting of 7% of the total possible set of event attributes; in the sculpted case, DG cells were innervated only by the EC_in cells coding this attribute set.