MNEMONIC REPRESENTATIONS OF TRANSIENT STIMULI AND TEMPORAL SEQUENCES IN THE RODENT DENTATE GYRUS IN VITRO ROBERT A. HYDE

Transcription

1 MNEMONIC REPRESENTATIONS OF TRANSIENT STIMULI AND TEMPORAL SEQUENCES IN THE RODENT DENTATE GYRUS IN VITRO by ROBERT A. HYDE Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy Dissertation Advisor: Ben W. Strowbridge, Ph.D. Department of Neurosciences CASE WESTERN RESERVE UNIVERSITY January 2013

2 CASE WESTERN RESERVE UNIVERSITY SCHOOL OF GRADUATE STUDIES We hereby approve the dissertation of Robert A. Hyde candidate for the doctoral degree*. Diana Kunze, Ph.D. (chair of the committee) Ben Strowbridge, Ph.D. Roberto Galán, Ph.D. Dominique Durand, Ph.D. John Leigh, M.D. September 21, 2012 *We also certify that written approval has been obtained for any proprietary material contained herein. ii

3 Table of Contents Signature Sheet...ii Table of Contents... iii List of Figures... iv Acknowledgements...v List of Abbreviations... vi Abstract...1 Chapter 1: Introduction...4 Memory and the hippocampal formation...4 Anatomy of the hippocampal formation...8 Spatial and episodic memory in the hippocampal formation...16 Working memory in the hippocampal formation...20 Computational models of information storage in the hippocampus...23 The hippocampal formation and contextual memory...26 Chapter 2: Mnemonic representations of transient stimuli and temporal sequences in the rodent hippocampus in vitro...37 Summary...37 Introduction...37 Results...40 The dentate gyrus supports multiple distinct neural representations...40 Origin of recall errors...45 Time course of response separability...47 Dentate gyrus circuits encode temporal sequence order...48 Population representations of biological information in the dentate gyrus...51 Discussion...52 Information representation in the dentate gyrus in vitro...53 Relationship to delay-period activity recorded during working memory tasks...58 Methods...63 Chapter 3: Discussion Mechanisms of transient information storage: cellular processes Mechanisms of transient information storage: network processes Mechanisms of sequential information storage: cellular processes Mechanisms of sequential information storage: neural networks Persistent activity and sequential information: cellular processes Persistent activity and sequential information: neural networks Persistent activity and sequential information storage: synaptic mechanisms Sequential information storage: state-dependent networks A proposed mechanism for information storage in the dentate hilus Bibliography iii

4 List of Figures Figure 1-1 Anatomy of the hippocampal formation...31 Figure 1-2 The dentate gyrus...33 Figure 1-3 Mossy cell association/commissural projection...35 Figure 2-1 Persistent synaptic barrages evoked by multiple PP stimuli...69 Figure 2-2 Short-term hilar representations of multiple PP stimuli...71 Figure 2-3 Example of synaptic barrages evoked by different stimulus locations...73 Figure 2-4 Response classifciation as a function of analysis window duration...75 Figure 2-5 Computing the OVL statistic...77 Figure 2-6 Prediction of stimulus identity from hilar population responses...79 Figure 2-7 Time course of hilar population responses...82 Figure 2-8 Classification accuracy across different times...84 Figure 2-9 Visual display of short-term information storage in the dentate gyrus...86 Figure 2-10 Short-term representations of temporal sequences in hilar neurons...88 Figure 2-11 Comparison of responses to the same stimulus presented in different sequences...91 Figure 2-12 Distinct representations of forward and reverse sequences with 8 s intervals...93 Figure 2-13 Plot of context dependence of responses to sequential stimuli...95 Figure 2-14 Sequence representations are robust to perturbation of stimulus interval...97 Figure 2-15 Population representations of stimulus and sequence identity in the dentate gyrus...99 Figure 3-1 Feedforward and feedback neural networks Figure 3-2 A hypothetical mechanism for maintaining contextual information in the dentate gyrus iv

5 Acknowledgements This dissertation would not have been possible without the help of many people to whom I am truly grateful. Foremost, I would like to thank my advisor, Ben Strowbridge, for providing me the opportunity to work in his lab, encouraging me daily in my scientific endeavors, and challenging me to seek excellence. My thesis committee Diana Kunze, Dominque Durand, Roberto Galán, and John Leigh offered great advice and guidance throughout my graduate career. In particular, I would like to thank Roberto Galán for very helpful discussions on a number of analysis methods. I would also like to acknowledge the strong mentorship provided my former research advisor, Stefan Herlitze, during my first years in graduate school, as well as the helpful advice and support of the members of his lab Melanie Mark, Takashi Maejima, Davina Gutierrez, Eugene Oh, and Viral Shah. Phil Larimer provided helpful encouragement and insights on Chapter 2 as well as on laboratory research in general. Daily work in the Strowbridge lab would not have been as enjoyable as it was without the friendship of the other lab members Loren Schmidt, Ross Anderson, and Isaac Youngstrom. I would like to thank all the members of my family who have supported me in incalculable ways throughout my graduate school career, especially my wife, Mary, whose love and encouragement has carried me for many years. Finally, this dissertation is dedicated to those people in my life that I have had the great fortune to know as teachers. Without their passion to that immortal profession, this work would not have been possible; it is as much a product of their efforts as it is of mine. v

6 List of Abbreviations ACSF: artificial cerebrospinal fluid AHP: after hyperpolarization AMPA: α-amino-3-hydroxy-5-methylisoxazole-4-propionic acid hydrate AP: action potential D-APV {NMDA receptor antagonist}: D-2-amino-5-phosphonovalerate BAPTA {fast, selective Ca 2+ chelator}: 1,2-bis(2-aminophenoxy)ethane-N,N,N',N' tetraacetic acid CA1: cornu ammonis subfield 1 of the hippocampus CA2: cornu ammonis subfield 2 of the hippocampus CA3: cornu ammonis subfield 3 of the hippocampus DG: dentate gyrus of the hippocampus EC: entorhinal cortex EGTA {Ca 2+ chelator}: O,O'-bis(2-aminoethyl)ethyleneglycol-N,N,N',N'-tetraacetic acid EPSC: excitatory postsynaptic current EPSP: excitatory postsynaptic potential fmri: functional magnetic resonance imaging GABA: gamma-aminobutyric acid GC: granule cell IML: inner molecular layer of the dentate gyrus of the hippocampal formation IPSC: inhibitory postsynaptic current IPSP: inhibitory postsynaptic potential MK801 {noncompetitive NMDA receptor antagonist}: (5S,10R)-(+)-5-methyl- 10,11-dihydro-5H-dibenzo[a,d]cycl ohepten-5,10-imine maleate MOPP: molecular layer interneuron cell with its dendritic and axonal domains confined to the perforant path MRI: magnetic resonance imaging NBQX {AMPA/kainate receptor antagonist}: 1,2,3,4-tetrahydro-6-nitro-2,3-dioxobenzo[f]quinoxaline-7-sulfonamide NMDA: N-methyl-D-aspartate NMG {cell-impermeant Na + substitute}: N-methyl-D-glucamine PP: perforant path SGC: semilunar granule cell TEA {potassium channel blocker}: tetraethylammonium vi

7 Mnemonic Representations of Transient Stimuli and Temporal Sequences in the Rodent Dentate Gyrus In Vitro Abstract by ROBERT A. HYDE The brain is constantly engaged in processing, maintaining, and recalling information. The ability to maintain information for a short period of time is known as working memory, and behavioral experiments in vivo have highlighted the role of persistent neural activity in several brain regions, including the hippocampus, in encoding information during working memory tasks. However, how that persistent activity arises in vivo is currently unknown. Recently, it was shown in hippocampal brain slices that persistent activity can be evoked in the dentate gyrus. In response to brief stimulation of the perforant path, semilunar granule cells exhibit intrinsic plateau depolarizations driven by NMDAR and L- and T-type voltage-dependent calcium channel conductances, driving persistent synaptic barrages onto hilar cells. These barrages last for 10s of seconds, resembling the persistent activity observed in vivo during working memory tasks. In this dissertation, I test the hypothesis that this in vitro preparation can maintain information for multiple stimulus locations over 10 seconds, as required 1

8 by neural circuits engaged in vivo during working memory tasks. As the hippocampal formation is known to maintain the identity of information context as well as content, I also test the hypothesis that this preparation is able to maintain representation of stimuli presented in sequences separated by short intervals. Finally, I examine the robustness of this in vitro preparation to small perturbations in sequence interval and test the hypothesis that both stimulus and sequence representations are maintained by a population code. In the first chapter of this dissertation I discuss classic studies on the neurobiology of memory and the anatomical organization of brain regions, especially the hippocampal formation, involved in mnemonic representations of experience. I then discuss how persistent activity has emerged as the neural correlate of working memory, and how experimental and computational studies of persistent activity have sought to uncover its origins in several cellular and network mechanisms. In the last part of the chapter I discuss experimental studies of maintaining the context of sequential information. In the second chapter, l discuss the results of investigating my hypothesis that small cell assemblies of the dentate hilus can maintain the identity of a stimulus. I demonstrate that stimulus representations decay over time, and I show that the network can respond to individual stimuli as well as sequences of stimuli. I discuss results suggesting that persistent activity in the dentate gyrus enables hilar cells to maintain noncommutative representations of sequential stimuli. These results implicate the dentate gyrus as a short-term processor of both individual stimuli as well as stimuli occurring in specific contexts. 2

9 In the final chapter, I discuss what previous computational and experimental studies have indicated might be required of the properties of individual cells and their organization into networks to allow the maintenance of these various forms of information. Finally, in the last section of that chapter I will suggest a hypothetical mechanism whereby the dentate hilus can encode the spatio-temporal patterns of stimuli reported in this dissertation. 3

10 Chapter 1: Introduction Neural circuits are organized to encode, maintain, and retrieve information necessary for behavioral and cognitive demands. Despite intensive investigation in a number of brain areas, it has been difficult to address how the properties of a local circuit give it its distinct function (Lee and Reid, 2011). In an area of the brain such as the hippocampus, whose role in the consolidation of declarative memory has long been recognized (Scoville and Milner, 1957; Squire et al., 2004; Squire and Wixted, 2011), the study of the physiological properties of individual cells, their connectivity to other cells within the network, and how their properties may be modified all underlie efforts to understand how the hippocampus functions as a mnemonic system. In this thesis I use an in vitro experimental approach to explore how the intrinsic cell properties and connectivity of a subfield of the hippocampus, the dentate gyrus, enable the circuit to maintain short-term representations of several categories of information. Memory and the hippocampal formation In both awake and sleeping states, the adult human brain is constantly engaged in recalling, storing, and consolidating experiences. The formation and accurate recall of such information for daily demands constitutes memory (Baddeley, 2010). Research on the anatomical and cellular systems underlying memory has been pursued with great interest over many decades, and it has become clear that not only are there a variety of memory systems in the brain, but that specific 4

11 brain regions and cellular mechanisms are responsible for encoding different categories of information (Scoville and Milner, 1957; Brunelli et al., 1976; Constantinidis and Procyk, 2004; Squire and Wixted, 2011). Memory mechanisms can be distinguished by their temporal extent as well as by their content (Squire et al., 2004; Baddeley, 2010). Short-term, or working memory, refers to the memory for information on a timescale of seconds to hours (Baddeley, 2010), while long-term memory refers to a system for maintaining information over the lifetime of an organism (Kandel, 2009). Memory content can be considered either nondeclarative or declarative (Squire, 1992). Nondeclarative memory refers to the ability to recall or perform procedural skills and habits, and includes simple forms of conditioning, habituation, and sensitization. This kind of memory is very different from the everyday usage of the term memory, which falls under the category of declarative memory. Declarative memory is the memory for facts and events. As a system, declarative memory is representational: remembered material is flexible and can be easily compared and contrasted. Representations are accessible to awareness and may modulate behavioral performance in a variety of contexts (Squire et al., 2004). For many years, it was unclear how memory was organized in the brain. Some researchers, including Karl Lashley, who removed neural tissue to identify the brain regions responsible for the engram, considered that mnemonic functions were widely distributed throughout the brain (Bruce, 2001), while others, including Wilder Penfield and Donald Hebb (Hebb, 1949), hypothesized 5

12 that there was a memory region localized to a specific area of the brain (Squire and Wixted, 2011). One of the important advancements in understanding memory came from the study of a patient, commonly known as H.M., who nearly half a century ago underwent a bilateral medial temporal lobectomy for treatment of epilepsy (Scoville and Milner, 1957). Following his surgery, H.M. became seizure-free but also developed profound anterograde and some retrograde amnesia. However, a vast array of perceptual, motor, cognitive and mnemonic functions remained intact. H.M. was extensively studied over many years and it became clear that the hippocampal formation, as part of the medial temporal lobe, does not serve as a permanent repository for memory but that it plays a time-limited role in memory (Scoville and Milner, 1957). Specifically, the function of the hippocampus has been framed in terms of its necessity for the consolidation of long-term declarative memory. Although H.M. s lesions provided a leap forward in understanding memory, the fact that his lesion included the entire medial temporal lobes precludes attempts to ascribe his memory deficits solely to the loss of his hippocampi (Corkin et al., 1997). More recently, it has become apparent that patients with selective damage to the hippocampus proper exhibit less profound memory deficits than those exhibited by H.M. (Squire and Wixted, 2011). Similar lesion studies in nonhuman primates (Alvarez et al., 1995) are consistent with these findings. Thus, it appears that the hippocampal formation may underlie the storage of certain kinds of declarative memories, but not others. Following Tulving s influential division of declarative memory into semantic and episodic categories (Tulving, 2002), studies in 6

13 patients with well-defined anatomical lesions have elucidated a role for the hippocampal formation in the maintenance of episodic memory far more so than semantic memory (Vargha-Khadem et al., 1997). Semantic memory refers to the conscious memory of specific facts, without any specific autobiographical associations. It is thought that semantic memory may depend on other structures within the medial temporal lobe (Eichenbaum et al., 1996; Vargha- Khadem et al., 1997; Brown and Aggleton, 2001). By contrast, episodic memory is the conscious memory of specific events in one s past (Tulving, 2002). A critical aspect of episodic memory is its transient nature. Although rehearsal or reactivation of an individual event may modulate its memory trace, the actual event happens once. Thus, the neural activity that underlies episodic memory must be able to encode representations of individual experiences rapidly and reliably (Knierim et al., 2006). Further, neuronal representations of episodic memory must encode not only information about a tobe-remembered stimulus, but also its spatiotemporal context (Manns et al., 2007; Paz et al., 2010). It should be noted, however, that while it is clear that all structures of the medial temporal lobe contribute to declarative memory, the distinction between brain regions involved in semantic and episodic memory is still a matter of ongoing debate (Tulving and Markowitsch, 1998; Squire et al., 2004), precluding definitive treatment of the cellular and network properties underlying each of these memory systems. 7

14 Anatomy of the hippocampal formation Before further considering the functional importance of the hippocampal formation in encoding diverse forms of information, it is necessary to address its structure, as its anatomical organization gives rise to a number of predictions about its overall function (Aggleton, 2012). The hippocampal formation is a widely conserved layered structure that is intimately involved in processing cortical information (Acsády and Kali, 2007). It is one of the most heavily investigated regions of the brain, not only for its crucial role in representing memories but also for its exquisite anatomical organization. In rodents, the hippocampal formation arcs in a C-shape curve from its most rostrodorsal extent at the midline of the brain near the septal nuclei, superioposteriorly to the thalamus, and then to its most caudoventral extent in the temporal lobe. The hippocampus (from the Greek for seahorse ) received its name from this characteristic shape, which can be readily recognized in the brains of many animals. Its relatively simple trilaminar organization has also long facilitated its study in neuroscience (Andersen et al., 2006). Neuroanatomically, the hippocampal formation is best appreciated as an extensive loop of interconnected regions that serve to process information coming from, and returning to, the entorhinal cortex (EC). The hippocampal formation itself consists of several distinct regions (Fig. 1-1). Following the nomenclature of Lavenex and Amaral (Andersen et al., 2006), the hippocampal formation consists of the dentate gyrus (DG), the hippocampus proper, 8

15 subiculum, presubiculum, parasubiculum, and entorhinal cortex. The hippocampus proper consists of the CA1, CA2, and CA3 subfields. As the beginning and end point of an extensive loop of information processing, the entorhinal cortex plays a critically important role in the flow of information through the hippocampal formation. Like most of neocortex (Ramón y Cajal et al., 1995), the entorhinal cortex is a laminated structure consisting of multiple cell layers. The entorhinal cortex can be divided into two regions: the lateral entorhinal area (LEA) occupies the rostrolateral portion of the rodent entorhinal cortex, while the medial entorhinal area (MEA) occupies the remaining area (Andersen et al., 2006; van Strien et al., 2009). These anatomical divisions give rise to important functional distinctions, as discussed below (Hargreaves et al., 2005). Layer II stellate and pyramidal cells in the EC give rise to extensive associational projections to other layer II neurons and send a few collaterals to deeper layers of the entorhinal cortex (Witter, 2007). However, their most important contribution to hippocampal circuitry is to the perforant pathway (Fig. 1-1b). The perforant path receives its name from the course it takes from the entorhinal cortex as it courses through the subicular complex and perforates the hippocampal fissure before synapsing onto cells of the hippocampal formation (Andersen et al., 2006). Axons of layer II cells provide the projection to the dentate gyrus and CA3, while axons of layer III cells project to CA1 and subiculum. Considering the spatial extent of this projection from EC to the four major regions of the HF, it is clear that the perforant pathway is likely the most 9

16 important means of relaying information from entorhinal cortex onto the hippocampal formation (Witter, 2007). It is not the exclusive one, however, as entorhinal fibers can also reach CA1 via the alveus (Andersen et al., 2006). In the dentate gyrus, perforant path fibers synapse onto the dendrites of granule cells, the principal cell type of the dentate gyrus. Much like the subfields of the hippocampus proper, the dentate gyrus consists of a principal cell layer (the granule cell layer) interposed between two adjacent regions (the molecular layer and the hilus, Fig. 1-2). Granule cells are polarized glutamatergic neurons whose dendrites extend into the molecular layer and whose axons arborize within the hilus (Amaral et al., 2007). Granule cell axons are known as mossy fibers and they form excitatory synapses onto CA3 pyramidal cells (Claiborne et al., 1986), hilar interneurons (Acsády et al., 1998), and mossy cells (Scharfman et al., 1990). The molecular layer consists primarily of granule cell dendrites, perforant path fibers of passage, and a relatively small population of inhibitory interneurons. The inner molecular layer, however, contains another type of glutamatergic cell type, the semilunar granule cell (Williams et al., 2007). This type of cell was originally described by Ramón y Cajal (Ramón y Cajal et al., 1995) and has only recently been re-discovered (Williams et al., 2007; Gupta et al., 2012). It will be discussed at length in chapters 2 and 3 of this thesis. The third region of the dentate gyrus, the hilus, consists of a glutamatergic cell type, the mossy cell, as well a number of hilar interneurons (Amaral, 1978). Mossy cells receive glutamatergic inputs from granule cells (Acsády et al., 1998), semilunar granule cells (Williams et al., 2007), as well as GABAergic inputs from 10

17 hilar interneurons (Scharfman, 1995; Larimer and Strowbridge, 2008). Mossy cells innervate hilar interneurons (Scharfman, 1995; Larimer and Strowbridge, 2008) and also project axons as the associational/commissural pathway (Fig. 1-3) to distant ipsilateral and contralateral lamella, where they form excitatory synapses onto the proximal dendrites of granule cells (Buckmaster et al., 1996). An unusual characteristic of the dentate gyrus is its abundance of interneurons. There are over 20 different subtypes of interneurons based on their morphological characterization in the hilus alone (Amaral, 1978). There are five major types of GABAergic interneurons that synapse onto granule cells: hilar interneurons distributed in conjunction with the termination field of the perforant path (HIPP), hilar interneurons in the termination field of the commissural and associational pathway (HICAP), dentate basket cells (Ribak and Shapiro, 2007), interneurons with somata in the molecular layer and axons in the termination field of the perforant path (MOPP), and axo-axonic cells (Halasy and Somogyi, 1993). It has been hypothesized that the diversity of types of interneurons provides a means to regulate the gain and potentiation of individual excitatory inputs (Halasy and Somogyi, 1993). Basket cells, in particular, have been implicated in pattern separation, long considered a primary function of the dentate gyrus (discussed below), as well as temporal lobe epilepsy (Sloviter, 1991). As will be clarified in chapter 3 of this dissertation, the presence of an abundant variety of hilar interneurons has important functional consequences for information processing in the dentate gyrus. 11

18 The next regions in the flow of hippocampal information are the subfields of the hippocampus proper: CA3, CA2 and CA1. In all of these subfields, the principal cell layer is the pyramidal cell layer. CA3 is much larger in area than CA2, which is a very narrow region characterized by pyramidal cells that lack thorny excrescences on their proximal dendrites. Located deep to this layer is stratum oriens, which contains the basal dendrites of the pyramidal cells and some classes of interneurons, as well as some CA3 recurrent collaterals. Deep to the stratum oriens is a thin zone known as the alveus, which carries a number of fibers to other cortical and subcortical structures (Andersen et al., 2006). Superficial to the pyramidal cell layer in CA3 is stratum lucidum, a largely acellular region containing the mossy fibers. It is absent in CA2 and CA1. Superficial to stratum lucidum is stratum radiatum, the site of the well-known CA3-CA3 associational connections. Detailed analysis of these recurrent connections using intracellular labeling techniques has demonstrated that a single CA3 pyramidal neuron may contain as many as 11,000 synaptic boutons with other CA3 pyramidal cells (Sík et al., 1993; Wittner et al., 2007). Finally, stratum lacunosum-moleculare comprises the most superficial layer and is the stratum where PP fibers terminate. As in the dentate gyrus, there are a number of interneurons scattered throughout all these layers (Freund and Buzsaki, 1996); these inhibitory neurons receive synaptic input from fibers that also innervate the apical or basal dendrites of pyramidal cells in those layers. Although the vast majority of excitatory fibers of CA3 synapse onto CA1 pyramidal cells or onto CA3 cells as recurrent collaterals (Andersen et al., 2006; Wittner et al., 2007), 12

19 there is also evidence for a backprojection from CA3 pyramidal cells (especially those located proximal to the dentate gyrus) onto hilar cells (Scharfman, 1994), which may have important implications for hippocampal processing (Scharfman, 2007). The remaining subfield of the hippocampus, CA1, is similar in structure to CA3/CA2. Like CA2, it lacks a stratum lucidum, but its pyramidal cell layer is more tightly packed. CA1 pyramidal cells tends to be smaller and their dendritic organization tends to be more homogeneous than that of CA3/CA2 cells (Ramón y Cajal et al., 1995; Pyapali et al., 1998). The Schaffer collaterals, well-known for their role in studies of long-term potentiation (Nicoll et al., 1988; Kullmann and Nicoll, 1992), form synapses onto the dendrites of CA1 pyramidal cells in stratum radiatum. Unlike in CA3, there are no documented associational projection between CA1 pyramidal cells. Axons of CA1 pyramidal cells travel in the alveus or stratum oriens toward the adjacent subiculum and to the deep layers of the entorhinal cortex (Andersen et al., 2006). The subicular complex (consisting of subiculum, presubiculum, and parasubiculum) is the final stage in the flow of information in the hippocampal formation. Subicular projections to layer I of presubiculum and layers I and II of parasubiculum comprise a minor output pathway, while most neurons project to the deep layers of entorhinal cortex, thus completing an anatomical loop originating in the more superficial layers of the EC (Andersen et al., 2006). Among almost all regions of the hippocampal formation, projections between regions are relatively well organized into a three-dimensional 13

20 topographical arrangement. In the DG and CA3, for example, perforant path fibers terminate in the outer two-thirds of the molecular layer and stratum lacunosum moleculare, respectively, with fibers originating from the lateral EC terminating in the more superficial layers of this zone and fibers originating from the medial EC terminating in the deeper layers (Witter, 1993; van Strien et al., 2009). There are conflicting reports on the transverse distribution of the layer II PP projection, however. Some studies indicate that in rats, the lateral PP projects preferentially to the suprapyramidal blade of the DG while the medial PP either does not show a preference or predominantly terminates in the infrapyramidal blade (Hargreaves et al., 2005; Witter, 2007). CA3 cells located closest to the dentate gyrus project more heavily to CA1 cell located septal to their location (Andersen et al., 2006). In CA1, proximally located pyramidal cells project to the distal portion of subiculum, whereas distally located CA1 cells project to more proximal portions of subiculum. The projections from CA1 and subiculum to the entorhinal cortex are organized topographically (Andersen et al., 2006). Although far less understood than the hippocampus proper, the subiculum appears to possess a different cellular organization than that of CA3 or CA1: there is a crude columnar and laminar organization. In addition, there is a lamellar organization to the hippocampus, as it had been observed that transverse slices ( lamellae ) obtained for electrophysiological studies (Andersen et al., 1971; Witter, 2007) preserve much of the connections between the four major regions of the hippocampal formation. This so-called trisynaptic circuit (Fig. 1-1b) consists of synapses between EC and DG, between DG and CA3, 14

21 and between CA3 and CA1 (Andersen et al., 1971; Acsády et al., 1998). However, as has been mentioned elsewhere, there are major projections from CA1 to subiculum and EC, as well as connections between DG and distant lamella, and although a useful description for the major connections in the hippocampus, the trisynaptic circuit model fails to capture much of the threedimensional organization critical to the function of the hippocampus. A number of observations on the connectivity of the hippocampal formation merit attention. The hippocampal formation exhibits a high degree of convergence at the same time that it also exhibits a high degree of divergence. In the transverse plane of the HF, for example, there is an extraordinary convergence of granule cells onto mossy cells in the dentate gyrus, but mossy cells extend associational/commissural fibers throughout the septotemporal extent of the dentate gyrus (Amaral et al., 2007). This divergence of mossy cell activity enables information at a single septotemporal level to influence more than 50% of the entire formation (Amaral and Witter, 1989; Buckmaster et al., 1996). It is clear that there is not only a serial or sequential unidirectional flow of information from one region to the next, but that there are parallel pathways as well (Acsády et al., 1998; Witter et al., 2000). For example, PP fibers synapse onto CA3 pyramidal cells monosynaptically but there is also a disynaptic connection through the mossy fibers of granule cells. There are also prominent associational connections in the DG, hippocampus, and entorhinal cortex, raising the possibility of polysynaptic activation. These circuit features suggest that each region does not depend exclusively on an immediately preceding region for input, 15

22 but rather that the regions function in concert with, but also partly independent of, each other (Witter et al., 2000). Spatial and episodic memory in the hippocampal formation Although the anatomical structure of the hippocampal formation suggests that it is able to process cortical information through a highly organized and complex network of distinct regions, it has been difficult to address experimentally the processing strategies of the hippocampal formation in forming episodic memories. Lesion studies in humans (Scoville and Milner, 1957; Vargha- Khadem et al., 1997) and nonhuman primates (Mishkin, 1982; Zola-Morgan and Squire, 1990; Alvarez et al., 1995) have indicated that the hippocampal formation is required for episodic memory, but these studies lack the functional resolution offered by electrophysiological studies in rodents. However, there are intrinsic challenges in evaluating the role the hippocampal formation in episodic memory in nonverbal animals (Tulving and Markowitsch, 1998; Tulving, 2002). In this respect, studies in rodents that have primarily explored the role of the hippocampal formation in spatial tasks have been instrumental in understanding its ability to encode experiences. Since place cells in the hippocampus were recognized (O'Keefe and Dostrovsky, 1971) in freely moving rats, a number of studies have investigated the activity of cells in the hippocampus during spatial exploration (Moser et al., 2008). Place cells are a type of cell that exhibit preferential firing in a relatively small area (the place field) of an animal s local environment, such as a test arena in laboratory studies. It has been shown that 16

23 neighboring place cells fire at distinct locations in the environment such that the entire environment is represented in the activity of these populations of cells (O'Keefe, 1976; Wilson and McNaughton, 1993). There is now evidence that such cells exist in all subfields of the hippocampal formation, although the majority of them are CA3 and CA1 pyramidal neurons (Barnes et al., 1990). The discovery of cells in other regions of the hippocampal formation and parahippocampal regions that exhibit activity in response to specific spatial stimuli has indicated that distinct regions differentially contribute to encoding information about the environment. In the entorhinal cortex, for example, there exist cells with multiple firing fields that form a triangular array, or grid, that tiles the animal s entire environment (Fyhn, 2004; Hafting et al., 2005). Similarly, cells in the parasubiculum show preferential activity to the head direction of an animal (Taube et al., 1990; Taube, 1998). There is thus an extended spatial representation system throughout the rodent brain, of which the hippocampal formation plays a critical part. Interestingly, it has also been reported that in addition to maintaining representations of a spatial environment, the rodent hippocampus and surrounding structures are critically involved in representing nonspatial information as well (Bunsey and Eichenbaum, 1996). In one study, rats were trained on a odor recognition task while single unit recordings were obtained in CA1 and CA3. The rats were trained to dig in sand for a buried cup containing a specific odor. The cup location and odor were independently varied, thereby permitting distinctions in unit firing activity to be discriminated on the basis of 17

24 these two variables. Of the 127 units analyzed, approximately 40% showed activity that was only correlated to odor identity, while 31% showed activity that depended on cup position. In addition, a proportion of the position-correlated cells showed responses to odors as well (Wood et al., 1999). These results and others in rodents have demonstrated that hippocampal neurons can maintain spatial and non-spatial representations, as well as combinations of both categories (Eichenbaum et al., 1999; Hampson et al., 1999). Similar results have been obtained in non-human primates (Doré et al., 1998) and humans (Squire et al., 2001; Levy et al., 2004). Such studies have prompted investigators to view spatial memory as a subset of a broader category of declarative memory functions served by the hippocampus and surrounding structures (Squire, 1992; Leutgeb et al., 2005; Moser et al., 2008). This view is clearly supported by anatomical studies indicating that the hippocampus receives information (via the entorhinal cortex) from nearly all regions of the cortex. The medial entorhinal cortex predominantly receives visual and spatial information from the perirhinal and postrhinal cortices (Burwell and Amaral, 1998), while the lateral entorhinal cortex predominantly receives olfactory, somatosensory and auditory information from the perirhinal cortex (Burwell and Amaral, 1998; Hargreaves et al., 2005). A similar dissociation is also present in the ventral and dorsal hippocampus, which have been shown to serve different functions. Lesions to the dorsal hippocampus impair performance on spatial memory tasks (Fanselow and Dong, 2010) while lesions to the ventral hippocampus have minimal effect. By contrast, lesions to the ventral 18

25 hippocampus impair performance on fear-conditioning (Czerniawski et al., 2009; 2011). This dissociation is consistent with other observations that the dorsal hippocampus is more important in processing spatial information while the ventral hippocampus is more important for representing the internal state (e.g. hunger, satiety, fear, etc.) of an animal (Kjelstrup et al., 2002; Bannerman et al., 2003; 2004; Fanselow and Dong, 2010). Thus the hippocampus is organized in such as way as to receive inputs from, and sends outputs to, brain regions implicated in a number of sensory systems, endowing it with the complex task of processing polysensory information. The anatomical organization of the hippocampal formation, together with its clear role in processing polysensory spatial and non-spatial information suggests that it is well-suited to combining disjoint sensory information into conjunctive representations that may then be propagated back to the cortex and subcortical structures. In this respect the hippocampal formation can be considered as a region for integrating information via parallel processing streams (Aggleton, 2012). The incorporation and integration of information from a variety of cortical regions suggest a mechanism for representing the experiences underlying a system of episodic memory (Dickerson and Eichenbaum, 2009; Devito and Eichenbaum, 2010). While behavioral studies combined with physiological recordings may help substantiate this theory of hippocampal function, the absence of a tractable model of mnemonic information processing has made it difficult to test how the hippocampal circuitry is able to form representations of combinations of stimulus inputs. 19

26 Working memory in the hippocampal formation In addition to its mnemonic role in maintaining information of the declarative type, the hippocampus also plays a distinct role in short-term, or working memory. Working memory refers to a brain system that provides temporary online storage and manipulation of information necessary for complex cognitive tasks such as learning and reasoning (Baddeley, 1992). The current conceptual framework of working memory recognizes several distinct modules for performing specific information processing routines, including a visual-spatial sketch pad, a phonological loop, and an episodic buffer under the control of a central executive (Baddeley, 2010). In a landmark study of working memory in non-human primates, it was shown that neurons in the prefrontal cortex exhibit preferential activity during the delay period of a delayed response task (Fuster and Alexander, 1971). Subsequent studies in nonhuman primates have confirmed the role of prefrontal cortical activity in a variety of behavioral tests of working memory (Kubota and Niki, 1971; Funahashi et al., 1989; 1993). It has also become apparent that although prefrontal cortex clearly exhibits activity related to the correct performance of these behavioral tasks, there are multiple working memory anatomical domains, not only within prefrontal cortex but in other regions as well, each with specific processing and content-specific storage mechanisms (Goldman-Rakic, 1988; Miyashita and Chang, 1988; Funahashi et al., 1989; Colombo and Gross, 1994; Goldman-Rakic, 1996). Indeed, it has been shown in nonhuman primates that neurons of the hippocampal formation also exhibit 20

27 delay-period activity during both nonspatial (Watanabe and Niki, 1985) and spatial (Colombo and Gross, 1994) working memory tasks. In particular, in a study of rhesus monkey using 2-deoxyglucose as a marker of metabolic activity found robust and selective signal enhancement in the molecular layer and granule cell layer of the dentate gyrus, as well as in stratum radiatum and stratum lacunosum-moleculare of CA3 and CA1, during delay matching to sample tasks (Friedman and Goldman-Rakic, 1988). These results suggest that specific circuits within the hippocampal formation are active during working memory, highlighting the need for detailed studies of the network and cellular mechanisms underlying this mnemonic process. In particular, delay period activity is typically extinguished by a behavioral response, but it is currently unclear what mechanisms (e.g., inhibition) specifically attenuate delay-period activity to their pre-stimulus level. It is also unclear how long self-sustaining activity in working memory may last in response to transient sensory stimuli (Fuster and Alexander, 1971; Funahashi et al., 1989). Unlike studies of episodic memory, working memory studies have been comparatively well-established both in primates and rodents. As in primates, there is a general view that working memory involves a number of anatomically distinct cortical and subcortical structures (Goldman-Rakic, 1988; Constantinidis and Procyk, 2004). In rodents there is evidence that both prefrontal cortex (Batuev et al., 1990; Baeg et al., 2003) and hippocampus (Hampson et al., 1993; Hock and Bunsey, 1998; Maruki et al., 2001) are recruited during working memory tasks. Moreover, there is evidence that these structures serve different 21

28 aspects of working memory (Lee and Kesner, 2003; Yoon et al., 2008), suggesting that working memory in the rodent, as in the primate, involves not only the prefrontal cortex but the hippocampal formation as well (Fell and Axmacher, 2011). Although lesions to the prefrontal cortex certainly impair working memory in rodents and primates, working memory function is not completely abolished. (Yoon et al., 2008). This view is also supported by studies of human patients with lesions of the prefrontal cortex (Ferreira et al., 1998). Although these patients demonstrate deficits in working memory, there is a significant preservation of working memory capabilities in both spatial and non-spatial tasks. Similar conclusions have also been found in rodents (Jones, 2002). A neural correlate of working memory is persistent activity. Persistent activity has been found throughout the cortex (Fuster and Alexander, 1971; Fuster and Jervey, 1981; Miyashita and Chang, 1988; Funahashi et al., 1989; Pesaran et al., 2002) and hippocampal formation (Watanabe and Niki, 1985; Zola-Morgan and Squire, 1990; Colombo and Gross, 1994), including the dentate gyrus (Watanabe and Niki, 1985), in monkeys performing a working memory task. Similar delay-period activity has also been found in cortex and hippocampus of rodents performing a working memory task (Batuev et al., 1990; Hampson et al., 1993; Baeg et al., 2003; Erlich et al., 2011). Thus, the presence of persistent activity in anatomically defined regions known to be critical for working memory suggest that such neural activity may underlie the maintenance of attended information. 22

29 Computational models of information storage in the hippocampus Given its central role in both declarative and working memory systems, a number of studies have sought to address how the circuitry of the hippocampus enables it to maintain abstract representations of polysensory information. By far the most influential work in this respect was that of David Marr, who in 1971 proposed a conceptual model of the hippocampus as a content-addressable memory system (Marr, 1971). Marr contended that spatiotemporal patterns of information arriving from neocortex could be stored in the modifiable connections between cells in the hippocampus. When a partial version of a stored pattern is presented later, activity propagating along the previously strengthened pathways would reinstate the complete pattern. This process of reconstructing the complete pattern from partial information is known as pattern completion. Critical to Marr s theory is his requirement of two intermediate layers (viz., granule cell mossy fibers and CA3 pyramidal cells) as well as collateral fibers (viz., recurrent collaterals of CA3) to form simple representations of inputs (Willshaw and Buckingham, 1990). In Marr s view, recurrent connections (the collateral effect ) between CA3 pyramidal cells enable that subfield to perform pattern completion. Marr s ideas were later expanded further into computational models in a number of theoretical studies (McNaughton and Morris, 1987; Rolls, 2007; 2010). Critical to many of these models is the conception of CA3 function as an autoassociative network (Hopfield, 1982). As many synapses in the hippocampus show associative modification by long-term potentiation (Nicoll et al., 1988; Kullmann and Nicoll, 1992), and as these associative changes are thought to be critical to 23

30 learning, inputs that are preferentially strengthened by certain patterns of activity may be able to encode specific memories. CA3 is particularly well-poised to generate those representations as it receives inputs from EC both directly (via the PP) and indirectly (via the mossy fibers). PP fibers provide numerous but relatively weak connections onto a single CA3 pyramidal cell, while mossy fibers provide powerful but sparse synaptic connections. In response to a given stimulus pattern in EC, mossy fibers act as a teaching input that selectively strengthens the synaptic connections between CA3 pyramidal neurons involved in the representation (Rolls, 2007). Thus, when a partial portion ( the cue ) of the pattern is subsequently presented, the direct PP connection will be sufficiently able to reinstate the full representation via the recurrent collaterals between CA3 pyramidal neurons (Sík et al., 1993). This model has found support in experimental studies (Nakazawa et al., 2003). According to Marr, in order for CA3 to function well as a contentaddressable memory system, it needs to be able to maintain or process a large capacity of information it needs to have a high capacity of memories. Given the finite size and number of cells in the hippocampus, it would be advantageous if incoming information were sparsened. Similar inputs would need to be disambiguated ( pattern separation ). One way that such disambiguation could occur is to have fewer cells maintain the representation of a given input. Experimental studies in rodents has shown that, following exposure to a novel environment, only 2% of granule cells exhibited activity as measured by the immediate early gene Arc (Jung and McNaughton, 1993; Chawla et al., 2005). 24

31 When the animal was placed in a different novel environment, however, a second distinct, but similarly-sized subset of granule cells, were positive for the Arc transcript (Chawla et al., 2005). Presumably the increased expression of the Arc transcript is driven by Ca 2+ entry during bursts of action potentials, although this hypothesis was not tested in the study. Although Arc enhanced expression argues that sparse activation of granule cells may provide an useful means of achieving pattern separation, these studies cannot verify whether granule cells are inhibited by feedforward or feedback pathways or whether granule cells activity is attenuated by other, possibly cell autonomous, mechanisms. Presumably physiological studies in brain slices can help clarify these issues. In addition, physiological studies can clarify the kind of activity present (e.g., EPSPs, single APs, burst of APs) in granule cells during pattern separation, while immediate early gene studies cannot (Scharfman et al., 1990; Williams et al., 2007). Another way of disambiguating overlapping inputs to the DG, however, is decorrelation. A computational model of the dentate gyrus has shown that hilar cells may be well-suited to this function. In this model (Myers and Scharfman, 2009), feedback projections from hilar cells in response to PP inputs are critical to distinguishing similar inputs. Because a single pattern of PP activity leads to excitation of granule cells that, in turn, results in activation of mossy cells projecting to granule cells in distant lamellae, similar inputs within the same lamella would result in activation of granule cells that are more spatially separated along the septotemporal axis. At the same time, hilar interneurons 25

32 excited by the same stimulus pattern provide feedback inhibition onto neighboring granule cells such that only those granule cells with the strongest excitatory connections would be activated. In this way, the dentate gyrus could function as a pre-processor of information coming from the entorhinal cortex (Myers and Scharfman, 2009). While this prediction of hilar cell function from theory has also received experimental support in several studies (Leutgeb and Moser, 2007), there is a longstanding debate in the field of temporal lobe epilepsy regarding the function of hilar cells. One view is that, in vivo, mossy cells provide disynaptic feedback inhibition to granule cells via basket cells, and that in temporal lobe epilepsy there is a loss of mossy cells, leading to a reduction in excitatory currents onto basket cells. Thus the basket cells become dormant, unable to attentuate granule cell hyperexcitatability (Sloviter, 1991). On the other hand, there is a view that in temporal lobe epilespy there is a loss of hilar interneurons and therefore granule cell hyperexcitability is due to increased excitatory drive directly from mossy cells (Dudek and Sutula, 2007). The hippocampal formation and contextual memory The aforementioned studies have highlighted the role of the hippocampus in maintaining spatial, episodic, and short-term representations of information and how the anatomical substrate of the hippocampal formation may permit it to process a variety of information from incoming cortical regions. As a processor of polysensory information, the hippocampal formation is thought to recombine and associate multiple sensory representations into newer constructs that can 26

33 then be maintained in downstream cortical regions. A current view of the function of the hippocampal formation is that it can thereby form a conjunctive spatial-nonspatial code for representation of experience (Vinogradova, 2001; Moser et al., 2008). In rodents, for instance, the separation of spatial and nonspatial information is maintained in the parahippocampal region until it is combined in the hippocampus (Dickerson and Eichenbaum, 2009). There has been a growing appreciation that the hippocampus role in the formation of memories may depend primarily on its role in binding sequential experiences into content-addressable representations (Tulving and Markowitsch, 1998). One of the implications of such a theory is that the hippocampus is able to generate a general a representation of context (Dickerson and Eichenbaum, 2009). This contextual encoding would be revealed not only in the static encoding of combinations of stimuli but also in the temporal structure of information presented sequentially over time; thus a major function of the hippocampus would be its ability to maintain representations of temporal sequences (Agster et al., 2002; Lehn et al., 2009). Consistent with this theory, a number of studies have demonstrated a critical role of the hippocampus for the temporal order of spatial items as well as a nonspatial items (Kesner and Novak, 1982; Chiba et al., 1994; Agster et al., 2002; Fortin et al., 2002; Komorowski et al., 2009). In one study, rats were trained on an eight-arm radial maze to examine the memory for the temporal order of arm locations. Rats were allowed to visit a random sequence of arm locations and then tested for their ability to recall which one of two arm locations 27

34 had been visited earlier in the study. Control animals were able to recall accurately the arm locations even when the animal had visited as many as six arm locations between the test arm and final arm in the sequence. However, animals with hippocampal lesions performed at chance (Chiba et al., 1994), suggesting that the hippocampal formation was required for the accurate recall of the temporal sequence of spatial locations. In other studies, rats were trained to recognize a series of five randomly selected odors. Memory for each series was then probed in a single-choice test in which a reward was given to the animal for selecting an odor that had appeared earlier in the series (Fortin et al., 2002; Kesner et al., 2002). Animals that subsequently received lesions to the hippocampal formation performed at near-chance levels compared to control animals. However, lesioned animals performed as well as control animals in selecting the odor that was not present in the series, suggesting that recognition of the odor was unimpaired (Fortin et al., 2002), while memory for temporal order of items in the sequence was extinguished. A study in human patients with intractable epilepsy has elucidated the role of hippocampal neurons in maintaining contextual memories for the order of temporal events (Paz et al., 2010). In this study, extracellular electrodes were implanted in the brain and single units isolated while the patients were presented with repeated presentations of cinematic episodes. Images were repeatedly presented under the hypothesis that, as the clip is seen over multiple trials, there is a gradual formation of memory for the temporal order of events within each 28

35 clip. The investigators of this study were able to show that neurons in the hippocampus, but not in the amygdala or cortex, showed increased correlated activity with repeated presentations of a series of episodes and that this correlated activity emerged after only two or three repetitions. Thus, hippocampal cell assemblies were able to encode rapidly the temporal relationships between stimuli within a given episode and maintain that activity over the duration of the clip (Paz et al., 2010). In a related study in humans using fmri, subjects watched a novel movie and later were asked to rearrange scenes from the movie in the order in which they had occurred. A control group organized the scenes by logically inferring their order. In the experimental group, but not the control group, there was enhanced signal in the right hippocampal formation that was positively correlated with the accuracy of sequence recall (Lehn et al., 2009). In humans, as in other animals, several experiments thus support the role of the hippocampal formation in maintaining the spatiotemporal context of experienced information. As has been mentioned earlier, however, the absence of a tractable model system for investigating the cellular and network properties that may allow the hippocampal formation to maintain information at short time scales and encode the context of sequentially ordered stimuli has made it difficult to understand its complex functions. In this respect a recent report that the dentate gyrus may be critical for discriminating spatial context of objects in rodents suggests that it may be a region of the hippocampal formation suitable for study in physiological studies of maintaining contextual information (Lee and Solivan, 2010). In that 29

36 study, rats encountered several objects in different locations and were rewarded for discriminating correct pairs of object and places. Different objects were placed in two locations of a radial arm maze and the distance between objects systematically varied. Rats with colchicine injections to the dentate gyrus performed poorly when the objects were in neighboring arms, whereas objects located in distant arms produced results more similar to control animals. These results highlight the role of the dentate gyrus is maintaining accurate representations of information presented with overlapping spatial contexts (Lee and Solivan, 2010). 30

37 Figure 1-1 Anatomy of the hippocampal formation a, Horizontal section (lamella) of the hippocampal formation. The major regions of the hippocampal formation include the entorhinal cortex (EC), dentate gyrus (DG), cornu ammonis 3 (CA3), cornu ammonis 1 (CA1), and the subiculum (S). b, The hippocampal circuit. EC layer II and layer III neurons project as the perforant path to the DG, CA3, CA1 and S. The canonical tri-synaptic circuit (Andersen et al., 1971) is shown in red. 31

38 a CA1 CA3 DG S EC b 32

39 Figure 1-2 The dentate gyrus The dentate gyrus consists of three regions: the granule cell layer (GCL); the molecular layer (ML); and the hilus. The somata of granule cells (blue) reside within the GCL. Mossy cells (MC, red) and hilar interneurons (HI, black) reside within the hilus. 33

40 34

41 Figure 1-3 Mossy cell association/commissural projection Mossy cells of the dentate hilus influence information processing along the septotemporal axis of the dentate gyrus by projecting to granule cells in distant ipsilateral and contralateral lamellae. Mossy cell (red), granule cell (blue). 35

42 Septal Temporal 36

43 Chapter 2: Mnemonic representations of transient stimuli and temporal sequences in the rodent hippocampus in vitro Summary A primary function of the brain is to store and retrieve information. Except for working memory, where extracellular recordings demonstrate persistent discharges during delay-response tasks, it has been difficult to link memories with changes in individual neurons or specific synaptic connections. Here, we demonstrate that transient stimuli are reliably encoded in the ongoing activity of brain tissue in vitro. We found that the patterns of synaptic input onto dentate hilar neurons predict which of four pathways were stimulated with an accuracy of 76% and performed significantly better than chance for >15 s. Dentate gyrus neurons also could accurately encode temporal sequences using population representations that were robust to variation in sequence interval. These results demonstrate direct neural encoding of temporal sequences in the spontaneous activity of brain tissue and suggest a novel local circuit mechanism that may contribute to diverse forms of short-term memory. Introduction A fundamental property of the central nervous system is the ability to encode and retrieve information. In mammals, declarative memory function is typically divided into behavioral tasks that promote short- or long-term storage of items such as names, places and specific temporal sequences (Fortin et al., 2002; Averbeck and Lee, 2007). While the specific cellular origin of individual long- 37

44 term declarative memories has remained elusive, a large literature highlights the importance of several critical brain regions, including the prefrontal cortex (Funahashi et al., 1989) and the hippocampal formation (Colombo and Gross, 1994) for encoding short-term, or working, memories. Extracellular unit recordings in these brain areas often display periods of persistent spiking activity at elevated frequencies when animals are required to retain transiently presented sensory information. This delay-period activity typically lasts for seconds and is extinguished when the animal initiates a behavioral response to indicate whether it correctly remembered the transient stimulus (Fuster and Jervey, 1981; Funahashi et al., 1989). The reduction in persistent spiking during error trials trials in which the animal made an incorrect behavioral response following the delay period argues that delay-period activity in those specific neurons reflects activity in neural circuits encoding that short-term memory, rather than a response to the cue stimulus or a motor plan associated with the behavioral output. These classic studies in primates, along with more recent reports in rodents (Batuev et al., 1990; Hampson et al., 1993; Baeg et al., 2003), remain some of the clearest examples in which the spiking activity of individual neurons can be directly tied to specific memories. While abundant evidence relates delay-period persistent spiking activity to short-term memories, the underlying cellular mechanisms that mediate this firing mode have not been determined. Donald Hebb postulated in 1949 (Hebb, 1949) that short-term memories could be maintained by reverberant activity that circulates through networks of interconnected excitatory neurons. Indeed, 38

45 transient stimuli can trigger persistent spiking activity in artificial neural networks that contain recurrent excitatory connections (Seung, 1996; Camperi and Wang, 1998; Compte et al., 2000; Miller et al., 2003), although maintaining stable firing frequencies in individual circuit elements during the storage operation requires highly precise tuning of synaptic connection strengths (Seung, 1996; Seung et al., 2000; Mensh et al., 2004). Some types of CNS neurons also contain constellations of intrinsic conductance that enable them to maintain simple memories by firing persistently for several seconds following transient depolarizing stimuli even when pharmacologically isolated from all other neurons (Egorov et al., 2002; Pressler and Strowbridge, 2006). Computational models that combine individual simulated neurons with intrinsic persistence and recurrent excitatory connections are capable of generating persistent firing modes without the high connection weight tuning precision required to enable simple associative network models to fire persistently (Camperi and Wang, 1998; Seung et al., 2000; Koulakov et al., 2002; Goldman et al., 2003). While biological experiments have demonstrated examples of both intrinsic persistence and recurrent excitatory connections in different brain regions, the absence of a tractable in vitro system capable of short-term information storage has limited the opportunities to determine which specific mechanisms are required to generate short-term mnemonic representations of transient stimuli. Here, we demonstrate that four distinct patterns can be reliably encoded within the spontaneous synaptic activity in conventional rodent hippocampal slice preparations. Each pattern is evoked by briefly activating a different subset 39

46 of entorhinal input axons (the perforant path, PP) using an array stimulation electrode; information is read out by recording intracellularly from two or three dentate hilar neurons, which sample the activity of both dentate granule cells (Acsády et al., 1998) and semilunar granule cells (SGCs; Fig. 2-1a), a recentlydiscovered excitatory cell type that responds to PP stimulation with graded depolarizing plateau potentials (Williams et al., 2007; Larimer and Strowbridge, 2010; Gupta et al., 2012). We find that synaptic barrages evoked in downstream hilar neurons by persistent firing in SGCs reliably encode both the identity of individual stimuli and temporal sequences of PP stimuli. Decoding both individual stimuli and temporal sequences relied on population representations of synaptic inputs to dentate hilar cells. Short-term encoding of sequences was robust to perturbation of sequence interval, suggesting that contextual coding in these experiments arises from stimulus order rather than the delay between stimuli. While several in vivo recording studies suggested that neurons in the hippocampal formation are involved in short-term memory function (Hampson et al., 1993; Colombo and Gross, 1994), these results provide the first demonstration that neural networks within the dentate gyrus are capable of encoding both multiple transient inputs and context-dependent information such as the order of temporal sequences. Results The dentate gyrus supports multiple distinct neural representations Transient stimulation of the perforant path (2 x 200 µs shocks) leads to sustained 40

47 increases in the frequency of spontaneous EPSPs recorded in dentate hilar neurons over multiple seconds (Fig. 2-1b), as described in a recent report. That study found that the circuitry contained in a conventional hippocampal slice preparation could encode two distinct activity patterns. Here, we first asked whether more than two patterns could be encoded in the dentate gyrus and what factors govern coding accuracy. Hilar cell responses to stimuli at different PP locations were recorded after placing an array electrode (4 contacts, 115 µm spacing) in the middle molecular layer of the dentate gyrus. Activation of each electrode contact (A, B, C, or D) individually typically elicited synaptic barrages (average EPSP frequency of spontaneous EPSPs in barrage = 18.7 ± 0.5 Hz vs. 3.0 ± 0.1 Hz immediately before PP stimulation; P < ; n = 33 cells). The 3.5 min interval between episodes was significantly longer than synaptic barrages triggered by PP stimuli (decay tau ~8 s, (Larimer and Strowbridge, 2010)) and we did not observe steady changes in baseline EPSP frequency during experiments. While responses evoked by different electrodes in this experiment had different mean EPSP frequencies, the range of EPSP frequencies in each response overlapped (Fig. 2-1c-d) and were only infrequently statistically separable (see below). We next asked if the separation of responses evoked by stimuli at different locations could be increased by sampling the hilar network more densely using multiple simultaneous intracellular recordings (Figs. 2-2b and 2-3). Using this approach, responses to each stimulus in an experiment could be plotted in the 3- dimensional space formed by the frequency of EPSPs recorded in each of three 41

48 hilar neurons during the initial 4 s after the stimulus (Fig. 2-2b). In most experiments, including the example shown in Fig. 2-2b, responses to each stimulus type were clustered in different regions within the 3-dimensional plot of EPSP frequency in each hilar cell. By analyzing synaptic responses of hilar networks in EPSP frequency space, we can directly apply standard statistical and clustering methods along biologically-meaningful dimensions without requiring dimensionality reduction techniques such as principal component analysis. Since the presentation order was randomized, response clustering was unlikely to represent short-term plasticity associated with repeated stimulation at the same location. While many experiments yielded visually distinct response clusters from each stimulus location, as in Fig. 2-2b, linear discriminant analysis (LDA) offers a rigorous method to determine whether specific pairs of responses to different stimulus locations (A/B, B/C, etc) were statistically separable. (We initially tried multiple different separation methods, but LDA offered the most consistent and straightforward results.) Statistical significance on 6 pairwise LDA tests are required to completely separate the responses to all four stimuli (P < for each LDA test, reflecting a Bonferroni correction for multiple comparisons). This criterion was met in 50% of the experiments analyzed (6/12; EPSP frequencies assayed in initial 4 s of each response; Fig. 2-2c), indicating that the ability of the dentate gyrus to generate short-term representations of 4 transient inputs is robust. Responses to three stimuli were statistically separable in half of the remaining 6 experiments (3/12 experiments). On average, we obtained 4.8 ±

49 out of the 6 separation planes required to separate four response patterns, compared with 0.83 planes expected by chance (P < 0.001; n = 12 experiments, 9 triple recordings and 3 paired recordings; see Methods for details). Our ability to statistically separate four patterns was greatest (5.6 ± 0.7 significant planes) when we only considered the 9 experiments with simultaneous triple recordings. Eliminating data from one or two neurons in the 9 triple recording experiments substantially degraded the statistically separate responses to different stimuli (Fig. 2-2d); synaptic barrages recorded from one hilar neuron typically were only able to discriminate between two out of the four response patterns (1.7 ± 1.4 statistically significant planes; Fig. 2-2d). The ability of hilar networks to create mnemonic representations of transient stimuli was relatively independent of the window duration over which EPSP frequency was analyzed (Fig. 2-4), with only a minor (13.8 %) reduction in the number of significant LDA planes obtained with 1 s analysis windows, compared with 4 s windows used in the initial analysis. We next conducted a separate set of triple recordings at relatively depolarized membrane potentials (~-60 mv) using two stimulus locations to ask if stimulus identity also was represented in the spike output of hilar cells. In each experiment in which responses to A and B stimuli were separable using LDA of EPSP frequency across the three cells (n = 3; LDA p values ranged between 2.0 x 10-9 and 2.2 x 10-4 ), we found stimulus-specific suprathreshold responses based on spike frequency assayed in the same 4 s window (LDA p values ranged between and 0.016; mean spike frequency 3.0 Hz following PP stimuli versus 0.06 Hz in 43

50 control conditions). These results demonstrate that both the synaptic input to hilar cells and the resulting spike output can represent stimulus identity. We also found significantly more variation between responses from different stimulus locations than responses evoked by the same location using an omnibus measure of response separation (OVL, 12/12 experiments with p(ovl) < 0.05; Fig. 2-2e and Fig. 2-5) that is used to assay overlapping distributions. Responses to the four stimuli were not separable based on LDA of baseline (prestimulus) EPSP frequencies (mean 1.0 plane; significantly different from the number of planes from actual stimulus responses; P < ). The stimulus intensities used to evoke synaptic barrages were not significantly different on each contact (F = 1.23; P > 0.05; one-way ANOVA; n = 180). These results demonstrate that transient stimuli activating 4 different sets of entorhinal inputs to the dentate gyrus resulted in at least 3 distinct patterns of hilar synaptic activity in a majority of experiments; all four responses were statistically separated in 50% of the experiments (Fig. 2-2c). We next assayed the accuracy of recall based on comparing individual responses to the average response recorded at each stimulus location. We first tested the accuracy of response classification by predicting stimulus identity based on the nearest average response (centroid) computed from the first 4 s of each response (Fig. 2-6a). In this decoding method, if a response triggered by an A stimulus was closest to the average response to all A stimuli, the analysis would generate a correct A prediction. By contrast, if the response triggered by that stimulus was closest to the average response of all C trials in that 44

51 experiment, then an incorrect prediction (that the response was triggered by C instead of the correct answer, A) would be recorded. Over all 12 experiments, this method predicted which one of the 4 possible stimulus locations was activated with an accuracy of 76.1% (n= 180 trials with 137 correct and 43 incorrect predictions). This prediction accuracy is significantly greater than the 25% accuracy expected by chance (S.D. = 3.2%; P < ; also significantly greater than the accuracy obtained following shuffling stimulus identities; P < ; Fig. 2-6b). We did not observe a correlation between the physical distance between pairs of stimulating electrodes tested and the Euclidean distance between the resulting synaptic barrages (Fig. 2-6c). Neither was there a correlation between the inter-electrode distance and the accuracy differentiating hilar responses (e.g., A/B was differentiated as accurately as A/D; Fig. 2-6d). These results demonstrate that the dentate gyrus can generate distinct and repeatable population responses patterns even when nearby stimulation electrodes are activated, arguing that separation of network states in EPSP frequency space is not related to the physical separation between the stimulating electrodes used. Origin of recall errors We next focused on the 43 trials with classification errors to determine the factors that govern accuracy of short-term information storage in the dentate gyrus. In Fig. 2-6e, we plotted the distance between response pairs that were always correctly classified (black symbols) and those that were misidentified at least 45

52 once in an experiment (red symbols). While the range of mean Euclidean distances between responses that were accurately predicted spanned 50 Hz, all classification errors arose from experiments where average responses between pairs of stimuli were separated by less than 30 Hz. The distance between average responses that were misidentified (10.2 ± 1.0 Hz; n = 43) was significantly less than the distance between responses that were always correctly classified (21.6 ± 1.3 Hz; n = 44; P < 0.002). Half of all errors occurred when discriminating between two average responses separated by less than 7 Hz (Fig. 2-6f). While restricting recall tests to trials with well-separated average responses improved accuracy (data not shown), there was a relatively poor correlation between the distance between pairs of average responses and the accuracy differentiating between those responses (R 2 = 0.13; P > 0.05; Fig. 2-6g). This model of variation in accuracy also did not predict the average separation between centroid pairs that were always classified correctly (black symbol). However, scaling the distance between responses by the pooled variance of each response improved the correlation with accuracy (P < ; F = 33.1; Fig. 2-6h). Moreover, this model, based solely on imperfectly classified trials, predicted the scaled distance between perfectly classified trials (black symbol). These results suggest that much of the variability in recall accuracy (62% of the variance) can be explained by a simple two-component model based on the Euclidean distance between average responses and the scatter around those average responses. 46

53 Time course of response separability Synaptic barrages evoked by PP stimuli decayed with an average tau of ~ 8 s and decreased to 39.8% of the initial frequency after 18 s in the present study (n = 112 cells; Fig. 2-7a). When displayed in 3-dimensional EPSP frequency space (Fig. 2-7b), average responses evoked by different stimulating electrodes followed distinct trajectories toward the origin and did not coalesce even though overall EPSP frequency decreased continuously in all responses. We next asked how long hilar responses evoked by different electrodes remain statistically separable and found that hilar networks could still represent information > 4 s after transient molecular layer stimulation, despite decreasing synaptic drive. The mean number of separable responses decreased from 3.25 to 2.25 over 18 s (black symbols in Fig. 2-7c). Three experiments maintained perfect response separation (4 separable responses, 6/6 significant planes; red plot in Fig. 2-7c) for 8 s. Overall classification accuracy decreased to 55.3% at 18 s after stimulation (n = 12 experiments, Fig. 2-7d). The duration representations remained distinct was inversely correlated with initial pattern overlap (Fig. 2-7e; R 2 = 0.62; F = 15.4; P < 0.005), suggesting that hilar networks reliably encode over time stimuli that create highly disparate responses. We also tested whether responses to different stimuli could be differentiated across different time windows (e.g., A at 5 s versus B at 8 s). The analysis presented in Fig. 2-8 suggests that responses to different stimuli remain distinct across different times (no statistically significant difference between same-time and 47

54 different time pairwise comparisons; P > 0.7; K-S test). The ability of the rodent dentate gyrus to encode four representations following transient molecular layer stimulation is shown graphically in Fig Each possible stimulus location is represented by a character (A-D) whose contrast is a function of the distance to the average response (Fig. 2-9a, see Methods). A trial with a population response very close to the mean response to stimulus B would be represented by a dark B superimposed on dim A, C, and D characters. This method illustrates both the high accuracy of dentate representations of stimulus identity (only 1 of 14 consecutive trials was incorrect) and the persistence of the population coding in the dentate hilus (Fig 2-9b). This visual analysis was cross-validated in Fig. 2-9c, where half the data set was used to compute mean response centroids and the other half used to generate the character display. Dentate gyrus circuits encode temporal sequence order Besides representing individual stimulus positions, hilar networks reliably encoded different temporal sequences. Distinct hilar up-states were evoked by activation of the same four stimuli in different order (ABCD vs. DCBA; Fig. 2-10ab). Hilar synaptic responses to forward- and reverse-order sequences were clustered and reproducible (Fig. 2-10c; 5 s intervals) and could be statistically separated using the same LDA methods used to analyze responses to individual stimuli. Responses to forward- and reverse-order sequences were statistically separable in 10 of 11 hilar triple recording experiments (at least P < 0.05; Fig. 2-48

55 10d). Sequence separability did not reflect simply the differences between the final responses in each sequence. In 8 of the 11 experiments, the sequence response was statistically different from the response to the final stimulus presented in isolation. Sequence identity was accurately predicted on the basis of hilar synaptic responses (classification accuracy = 95.6 ± 1.9 %; significantly greater than expected by chance; P < ; n = 11 experiments). Synaptic barrages associated with sequential stimulation were blocked by the NMDA receptor antagonist MK801 (10 µm; Fig. 2-10e), as described previously for responses to single location stimuli (Larimer and Strowbridge, 2010). Blockade of NMDA receptors did not affect the basal frequency of AMPA receptormediated EPSPs recorded in hilar neurons (P > 0.05; Fig. 2-10e pre-sequence bars), suggesting that this treatment suppressed stimulus-evoked network upstates. Hilar representations of temporal sequences did not arise from simple summation of invariant responses to individual stimuli. Rather, responses to each stimulus varied greatly depending on the temporal order within the sequence (Fig. 2-10f), suggesting that responses to specific stimuli reflected sequence history. While the magnitude of the change in the hilar response tended to decline with each successive stimulus in the train (20.2% mean reduction during each interval), the average magnitude change following the fourth stimulus was still significantly above the basal frequency of spontaneous EPSPs (3.3 ± 0.5 Hz; P < ; n = 11 experiments). The history dependence of responses to stimulus sequences was evident when vector representations of 49

56 responses were added head-to-tail (Fig. 2-10g) and also when responses to the same stimuli were plotted from the same origin (Fig. 2-11). Over 11 triple recording experiments, hilar sequence responses significantly diverged following the second stimuli (Fig. 2-10h) and failed to converge at the end of the sequence. (Convergence would be expected if responses to sequential stimuli were commutative.) Increasing the inter-stimulus interval from 5 to 120 s eliminated most of the history dependence (Fig. 2-10g-h), suggesting a temporal limit to the ability of the dentate gyrus to encode sequential stimuli in vitro. We also found statistically-significant sequence-specific hilar responses when forward and reverse sequential stimuli were repeated at 8 s intervals (3 of out 3 experiments with LDA P values < 0.05; Fig. 2-12) but only infrequently using 2 s intervals (1 of 3 experiments with P < 0.05; Fig. 2-13). Sequence classification accuracy (forward or reverse) also was robust to small perturbations of inter-stimulus interval. Intracellular responses of hilar cells to shortened DCBA stimulus trains (4 s instead of the standard 5 s interval) more closely resembled responses to standard DCBA than to ABCD stimulus trains (Fig. 2-14a). Population responses to shortened (4 s) DCBA trains also overlapped with responses to 5 s DCBA stimuli in EPSP frequency space (Fig. 2-14b) and were closer to the centroid of the 5 s DCBA response than to the 5s ABCD response (Fig. 2-14c). In 5 experiments tested, the plane generated by LDA of sequences with 5 s inter-stimulus intervals accurately classified 4 s reverse sequences (85.4 ± 9.0 % correct; purple bar in Fig. 2-14d; not significantly different than the overall accuracies classifying 5-s interval stimulus 50

57 trains; green and orange bars in Fig. 2-14d; P > 0.05). These results demonstrate that synaptic responses recorded in small groups of hilar neurons in vitro can encode both stimulus and sequence identities and are robust to variations in sequence interval. Population representations of biological information in the dentate gyrus Hilar representations of stimulus information could arise from a simple coding strategy based on the magnitude of the distinct EPSP frequencies across hilar cells. Alternatively, stimulus location and sequence information could be encoded by diverse population activity patterns reflecting differential responses in individual hilar neurons (Fig. 2-15a). We first asked if eliminating the information associated with differences in vector magnitude impaired information encoding. Responses to each of 4 stimulus locations remained well clustered following vector normalization (two examples shown in Fig. 2-15b). The wide distribution of vector directions over 9 triple recording experiments (Fig. 2-15c) suggests there was minimal systematic bias to generate clusters in one particular region. Eliminating response magnitude information only slightly reduced the number of significant LDA separation planes (from 4.8 ± 0.5 to 3.9 ± 0.7; P > 0.05; n = 12; Mann-Whitney; Fig. 2-15d) while eliminating vector direction information (retaining only vector magnitudes) significantly impaired response discrimination (1.4 ± 0.4 planes; significantly less than both actual and normalized; P < 0.01). Normalized population responses also tracked actual discrimination performance well over time (Fig. 2-15e) and remained significantly above chance performance 51

58 over the time period analyzed. Hilar responses to sequential stimuli were similarly robust to normalization (10/11 experiments with separable forward and reverse responses following magnitude normalization; P < 0.05 from LDA), implying that the dentate gyrus uses primarily distributed population codes to represent stimulus information in these experiments. Sequence identity was classified accurately in both control and normalized conditions over 6 s and prediction accuracy remained significantly better than expected by chance over the time window tested (Fig. 2-15f). Discussion We find that dentate hilar neurons reliably encode information as distinct patterns of spontaneous synaptic activity that persist for seconds and that resemble the persistent activity patterns recorded in nonhuman primates (Funahashi et al., 1989; 1993; Colombo and Gross, 1994) and rodents (Hampson et al., 1993) during cross-modality working memory tasks. The same cellular mechanism that encodes isolated stimuli appears also to generate reliable contextual representations of stimuli presented within temporal sequences. Responses of dentate hilar neurons accurately predicted the identity of sequences of stimuli, were distinct from responses to the final stimulus presented in isolation in most experiments, and were robust to perturbation of sequence interval. Our results suggest that both sequential and non-sequential information were represented by population codes in the dentate gyrus. 52

59 Information representation in the dentate gyrus in vitro While hilar synaptic barrages have been reported previously in response to PP stimulation (Larimer and Strowbridge, 2010), this study is the first to demonstrate that changes in the ongoing synaptic input to dentate gyrus neurons can encode both the identity of more than two stimuli presented individually and in temporal sequences. Stimulus-evoked synaptic barrages reliably represented information that could be decoded using a standard classification method (Crowe et al., 2010) assaying the Euclidean distances between a response in a single trial and the four centroids that reflect average responses to the four stimuli. The classification accuracy we found using this unbiased decoding approach (~76% correct trials) was significantly greater than expected by chance; high classification accuracy was maintained when we constructed a naïve decoder based on centroids computed from half of the data set (e.g., Fig. 2-9b). The success of this simple, distance-based decoder was unexpected given the small number of nearby hilar cells sampled and argues that the ability of the dentate gyrus to generate multiple distinct activity patterns is robust and is preserved in acute hippocampal slice preparations. While experimental limitations restricted these experiments to assaying closely-spaced groups of neurons, asking whether widely-separated populations of hilar cells also can represent information should be possible in future experiments using optical activity probes. The success of the decoding method employed in this study suggests that the separation between different population response patterns is typically large enough to overcome noise and any potential synaptic plasticity associated with 53

60 repeated trials in the same experiment. Most decoding errors occurred in experiments where pairs of response centroids occupied similar locations within EPSP frequency space (Fig. 2-6e-f). A simple model that included both the separation between centroids and the dispersion of individual responses around each centroid (Crowe et al., 2010) predicted the minimal inter-centroid distance required for reliable (error-free) decoding (Fig. 2-6h). In addition to assessing accuracy using a Euclidean distance-based decoder, we used two independent methods (overlap coefficient, OVL and linear discriminant analysis, LDA) to test whether hilar population responses evoked by different stimuli were distinct. These methods demonstrated that responses evoked by individual stimuli or temporal sequences of stimuli were statistically separable in most experiments. Our experiments took advantage of intracellular recordings to reveal the underlying synaptic events that excite hilar neurons, rather than inferring synaptic drive from extracellularly recorded spiking activity. One tradeoff with this approach, however, is the limited time over which three simultaneous intracellular recordings could be maintained (typically min). This limitation restricted our ability to gather the hundreds of repetitions typically required for rigorous information-theoretic analysis. Previous work demonstrated that PP-evoked synaptic barrages are unlikely to result from recurrent excitatory connections. The majority of mossy cell axonal projections are to contralateral and distal ipsilateral granule cells (Buckmaster et al., 1996) projections that are largely absent in the transverse hippocampal slices employed in this study. The local projections of 54

61 glutamatergic hilar mossy cells are primarily to GABAergic interneurons and only infrequently (0.5% connection probability) form recurrent monosynaptic connections onto other nearby mossy cells (Larimer and Strowbridge, 2008). Instead, hilar synaptic barrages appear to result from persistent spiking activity in semilunar granule cells (SGCs), a newly-discovered excitatory cell type in dentate molecular layer (Williams et al., 2007; Gupta et al., 2012). SGCs in the molecular layer respond to transient PP stimuli with depolarizing plateau potentials that require L- and T-type voltage-gated calcium channels (Larimer and Strowbridge, 2010). Granule cells, by contrast, are typically inhibited by PP stimuli that trigger hilar synaptic barrages. Both granule cells and SGCs are polarized neurons whose axonal arbors are excluded from regions containing their own somatodendritic compartments (Williams et al., 2007; Gupta et al., 2012), indicating that neither cell type forms recurrent excitatory connections. Antagonists of calcium channels, as well as NMDAR blockers, abolished both plateau states in SGCs and synaptic barrages in downstream hilar neurons, without abolishing fast AMPAR-mediated excitatory synaptic transmission (Larimer and Strowbridge, 2010). Our ability to trigger statistically separable hilar population responses may result from recruitment of persistent spiking activity in different subsets of SGCs that project diffusely to hilar neurons. Alternatively, the distinct patterns we observe may result from stimulus-specific graded plateau depolarizations (Larimer and Strowbridge, 2010) within a uniform subset of SGCs or from differential activation of inhibitory local circuits. Further experiments monitoring larger subsets of dentate gyrus neurons will be required to determine 55

62 the extent of the SGC network excited by different stimuli in order to distinguish among these hypotheses. Previous in vitro work in neocortical slices (Shu et al., 2003) demonstrated that proportionally balanced excitation and inhibition enables self-sustaining activity mediated by recurrent excitatory connections, termed up-states. Cortical up-states occur spontaneously in acute brain slices bathed in solutions that enhance excitability and also can be triggered by synaptic stimulation (Shu et al., 2003; MacLean et al., 2005). While these up-states can form representations of a single stimulus (whether a stimulus was presented or not), these network states are not synapse specific (Shu et al., 2003). By contrast, PP-evoked hilar synaptic barrages are highly synapse specific and are capable of representing at least four distinct patterns. Cortical networks also exhibit spontaneous neuronal avalanches that persist for 10s of ms (Beggs and Plenz, 2003), a much faster time scale than that of the persistent activity reported here. To our knowledge, no group has reported stimulus-specific avalanches and most studies have focused on spontaneous avalanches that follow highly variable trajectories. In addition to representing specific stimulus locations, we find that hilar synaptic barrages accurately represented temporal sequences of stimuli. Rather than simply trigger stereotyped, stimulus-specific activity patterns, stimulus sequences typically triggered new activity patterns that were distinct from those evoked by the final stimulus in the train (ABCD was different from D presented alone; Fig. 2-13). This context dependence of stimulus representation, where the 56

63 response to a specific stimulus was dependent on prior stimuli, also was apparent when synaptic inputs during sequence trains were analyzed (Fig. 2-10e). Surprisingly, classification accuracy for decoding temporal sequences was robust to a perturbation in sequence interval (Fig. 2-14d), suggesting that context dependence arises primarily from the relative order of the stimuli. Semilunar granule cells often receive inhibitory postsynaptic responses following hilar and molecular layer stimulation (Williams et al., 2007) that could provide a cellular substrate for context dependence. While sequence memory often involves the hippocampal formation (Fortin et al., 2002) and has been demonstrated in computational models of cortical networks (Abbott and Blum, 1996; Diesmann et al., 1999; Mongillo et al., 2008; Goldman, 2009), reliable short-term storage of temporal information has only been infrequently demonstrated at the cellular level. While Branco et al. (Branco et al., 2010) recently used focal glutamate uncaging at multiple sites along one dendrite to generate sequence-dependent intracellular responses, this mechanism presumably operates primarily through rapid cell-autonomous mechanisms. By contrast, the dentate gyrus circuits activated in the present study reliably encoded temporal sequences with substantially longer time intervals (seconds) and could be decoded accurately only by comparing activity across multiple simultaneously-recorded hilar cells. Other investigators have demonstrated distinct network states reflecting contextual temporal information that evolve following repeated training. While the state-dependent networks described in these studies can reliably encode temporal interval, it is not clear if 57

64 they can also represent sequence order (e.g., ABCD vs. DCBA). Multiple groups (Nikolic et al., 2009; Johnson et al., 2010) have identified state-dependent network responses in vitro (Johnson et al., 2010) and in vivo (Nikolic et al., 2009) that reflect the history of previously presented stimuli. While the contextual responses we report in the dentate gyrus in vitro likely reflect state-dependent processes, our results differ from previous reports in several respects. We find that hilar networks represent information over longer time periods (many seconds) compared with previous state-dependent studies (Buonomano and Maass, 2009; Nikolic et al., 2009; Johnson et al., 2010). Hilar networks can encode the order of temporal sequences, not only immediately following the sequence, but also for several seconds afterward. Finally, our results demonstrate rapid encoding of information while previous in vitro work required many repeated presentations of stimuli (Johnson et al., 2010), implying a gradual entrainment or learning process. Relationship to delay-period activity recorded during working memory tasks Much of our understanding of the cellular basis of short-term information storage is based on classic extracellular recording studies that demonstrated persistent spiking of neocortical (Funahashi et al., 1989; Batuev et al., 1990; Funahashi et al., 1993; Baeg et al., 2003) and hippocampal (Hampson et al., 1993; Colombo and Gross, 1994) units during the delay period of short-term memory tasks. These studies showed that the particular units recruited during the delay period depended on the stimulus type (e.g., spatial location) and that 58

65 delay-period activity was typically absent or reduced in trials when the preferred stimulus was presented and the subject made an incorrect behavioral response (Funahashi et al., 1989; 1993). Extracellular recording studies have demonstrated elevated firing frequencies for as long as 20 s during working memory tasks (Fuster and Jervey, 1981; Miyashita and Chang, 1988), arguing that cortical networks must contain mechanisms that support persistent activity over this time scale. However, neurons with a diversity of persistent firing durations typically contribute to delay-period activity during working memory tasks and only a fraction are typically active during the entire delay phase. The subset of delay-period neurons with prolonged persistent activity may play a critical role as summators that integrate activity from neurons that are active over different time scales (Batuev et al., 1990; Baeg et al., 2003). While it is difficult to determine the kinetics of the mechanisms underlying persistent activity during working memory tasks since the learned motor output typically extinguishes persistent firing (Funahashi et al., 1989), our experiments demonstrate that the circuitry contained in transverse hippocampal slices, including SGCs with intrinsic persistence, are sufficient to represent information over these long behavioral time scales. Multiple studies (Baeg et al., 2003; Barak et al., 2010) have demonstrated that compact representations based on population decoding approaches can accurately predict behavioral responses in working memory tasks. Similarly, preserving the heterogeneity in hilar cell responses while ignoring differences in the overall amplitude of responses (i.e., normalization) maintained accurate decoding in our experiments. By contrast, 59

66 average rate-based coding schemes that ignored the differential contribution of each neuron to the population response failed to classify accurately either location or sequence stimuli. One likely function of dentate gyrus neurons in vivo hypothesized by Marr (Marr, 1971) and subsequent researchers (Rolls and Treves, 1994) is to transform highly overlapping entorhinal input patterns into more distinct output patterns. While this potential functional role has only recently begun to be tested experimentally (Leutgeb et al., 2007; McHugh et al., 2007), our demonstration that multiple, reproducible population patterns can be generated in dentate gyrus neurons is consistent with the hypothesis that the neural networks in this brain region can stably represent multiple population patterns. Semilunar granule cells may play an important role during pattern decorrelation by generating a latch function that enables hilar neurons to maintain their activity following the transient presentation of a particular entorhinal input pattern. Future in vivo studies will be required to follow the downstream consequences of hilar persistent activity since most of the axonal projections of mossy cells extend to distant ipsilateral and contralateral granule cells (Deadwyler et al., 1975; Buckmaster et al., 1996), which are not present in transverse hippocampal slices. The largely feedforward local circuits that appear to support short-term storage of information in the dentate gyrus suggest a fundamentally different mechanism than the recurrent excitatory networks commonly employed in artificial neural network models of working memory (Seung, 1996; Camperi and Wang, 1998; Compte et al., 2000; Miller et al., 2003). Despite extensive efforts 60

67 to identify biological persistent activity modes mediated by recurrent excitatory synaptic connections (Aksay et al., 2007), very little experimental evidence supports the classic hypothesis that reverberant network activity functions to encode and maintain information proposed by Hebb (Hebb, 1949). Our finding that sequential stimulation did not reset hilar persistent activity to a new state corresponding to the most recent stimulus is also inconsistent with the predictions from theoretical studies employing attractor-based networks with recurrent connectivity (Brunel and Wang, 2001). At the other extreme, multiple groups have demonstrated neuron cell types that can fire persistently following transient stimuli through cell-autonomous intrinsic mechanisms (Egorov et al., 2002; Pressler and Strowbridge, 2006) and several theoretical studies have demonstrated that combining intrinsic persistence with recurrent excitatory connections relaxes the requirements for precise tuning of synaptic weights (Camperi and Wang, 1998; Seung et al., 2000; Koulakov et al., 2002; Goldman et al., 2003). The alternative intrinsic/feedforward mechanism we hypothesize enables hilar neurons to represent stimulus and sequence identity is related to a recent theoretical study (Goldman, 2009) arguing that purely feedforward circuits can generate prolonged network activity when the effective network time constant is extended by creating multiple serial processing stages. Presumably, the number of stages required to generate delay-period activity would be reduced if some stages of the feedforward network employed neurons with intrinsic persistence, such as SGCs. The dentate gyrus is unusual in the high frequency of spontaneous 61

68 synaptic activity recorded in hilar neurons and because of semilunar granule cells, a subtype of excitatory projection with multistability under standard physiological conditions. Populations of hilar neurons presumably can encode stimulus and sequence identity in our experiments because at least some divergent SGC-to-hilar neuron synaptic connections are preserved in conventional horizontal brain slices. Revealing persistent activity modes in the dentate gyrus was facilitated by the anatomical separation of multistable excitatory neurons (SGCs) and downstream hilar neurons. However, other examples of bistability and multistability of excitatory projection neurons in more heterogeneous brain regions have been reported following bath application of modulatory neurotransmitters, raising the possibility that the hybrid intrinsic/feedforward mechanism we propose in the dentate gyrus may be generalizable. For example, a subclass of neurons in entorhinal cortex are capable of persistent firing following transient stimulation in the presence of muscarinic receptor agonists (Egorov et al., 2002). To date, no group has reported whether activation of different synaptic inputs to entorhinal cortex leads to mnemonic encoding. Transient inputs also can trigger cell-autonomous persistent firing in neocortical neurons when intrinsic excitability is enhanced following cholinergic receptor activation (Krnjević and Phillis, 1963), providing a potential connection between intrinsic/feedforward persistent mechanisms we propose in the dentate gyrus and classic delay-period activity recorded in prefrontal cortical neurons in vivo. 62

69 Methods Animals. Horizontal slices (300 µm thick) of the ventral hippocampus were prepared from P14-25 Sprague-Dawley rats anesthetized with ketamine, as described previously (Larimer and Strowbridge, 2008; 2010). Slices were incubated at 30 C for 30 min and then maintained at room temperature until needed. All experiments were carried out under guidelines approved by the Case Western Reserve University Animal Care and Use Committee. Electrophysiology. All intracellular recordings were performed in a submerged recording chamber maintained at 30 C and perfused with an extracellular solution containing (in mm): 124 NaCl, 3 KCl, 1.23 NaH2PO4, 1.2 MgSO4, 26 NaHCO3, 10 dextrose, 2.5 CaCl 2, equilibrated with 95% O 2 /5% CO 2 (ph 7.3). Whole-cell patch clamp recordings were made using AxoPatch 1D amplifiers (Molecular Devices) and borosilicate glass pipettes (3-10 MΩ). Recording electrodes contained (in mm): 140 K methylsulfate, 4 NaCl, 10 HEPES, 0.2 EGTA, 4 MgATP, 0.3 Na3GTP, 10 phosphocreatine, adjusted to ph 7.3 and ~ 290 mosm. Individual neurons were visualized under IR-DIC video microscopy (Zeiss Axioskop FS1) prior to patch-clamp recording. We typically restricted our experiments to a set of two or three nearby hilar cells with inter-soma distances less than 100 µm. Intracellular recordings were low-pass filtered at 2 khz (FLA- 01, Cygnus Technology) and acquired at 5 khz using a simultaneously-sampling 16-bit data acquisition system (ITC-18, Instrutech) operated by custom software written in VB.NET (Microsoft) and Matlab (Mathworks). Intracellular voltages 63

70 were not corrected for the liquid junction potential. A matrix microelectrode consisting of 4 sharpened tungsten monopolar electrodes (115 µm spacing; FHC) was used for extracellular stimulation. Stimulus intensity was controlled by a custom-built constant-voltage stimulus isolation unit. Hilar barrages were evoked most reliably by paired stimuli (500 ms inter-stimulus intervals), as described previously (Larimer and Strowbridge, 2010). All population responses reported in this study reflect paired stimuli at this interval; responses to sequential stimuli (e.g., ABCD) reflect paired shocks at each matrix electrode contact at the indicated interval (paired shocks every s). Hilar recordings and population responses were typically stable over paired stimuli or temporal sequences of different stimuli (3.5 min delay between different stimuli in both experiment types, except for 120 s sequential stimuli experiments). We collected the data set of experiments using different stimulus locations in one continuous series after an initial set of pilot studies to optimize stimulus and recording conditions. Only experiments in which hilar cell recordings degraded or in which we obtained fewer than 3 responses from each stimulus type were excluded from the analysis. Data analysis and statistics. Hilar cells were identified by their intrinsic properties (mean spike time, spike clustering, and spike AHP) following 4 s-duration depolarizing and hyperpolarizing current steps, as described previously (Larimer and Strowbridge, 2008). Most recordings in this study were from presumptive hilar mossy cells. Spontaneous EPSPs were identified automatically using a 64

71 custom detection algorithm (Larimer and Strowbridge, 2008). Except where noted, all data are presented as mean ± SEM. Statistical significance was determined using Student's t-test, unless otherwise specified. Analysis of hilar population responses. EPSPs were typically detected in a 4-s window that began 0.5 s after the last shock of each paired stimulus. The mean EPSP frequency within this window in each recorded hilar cell was used to analyze population responses. We approached the question of whether population responses evoked by different stimulus electrodes were distinct using three complementary methods. First, we assessed the separability of population responses by calculating the overlap coefficient (OVL (Inman and Bradley, 1989; Clemons and Bradley, 2000); Fig. 2-5). In this method, we computed the distribution of Euclidean distances from individual population responses (the mean EPSP frequencies in the same 4 s analysis window applied to each simultaneously-recorded hilar cell) with the centroid calculated from all responses evoked by the stimulus electrode in the same experiment ( Cis distances). The distribution of these Cis distances was then compared with the distribution of Trans distances calculated from dissimilar data point/centroid pairs (e.g., A points to the centroid of all B points). The overlap coefficient OVL was computed from Gaussian fits of Cis and Trans distributions. The statistical significance of the OVL coefficient was determined by bootstrap methods (20,000 randomizations of Cis/Trans labels on each actual distance). All location experiments analyzed in this study yielded statistically distinct Cis/Trans 65

72 distributions (all OVL calculated reflected P < 0.05; Fig. 2-2e). Second, we used linear discriminant analysis (LDA (Fisher, 1936; Klecka, 1980; MacLeod et al., 1998)) to determine the number of statistically separable population responses in each experiment. Complete classification of population responses to all 4 stimulus locations (A-D) required 6 statistically significant pairwise LDA tests (A/B, B/C, etc), each with P < reflecting a Bonferroni correction for multiple comparisons. Plots of the number of discrimination planes in different conditions (Fig. 2-2d, Fig. 2-4, Fig. 2-15d-e) reflect the number of pairwise LDA tests with P < We estimated the number of planes expected by chance by randomly re-assigning EPSP frequencies in each hilar cell to different stimulus identities. We recalculated the number of statistically significant LDA planes following 30 iterations to generate an estimate of the number of planes expected by chance. A statistically significant OVL coefficient combined with 6/6 possible LDA tests (with P < ) represents a conservative test for response separability, a criterion that was met in 50% of the experiments we analyzed. If one experiment failed to generate 4 separable responses, it was re-analyzed to determine if three (3 appropriate planes with P < 0.016) or two (one plane with P < 0.05) responses could be isolated. We calculated the time course of response separability using a sliding 4-s window and repeated the LDA at 1 s intervals. Times indicated reflect the beginning of each 4-s time window. We also used LDA based on the mean EPSP frequency in 4-s windows that began 0.5 s after the last stimulus in each sequence (e.g., 0.5 s after the D in ABCD and after the A in DCBA) to determine if population 66

73 responses to different temporal sequences were significantly different. We employed the same time windows when analyzing sequences with both 4- and 5- s intervals. Finally, we assessed the accuracy of hilar population responses in predicting stimulus identity, assayed by nearest stimulus centroid. Classification accuracy was calculated for the entire data set from the ratio of trials classified correctly to the total number of trials tested. Sequence identity was predicted from the sign of the response distance to the LDA plane. Character intensity plots (Fig. 2-9) reflect the relative Euclidean distances between each population response and the four stimulus centroids, recomputed at 1 s intervals. Each location reflects the superposition of four characters (A,B,C,D) whose intensities are scaled to reflect their relative distance to stimulus centroids. The scaling function was normalized by the sum of the scaled distances to all four centroids: where d represents the Euclidean distance between a single response and a response centroid and x is the stimulus identity being tested. This transformation resulted in one very dark (high contrast) character superimposed on three light (low contrast) characters if a response was very close to a stimulus centroid. We then iterated through all four possible centroids at each time step: 67

74 , where p is character intensity, i represents stimulus identity, j represents the time bin, k represents the trial, and F C represents the centroid vector for one stimulus identity at one time bin. The relative distance between forward and reverse responses in Fig. 2-14c was calculated as the signed length of the projection of the population responses onto the difference vector: where δ represents the difference vector of the mean population responses for forward and reverse sequences and F is the population response vector. Data analysis was performed in Matlab (Mathworks) and Origin (OriginLab). 68

75 Figure 2-1 Persistent synaptic barrages evoked by multiple PP stimuli a, Diagram of experimental configuration. A four-position stimulating electrode was used to activate different perforant pathway (PP) segments while recording synaptic input in hilar neurons. b, Example of a hilar mossy cell synaptic barrage evoked by PP stimulation (arrow heads). The timing of the onsets of individual EPSPs is indicated by vertical lines. Enlargement of region indicated with horizontal bar shown in inset. c, Diagram of experimental protocol and EPSP analysis windows. d, Responses to three repetitions of each of the four stimulus positions (A-D) in one hilar mossy cell. Vertical lines indicate EPSP onset times. Mean EPSP frequency and range for each stimulus type indicated on right. Diagram of experiment protocol shown at top. 69

76 70

77 Figure 2-2 Short-term hilar representations of multiple PP stimuli a, Diagram of triple hilar cell recording configuration. b, Stimulation at different sites evokes distinct synaptic barrages in dentate hilar networks. Summary plot of mean EPSP frequency during initial 4 s of barrages following PP stimulation in one experiment. Black dots represent 14 consecutive single trials, crosses represent centroids of each trial type (stimulus A, B, etc). Shaded zones represent bounding ellipsoids used to illustrate variance associated with repeated responses to the same stimulus (60% confidence intervals). c, Histogram of experiments with 2, 3 and 4 statistically separable responses (out of 4 possible) over 12 experiments. No experiments had fewer than 2 separable responses. d, Plot of the number of statistically significant (P < 0.05) planes obtained by LDA in the 9 triple recording experiments (black bars) and following shuffling stimulus identity (grey bars). Results are presented from LDA computed on data from all three simultaneous recordings (right set of bars; n = 9) and when only one (n = 27) and two neurons (n=27) are considered. ** P < e, Plot of probabilities associated with the overlap coefficient (OVL) for the same 12 experiments. Dashed line indicates statistically significant OVL range (P < 0.05). See text and Fig. 2-5 for details. 71

78 72

79 Figure 2-3 Example of synaptic barrages evoked by different stimulus locations Raster plots of EPSP onset times recorded in three hilar cells simultaneously (2 s window beginning 0.5 s after the last stimulus). Responses from different cells identified by color. Three consecutive trials from each stimulus location presented in each box. Each row represents responses to a different stimulus location. Stimulus trials were evoked in pseudorandom order but displayed together for clarity. 73

80 74

81 Figure 2-4 Response classification as a function of analysis window duration Plot of the number of statistically significant LDA planes (P < 0.05) versus window duration for 4-site location stimulation experiments (n = 12; 6 possible planes). All EPSP frequency analysis windows began at the same time point (0.5 s following the last PP stimulation). * P < 0.05 (t-test). (No other comparisons with the number of planes generated using 4 s duration windows were statistically significant.) 75

82 76

83 Figure 2-5 Computing the OVL statistic a, Schematic of overlap method. Red indicates cis distances computed from Euclidean distances between individual responses and the mean response of that stimulus type. Blue indicates trans distances from individual responses to each dissimilar mean response. b, Example of overlap analysis in one experiment showing the distribution of cis (n = 19) and trans (n = 57) distances. Solid lines indicate Gaussian fits to each distribution (cis mean = 6.6 Hz; trans mean = 27.1 Hz). Overlap coefficient calculated from probability density distribution overlap (OVL = in this experiment). All distances computed from mean EPSP frequencies in 4 s window beginning 0.5 s after the last stimuli. c, Probability distribution obtained by shuffling distance labels (cis or trans) 20,000 times. In this experiment, the probability of obtaining the observed overlap coefficient by chance was < Statistical significance was considered at the 95% confidence interval. 77

84 78

85 Figure 2-6 Prediction of stimulus identity from hilar population responses a, Diagram of analysis method based on computing Euclidean distance from each response (black diamond) to the four centroids (filled circles) that represent average responses to each type of stimuli in that experiment. b, Plot of mean accuracy in predicting stimulus identity (A-D) based on the nearest average response centroid, compared with accuracies expected by chance. ** P < c, Plot of the average Euclidean distance between pairs of response centroids evoked by adjacent (115 µm separation) and non-adjacent (230 and 345 µm separation) stimulating electrode contacts. d, Plot of accuracy of predicting stimulus identity based on nearest centroid versus stimulus electrode separation. Pairs of stimuli evoked on adjacent electrode contacts were classified as accurately as stimuli evoked by non-adjacent pairs. Accuracy expected by chance was 50% in this pairwise analysis. e, Distribution of inter-centroid distances between pairs of responses that were correctly (black symbols) and incorrectly (red symbols) classified. Number of errors associated with each intercentroid distance plotted on X axis. Mean ± SEM of distance for correctly (black circle) and incorrectly (red circle) classified response pairs plotted next to each distribution. f, Cumulative distribution of Euclidean distances associated with classification errors. Half of the classification errors results from comparisons between response centroids separated by less than 7 Hz (arrow on X axis). g, Plots of accuracy classifying response pairs with occasional errors versus intercentroid distance. h, Plot of accuracy classifying response pairs with occasional errors versus distance/s.d. ratio. Solid lines represent linear regression fits in 79

86 both g and h. Black symbols represent the mean Euclidean distance (in g) and distance/s.d. ratio (in h) associated with response pairs that were always classified correctly. 80

87 81

88 Figure 2-7 Time course of hilar population responses a, Plot of decay of average EPSP frequency during synaptic barrages in 12 experiments (thin lines represent each experiment; symbols represent overall mean ± SEM). Arrow on Y axis indicates average baseline EPSP frequency. b, Plot of centroid position as response evolves over time in one experiment (18 sliding 4-s windows). Each point reflects mean EPSP frequency in three hilar cells over a 4-s window. c, Plot of the number of statistically separable responses (out of 4 possible) versus time. Symbols represent mean ± SEM (n=12 experiments). Red and blue lines represent best and worst three experiments, respectively (see text). d, Plot of classification accuracy over time (black symbols). Arrow along Y axis indicates accuracy expected by chance. e, The initial response overlap coefficient (OVL) is negatively correlated (R 2 = 0.62; P < 0.005; F = 15.4; n = 12 experiments) to the duration that responses evoked by different stimuli remained significantly distinct. 82

89 83

90 Figure 2-8 Classification accuracy across different times a, Example displays of pairwise classification accuracies in two experiments (both stim A vs. B comparisons). All accuracies were computed pairwise using linear discriminant analysis. b, Summary display for all pairwise comparisons (A vs. B, B vs. C, etc) from 9 triple recording experiments. The same color map for pairwise accuracy, shown on right, was used for displays in panels A and B. Black rectangles represent same-time comparisons; off-diagonal values represent comparisons across different times (e.g., 5 sec in Stim A vs. 10 sec in Stim B). All accuracy matrices were computed using 4 s analysis windows starting at the times indicated on the X and Y axes. c, Histogram of all same-time pairwise discrimination accuracies in 9 triple recording experiments. d, Histogram of across-time discrimination accuracies in the same 9 experiments plotted in C. e, Plots of cumulative distributions of the same-time and across-time discrimination accuracies. The two cumulative distributions are not statistically different (two-sample Kolmogorov-Smirnov test statistic = 0.18; P > 0.7). 84

91 85

92 Figure 2-9 Visual display of short-term information storage in the dentate gyrus a, Analysis method used to generate visual display. Character intensity was inversely related to distance to centroids for each stimulus. See Methods for details. b, Graphical representation of hilar population response to repeated trials of four different stimulus sites (indicated on left; data acquired in pseudorandom order and then re-arranged to cluster each stimulus type). c, Graphical representation of response classification using only untrained data. Half of the episodes were used to compute response centroids; character plots were generated from the remaining 50% of episodes. 86

93 87

94 Figure 2-10 Short-term representations of temporal sequences in hilar neurons a, Diagram of experimental configuration. b, Responses to three trials of forward (ABCD) and reverse (DCBA) temporal stimuli sequences recorded intracellularly from three hilar cells. Vertical lines represent EPSP onset times. c, Plot of individual responses to forward and reverse sequential stimulation (black circles) in EPSP frequency space (5 forward and 6 reverse, acquired in pseudo-random order; 4 s analysis window after final stimulus in each sequence). All points contained within bounding ellipsoids (60% confidence interval, as in Fig. 2-2b) centered on response centroids (black asterisks). d, Plot of probabilities that forward and reverse points are significantly different by LDA. Forward and reverse sequences were significantly different (P <0.05) in 10/11 experiments (filled symbols) and not different in one experiment (open symbol). e, Plot of the effect of MK801 (10 µm) on the mean EPSP frequencies in three sequential stimulation experiments. Population responses to forward and reverse sequential stimulation were separable using LDA in control conditions in each experiment (before MK801; all P < 0.05). Sequential stimulation triggered a mean increase of 17.8 Hz in control conditions in these three experiments, not statistically different than the 16.1 Hz increase observed in the larger set of 12 sequential stimulation experiments. Sequential stimulation failed to trigger an increase in EPSP frequency in MK801 (P > 0.05). * P < f, Plot of mean EPSP frequency across a triple recording for all forward and reverse sequences tested in one experiment. Horizontal bars indicate mean EPSP frequency for 88

95 baseline period and for 4-s windows following each stimuli within the sequence. g, Vector representation of average population response to forward and reverse sequences. Responses to each stimuli were combined head-to-tail and failed to converge for fast (5 s intervals, left) sequences but converged for slow (120 s intervals, right) sequences in a different experiment. h, Plot of average separation in head-to-tail vector representations of forward and reverse sequences (similar to g) over 11 experiments. Black bars represent results from 5 s sequences; purple bar represents forward/reverse separation at the end of 120 s sequences. * Significantly different; * P < 0.05; ** P < Mean ± SEM. 89

96 90

97 Figure 2-11 Comparison of responses to the same stimulus presented in different sequences a, Examples of vector responses to stimulus B (left plots) and C (right plots) presented within forward and reverse 5 s sequences. b, Examples of vector responses to B and C stimuli presented within 120 s forward and reverse sequences. All vectors plotted in panels A and B were computed from average responses to 5 forward and 5 reverse sequence trials and displayed from the same origin. 91

98 92

99 Figure 2-12 Distinct representations of forward and reverse sequences with 8 s intervals Plot of individual responses to forward (n = 6 trials) and reverse (n = 7 trials) sequential stimuli (8 s intervals) plotted in 3-dimensional EPSP frequency space. All points contained within bounding ellipsoids (80% confidence intervals) centered on response centroids. Forward and reverse responses were statistically separable in this experiment (P < 5 x 10-6 ; t-test). 93

100 94

101 Figure 2-13 Plot of context dependence of responses to sequential stimuli Summary plot of results from sequential stimulation experiments grouped by sequence interval. Average separation between forward and reverse sequences plotted along X axis (ABCD vs DCBA); average distance between sequence response and response to the last stimulus within the sequence when presented in isolation plotted along Y axis (D vs ABCD; A vs DCBA). Data plotted as mean ± SEM. Red symbol represents 5 s interval sequence experiments in which NMDA receptors were blocked using 10 µm MK

102 96

103 Figure 2-14 Sequence representations are robust to perturbation of stimulus interval a, Example of EPSP responses to forward and reverse sequential stimuli at 5 s intervals (ABCD, DCBA) and reverse stimuli at 4 s intervals (DCBA short) recorded in three hilar mossy cells. b, Plot of forward (black symbols within green ellipsoid), reverse (black symbols within orange ellipsoid) and reverse short (white points within purple ellipsoid) responses in EPSP frequency space in one experiment. Response centroids indicated by asterisks (reverse and reverse short centroid symbols overlap); colored ellipsoids represent 60% confidence intervals. c, Plot of scaled distance between forward (black symbols) and reverse (orange) 5 s interval sequence responses in one experiment. Responses to 4-s reverse responses (purple symbols) overlapped the 5 s reverse responses. Over 5 experiments, the mean scaled distance of the 4 s reverse points was 0.72; significantly different from 5 s forward responses (P < 0.005) and not significantly different from 5 s reverse responses (P > 0.05). d, Plot of classification accuracy for experiments with forward (green bar; n = 11), reverse (orange bar; n =11) 5 s interval sequences, and reverse sequences with 4 s intervals (purple bar; n = 5). Mean accuracies not statistically different from each other, P > Mean ± SEM. 97

104 98

105 Figure 2-15 Population representations of stimulus and sequence identity in the dentate gyrus a, Diagram illustrating rate and population coding strategies. b, Plots of responses to 4 different stimulus positions in normalized EPSP frequency space from two experiments. c, Plot of all responses from 11 experiments in normalized EPSP space. Gray regions represent surface of sphere with unit radius. d, Plot of number of statistically significant LDA separation planes (P < 0.05) from all experiments with 4 different stimulus locations ( Actual, n = 12), the number of planes in the same data set after normalizing vector magnitudes ( Normalized ), and after eliminating vector direction information ( Magnitude only ). Dashed line indicates mean number of significant LDA planes after shuffling stimulus identities in all 12 experiments. Mean ± SEM. ** Significantly different from shuffled stimulus identities; P < Both Actual and Normalized are significantly different from Magnitude only (P < 0.01); Actual and Normalized are not significantly different (P > 0.05). e, Plot of number of statistically significant LDA separation planes (P < 0.05; out of 6 possible) over time. Purple symbols indicate mean ± SD number of significant planes after shuffling the stimulus identities. Both actual (black symbols) and normalized responses (red symbols) remained significantly greater than shuffled throughout the time period examined (P < 0.05). f, Plot of accuracy of predicting sequence identity over time based on moving 4-s analysis windows in control conditions (Actual, black symbols), following response normalization (red symbols) and using only vector magnitude information (green symbols). Both control (Actual) 99

106 and Normalized conditions are significantly different from the Magnitude only condition at all time points. ** P <

107 101

108 Chapter 3: Discussion The experimental results presented in this thesis indicate that the dentate gyrus network is able to maintain a pattern of synaptic activity across three hilar cells for more than 10 seconds and that sequences of stimuli can also be represented for a similar time period. The ability of the hilar network to form these representations depends on persistent excitation of stimulus-specific mossy cell assemblies. Hilar representations of temporal sequences can persist for seconds, are robust to small perturbations in the interstimulus interval, and can faithfully represent the order of the sequence (Hyde and Strowbridge, 2012). While the specific cellular and network mechanisms that may account for the contextual representations observed in this dissertation have not yet been fully explored, a number of previous studies highlight the role of both individual cellular properties and network properties in maintaining representations of both individual stimuli and temporal sequences. In the following, I will discuss how previous computational models of information storage as well previous experimental studies have elucidated mechanisms of maintaining spatiotemporal patterns of information. Finally, I will propose a hypothetical mechanism, based on known properties of the hilar cell network, that may account for the findings in this dissertation. 102

109 Mechanisms of transient information storage: cellular processes Neurons of the CNS can exhibit a wide variety of dendritic morphologies, from compact arborizations to elaborate branching patterns (Spruston, 2008). The length, morphology, and constellation of conductances and receptors found in dendrites can profoundly affect the characteristics of propagating potentials generated by synaptic inputs (Branco and Häusser, 2010). In the 1960 s, Wilfrid Rall proposed that the output of a neuron in response to synaptic stimulation along its somato-dendritic compartment can depend greatly on the location and time course of those synaptic inputs (Rall, 1967). Because the electronic length of the dendritic arbor of a neuron passively shapes the waverform of propagating synaptic inputs in a manner determined by the length, diameter, and branching structure of the dendritic compartment (Rall, 1962), an individual neuron can produce unique outputs to a variety of individual stimuli. In addition, Rall showed that unique outputs were not limited to individual stimuli, but that neurons can transform combinations of distinct excitatory and inhibitory inputs into specific somatic potentials, elucidating the influential role of the dendritic compartments in the transfer function of the neuron (Rall, 1967). In this way neurons can act as nonlinear integrators of synaptic inputs and produce specific outputs in response to individual and combinations of stimuli. A number of experimental and theoretical studies have supported and extended Rall s theoretical predictions. In an in vitro slice preparation of the turtle spinal cord, it has been shown that local glutamate release onto distinct dendritic compartments of the same motor neuron can result in specific somatic 103

110 potentials (Skydsgaard and Hounsgaard, 1994). In these intracellular recordings, somatic potentials generated by glutamate application along one dendrite could be modulated by local application of GABA, TEA, and NMDA, but such pharmacological interventions did not affect the response to activating a different dendrite. These results suggest that although neurons often act as nonlinear summators of their inputs (Burke, 1967; Rall, 1967), dendrites can also exhibit non-trivial autonomous behavior. In a more recent study using two-photon Ca 2+ imaging and optical stimulation in the isolated intact rabbit retina, it was demonstrated that starburst amacrine cells can exhibit location-specific dendritic Ca 2+ transients (Euler et al., 2002). The authors of this study went on to show that, not only can different dendritic branches act locally and independently of one another, but that individual dendrites can exhibit selectivity to the direction of motion of a optical stimulus. Optical stimulation that moved centrifugally from the soma to the distal dendrite generated large Ca 2+ transients, whereas motion in the opposite direction failed to generate a response. It was concluded that, because starburst amacrine cells form inhibitory synapses onto retinal ganglion cells, this asymmetric pattern of dendritic activation could account for the longobserved direction-selectivity of the principal output cells of the retina (Barlow et al., 1964). The asymmetric activation of dendritic conductances in the retina indicates that dendrites alone may possess intrinsic mechanisms for generating specific responses to complex stimuli. Indeed, over the last few decades, it has become apparent that dendrites do not necessarily act as passive filters of synaptic inputs 104

111 but that they often can transform inputs due to the presence of a number of active conductances (Silver, 2010). For example, Na + conductances (Golding and Spruston, 1998; Oviedo and Reyes, 2002) and NMDA receptor conductances (Larkum et al., 2009) in the dendrites of pyramidal cells can amplify synaptic inputs. The presence of these active dendritic conductances can potentially alter the impact of synaptic inputs at different locations on the dendritic arbor (Cook and Johnston, 1999; Oviedo, 2005). In addition to Na + and NMDA receptor conductances, dendritic Ca 2+ conductances can amplify coincident synaptic inputs to generate dendritic spikes (Larkum et al., 2009). In the presence of back-propagating action potentials, there would be a lower threshold for these Ca 2+ dendritic spikes, which could enable the cell to associate spatially separated signals (Larkum et al., 1999). However, it has also been shown that the presence of active conductances does not necessarily result in complex transformations of synaptic inputs as such conductances can cancel each other out (Cash and Yuste, 1998), resulting in a dendrite that effectively acts as a passive filter. Finally, computational models of neurons containing similar distributions of active dendritic conductances indicated that even when different dendrites exhibit the same constellation of conductances, the geometric configuration of dendritic compartment can greatly affect the output properties of the neuron (Mainen and Sejnowski, 1996). Thus, the impact of spatially segregated synaptic inputs on a given neuron depends on a complex interplay of dendritic conductances, morphological features, and biophysical properties. 105

112 Mechanisms of transient information storage: network processes Although it is clear that neurons can vary widely with respect to their intrinsic electrical and morphological properties, neurons are organized into networks that themselves exhibit a rich variety of structure (Bullmore and Sporns, 2009; Reid, 2012). How synaptic activity between individual neurons forms the foundation for the extended computational hierarchy of the whole brain has been a guiding question for brain theorists (Marr, 1970; 1971; McCulloch and Pitts, 1990). McCulloch and Pitts were the first to propose that the all-or-none character of neuronal activity endowed networks of neurons with a vast computational ability (McCulloch and Pitts, 1947; 1990). While only considering the neurons themselves as simple integrative devices, McCulloch and Pitts showed that networks of excitatory and inhibitory elements could be combined in ways that could compute any logical function (e.g. XOR, NAND, etc.) As small networks of neurons performing a given logical operation could be connected with networks performing other logical operations, it was suggested that increasingly larger and more complex networks of neurons could compute increasingly more difficult computational problems. This emphasis on the material basis underlying computational properties of the brain has led to a number of connectionist (Rosenblatt, 1958), or artificial neural network, architectures (Rumelhart et al., 1986b; Arbib, 2003; Thomas and McLelland, 2008). All neural networks consist of input and output layers of processing units. The units in these layers may be connected directly to each other or indirectly through one or more hidden layers (Fig. 3-1) of processing 106

113 units (Thomas and McLelland, 2008). When input signals are delivered to the network at the input layer, the specific connectivity and activation parameters of the neural network are responsible for propagating that input into specific signals at the output layer. Networks may consist of excitatory and/or inhibitory units whose connectivity may range from purely feedforward to purely feedback at any layer in the network (Fig. 3-1). Finally, neural networks can be trained to respond to a specific set of input values by generating a desired set of output values ( supervised learning ) or can be trained to generate outputs given inputs whose identities are unspecified ( unsupervised learning ) (Duda et al., 2001). Given the appropriate parameters, neural networks can perform a variety of complex functions such as classification, prediction, clustering (Duda et al., 2001), and association (Hopfield, 1982) all of which are relevant to understanding the biological principles of sensory perception (Hebb, 1949; Hopfield, 1982; Arbib, 2003). One of the most famous examples of a feedforward network with supervised learning is the perceptron. Perceptrons were originally devised by Rosenblatt, who was inspired by the problem of pattern recognition in the retina (Rosenblatt, 1958). Influenced by Hebb s concept of learning in cell assemblies, in which the connections between cells are strengthened in a manner dependent on their activity (Hebb, 1949), a simple perceptron consisting of only an input and an output layer could be trained to produce a reliable output in response to repeated presentations of an arbitrary stimulus pattern. In this scheme, learning converges on a specific set of 107

114 connection weights between the input and output layers for a given stimulus (Rosenblatt, 1958; Thomas and McLelland, 2008). It was subsequently discovered that two-layer perceptrons exhibit a number of deficiencies. The computational abilities of perceptrons were rigorously mathematically analyzed by Minsky and Papert, who showed that simple two layer networks could only compute a small set of logical functions (Minsky and Papert, 1988). In particular, it was shown that only linearly separable inputs can be linearly classified; inputs that were not mutually orthogonal would interfere and not lead to distinct associations (Rumelhart et al., 1986b; Minsky and Papert, 1988). In addition, more computationally powerful functions such as XOR could not be determined without introducing multiple hidden layers. The problem of introducing such hidden layers, however, was that it was not understood how the synaptic connection weights of those layers could be modified to produce the desired outputs. A solution to such a problem came with the discovery of an algorithm that could effectively adjust synaptic weights in a multi-layer perceptron (Rumelhart et al., 1986b). If neurons in a multilayer perceptron change their synaptic weights based on the gradient descent of the sum of the squared differences between the desired output and an input target vector, and then adjust the weights of intermediate layers by backpropagating the error terms through the layers of the network, perceptrons can exhibit far more computational power than originally conceived (Rumelhart et al., 1986a; Arbib, 2003). Although advantageous in computational networks, the biological plausibility of the backpropagation learning rule is unclear (Rumelhart et al., 108

115 1986a; Stork, 1989). In the experimental data presented in this dissertation, however, the ability of the hilar cell network to maintain individual stimuli can be maintained for 10s of seconds without prior training or repetitions of stimuli, suggesting that biological feedforward networks may not require such learning mechanisms. While the individual stimuli presented in the experimental studies in this dissertation may drive different subpopulations of input neurons and thus resemble the spatiotemporal input patterns used in computational perceptrons learning mechanisms are presumably not required in the studies presented herein. Mechanisms of sequential information storage: cellular processes While it is clear that under certain conditions, both individual neurons and perceptrons may be able to represent individual stimuli, it is of interest for hippocampal information processing to understand the possible mechanisms individual neurons or neural networks may use to classify temporally organized patterns of information (Tulving and Markowitsch, 1998; Eichenbaum, 2004). A sequence is the simplest form of dynamic data and is a natural way of modeling temporal domains (Bianchini and Gori, 2001). Thus a way of addressing the manner in which individual neurons or neural networks could represent temporally organized information is to ask how neurons or neural networks can encode sequences of information. 109

116 It has been proposed that sequences of inputs on an individual neuron could be encoded in a manner analogous to how individual stimulus inputs are represented in the perceptron neural network (Gütig and Sompolinsky, 2006). The tempotron, as it has been called, represents an integrate-and-fire neuron that uses a supervised learning algorithm (Rubin et al., 2010) to classify input patterns into two categories. The tempotron receives N synapses, each with strength ω i, i = 1,..., N. Each input pattern consists of N sequences of spikes. When presented with a target pattern, the tempotron fires one or more output spikes, but when presented with a null pattern, the neuron does not spike. Much like the perceptron, the tempotron learns to respond to a target pattern by modifying its synaptic efficacies ω i whenever an error occurs. If the tempotron spikes in response to a null pattern, synapses are depressed by an amount that reflects their contribution to the erroneous output spike, while if the tempotron fails to spike in response to a target pattern, synapses are potentiated accordingly. Also by analogy with the perceptron, the tempotron uses a supervised learning algorithm to adjust synaptic weights based on the gradient descent of the error between the maximum voltage generated by erroneous patterns and the firing threshold (Gütig and Sompolinsky, 2006; Rubin et al., 2010). Although the absence of hidden layers in the tempotron absolves the requirement of a backpropagation algorithm as in the perceptron (Rumelhart et al., 1986b), questions about the biological plausibility of the learning rule in the tempotron merit discussion. While LTP and LTD are suitable mechanisms for 110

117 synaptic potentiation and depression required by the tempotron s learning rule, it is still necessary for a supervisory system to induce LTP or LTD when the target pattern is presented. Because LTP is under the control of a number of neuromodulatory systems (Foehring and Lorenzon, 1999; Otani et al., 2003; Chen et al., 2007; Hamilton et al., 2010), it has been suggested that if neuromodulation is induced during the target pattern, the tempotron may represent a biologically relevant classifier of sequences of information. However, it should be noted that the capacity of the tempotron for encoding sequences of information is limited by the number of synapses, the time constants of inputs as well as the postsynaptic integration window, and the duration of the sequences (Rubin et al., 2010). It has been postulated that sequences of information longer than ms (Rubin et al., 2010) may not be able to be classified by such a model, and that slower synaptic dynamics, multilayer architectures, or working memory mechanisms may be required (Gütig and Sompolinsky, 2006). Mechanisms of sequential information storage: neural networks Although the simplicity of the tempotron as a single neuron makes it an attractive mechanism for forming binary classification of sequences of information, neural networks have also been investigated for their potential in classifying spatiotemporal patterns (Rumelhart et al., 1986b). In addition to classifying individual inputs, such networks also classify sequences, although their capacity for doing so is not without constraints. A common approach has been to convert the temporal pattern into a spatial one by dividing the sequence into smaller 111

118 segments and then temporarily storing each segment in a buffer. These buffer segments could then be concatenated and presented to the neural network as a spatial pattern and learning algorithms such as back-propagation could then be applied (Rumelhart et al., 1986b). However, a number of problems exist with such a procedure: the buffer must be sufficiently large to accommodate the largest subsequence; the buffer must be specified in advance; and the conversion of temporal inputs into spatial ones obscures any potential correlations such as synchrony (Bressloff and Taylor, 1992). It has been proposed that introducing a hidden layer of time-summating neurons (Reiss and Taylor, 1991) can circumvent the problematic requirement of a static buffer. In this approach, the activation state of each neuron in the layer at time t 1 s is delayed and fed back onto the neuron to determine its activation state at time t (Jordan, 1986). Consequently, each neuron maintains an activity trace consisting of a decaying sum of all previous inputs to that neuron (Elman, 1990). The inclusion of such a layer eliminates the need for a buffer and permits the network to prolong information propagation at long temporal scales. Moreover, biologically implausible learning rules such as backpropagation are not required (Bressloff and Taylor, 1992), especially when autoassociative networks (Hopfield, 1984) are used to generalize sequence representation to arbitrarily complex sequences of information. The constraints imposed on neural networks to process sequentially organized information has led to the conclusion that such networks must be organized in such a way as to retain a memory of short-term information (Elman, 112

119 1990; 1991; Henson, 1998). Coincidentally, it has long been recognized that an important feature of working memory it its capacity for maintaining representations of serially ordered events at short time scales (Botvinick and Plaut, 2006; Botvinick and Watanabe, 2007). As in models for encoding sequential information, most computational models of working memory also require feedback, or recurrent, connections to sustain patterns of neuronal discharges (Camperi and Wang, 1998; Constantinidis and Wang, 2004; Compte, 2006). Although utilized for different purposes in sequence-encoding networks and simple network models of working memory for individual stimuli, feedback connections have long been considered integral to the function of both. Persistent activity and sequential information: cellular processes Two critical questions in understanding a network for working memory are: to what extent is the network capable of representing sequential information, and what cell-autonomous or network properties allow it do so? As previously discussed, experiments in nonhuman primates have established that critical to the function of working memory is the presence of persistent activity in neurons during a mnemonic behavioral task (Funahashi et al., 1989; Goldman-Rakic, 1995). It has been observed in a number of biological systems that such persistent activity can arise from cell-autonomous mechanisms as well as network mechanisms. In the entorhinal cortex, for example, layer V pyramidal neurons exhibit sustained persistent spiking activity in the presence of carbachol, an agonist of muscarinic receptors (Egorov et al., 2002), which is thought to be 113

120 due to activation of a self-sustaining Ca 2+ conductance that is tonically suppressed by a slow AHP (Fransén et al., 2006). Cholinergic modulation of CA1 pyramidal neurons has also been shown to prolong plateau potentials triggered by brief depolarization, presumably due to activation of a cyclicnucleotide gated conductance (Kuzmiski and MacVicar, 2001). It has also been shown in several cortical regions that pyramidal cell dendrites are capable of exhibiting broad spikes or plateau potentials in response to NMDA receptor stimulation (Schiller et al., 2000; Polsky et al., 2004). While cell-autonomous mechanisms can account for persistent activity modes in these brain areas, far more theoretical and experimental work has considered persistent activity arising from the structural properties of the network. Motivated by the observation that cortical cells in prefrontal cortex tend to exhibit recurrent connectivity, many investigations of persistent activity in neural networks have employed such a recurrent connectivity to account for the prolonged discharges of individual elements (Durstewitz et al., 2000; Wang, 2001; Brody et al., 2003b). However, recurrent networks of persistent activity have been implicated brain areas whose function is purely motor and not likely mnemonic. For example, it has long been recognized that bistable persistent activity in pre-motor neurons is critical to maintaining gaze and the stability of eye movements (Seung, 1996; Seung et al., 2000). While it is clear that simple recurrent networks may be able to generate patterns of persistent activity, it is unclear to what extent such networks might be able to maintain sequentially 114

121 organized inputs found in biological networks of working memory (Batuev, 1994; Baeg et al., 2003; Brody et al., 2003a). Persistent activity and sequential information: neural networks In order to account for persistent firing modes driven by temporally varying inputs, researchers have investigated a number of neural networks whose architecture is not constrained by recurrent connectivity motifs. For example, (Maass et al., 2002) investigated how a liquid-state machine, a neural network consisting of randomly connected elements that can represent, at any moment in time, present as well as past inputs, can generate a diversity of temporal patterns. Although that network can flexibly accommodate a range of complex sequences, its performance is constrained by the biophysical time constants of individual elements (Maass et al., 2002). Similar results have been obtained in neural networks that do not behave as liquid state machines but that still exhibit random connectivity (Buonomano, 2000). In order to accommodate a diversity of temporal responses at longer timescales, it has been suggested that nonrandom or specialized architectures may be required (Goldman, 2009). An important innovation in exploring persistent activity modes in networks exhibiting feedforward architectures has been re-evaluating the meaning of persistent activity. In a recent network model, persistent activity is not considered as the self-sustaining activity of individual neurons, but rather as a sequence of activity propagating through a feedforward network (Goldman, 115

122 2009). In this schme, a network consists of N neurons and each neuron receives input from earlier neurons and acts as a low-pass filter of this input with an exponential time constant τ. Each output neuron can receive its input directly or indirectly through any number of n = 1, 2, 3, or more intermediate stages formed by connections in the network. The total contribution of all pathways that travel through n intermediate stages can be described by a cumulative weight W n. Each output unit then generates a response that is dependent on W n. If there are multiple output units and appropriate synaptic weights between intermediate stages, such a network can persistent neural activity over a timescale much longer than the intrinsic neuronal or synaptic time constants. Interestingly, such a feedforward network architecture can actually be implemented as a feedback network as well, suggesting that extending the effective time constant of propagating activity may be a more influential factor in generating modes of persistent activity than requiring recurrent connections between individual elements. A hallmark of model is that such a network could generate a rich repertoire of patterns of temporal activity at time scales relevant in working memory (Durstewitz et al., 2000; Goldman, 2009). Persistent activity and sequential information storage: synaptic mechanisms While the above mentioned networks can exhibit, to varying degrees, properties required of working memory tasks, a great deal of research has indicated that synaptic plasticity can underlie memory functions in a number of systems (Bailey et al., 2004). In computational models of neural networks that use distributed 116

123 population representations (Georgopoulos et al., 1986) to encode information about a dynamic sensory stimulus, it has been shown that by modifying synapses in a manner inferred from experimental studies of LTP, a network can generate sequential modes of activity (Abbott and Blum, 1996). This result implies that with repeated presentations of a sequential stimulus, a network employing a temporally asymmetric synaptic modification rule, such as that found in NMDA-mediated LTP, could learn to respond in a spatiotemporally specific manner (Abbott and Blum, 1996) that could be relevant for motor planning. Although indicating an important role for NMDAR conductances in generating sequential representations of information, the training required of such a networks obscures its role in working memory processes (Baddeley, 1992). A computational model of an individual neuron in which trains of dendritic inputs produce regenerative release of calcium from internal stores (Ross et al., 2005) has also been implicated as a mechanism for generating epochs of persistent activity whose duration far exceeds conventional membrane time constants (Loewenstein and Sompolinsky, 2003). Although offering more biological plausibility than an LTP-based network, such a model provides persistent output activity while at the same time abolishing any temporal structure that may be present in its dendritic inputs (Loewenstein and Sompolinsky, 2003; Gütig and Sompolinsky, 2006). A number of computational models, however, have explored synaptic mechanisms between these two extremes. Based on the observation that neurons in prefrontal cortex, an area implicated in working memory (Funahashi et 117

124 al., 1989) tend to exhibit facilitating synapses (Thomson, 1997; Wang et al., 2006), Misha Tsodyks and colleagues explored how individual stimuli could be persistently represented in a recurrent network where synapses undergo shortterm plasticity (Mongillo et al., 2008). In this simulated network, a single transient excitatory stimulus is delivered to the network, resulting in an increase in population activity for the duration of the input and an increase in the connection strength of excitatory synapses. Following the stimulus offset, population activity returns to its baseline spontaneous rate, but the synaptic connection strengths remain elevated, mimicking the biologically observed phenomenon of residual calcium elevation in presynaptic terminals (Katz and Miledi, 1968). If the spontaneous baseline activity is artificially increased to a critical amount, transient stimulus inputs can recapitulate the persistent activity modes observed in other models of working memory (Durstewitz et al., 2000). As opposed to those models, however, the network requires that the synaptic connections between distinct non-overlapping subpopulations of neurons undergo potentiation by a Hebbian learning rule. Thus the capacity of the network for maintaining representations of distinct stimuli depends on the number of these subpopulations (Mongillo et al., 2008). At the same time, the presence of such subpopulations allows multiple items to be stored independently of each other, providing a natural way of representing sequences of transient inputs (Mongillo et al., 2008). A recent study in nonhuman primates (Barak et al., 2010) has provided some evidence that facilitating synapses may underlie information storage during 118

125 WM tasks. Animals were trained to a vibrotactile task (Brody et al., 2003a) in which two stimuli were presented to the index finger separated by a variable delay of ms. Following the second stimulus, the monkey was trained to press a lever to indicate which of the two stimuli had a higher frequency. Population activity profiles of many single units obtained from the monkey s prefrontal cortex during the task were then computed. It was shown that the population state accurately encoded the stimulus frequency immediately following the stimulus and at the end of the delay period, but that the frequency could not be as accurately represented during several parts of the delay period. In other words, the neurons exhibited variable tuning to the stimulus during the delay period. It was shown that a model incorporating facilitation-based neural mechanisms could account for the variable temporal profile of the tuning curves, suggesting that short-term synaptic plasticity might underlie the mnemonic properties of neurons in prefrontal cortex (Barak et al., 2010). Sequential information storage: state-dependent networks A number of groups have identified state-dependent network responses, where the trajectory of activity following a stimulus depends on previously presently stimuli (Buonomano et al., 1997; Buonomano, 2000; Buonomano and Maass, 2009). (Buonomano and Merzenich, 1995) developed a neural network model consisting of excitatory and inhibitory elements and incorporating slow IPSPs and paired-pulse facilitation. Excitatory and inhibitory elements were randomly connected in a integrate-and-fire network resembling cortical layers IV and III. 119

126 Inputs to the network were given as a two pulses separated by an interval of 80 to 280 ms. The authors observed that 25 to 50% of units projecting to an output layer could produce interval specific responses. They showed that because the slow IPSP and PPF impose time-dependent changes in the network, the network is in a different state at the arrival of the second pulse. More concretely, the firing probability of some of the units in the network had changed depending on the interpulse interval. Thus such a network can discriminate temporal information. However, the maximum time that can be discriminated (the maximum interpulse interval) is constrained by the time constants of the network: intervals longer than 280 ms cannot be discriminated. Interestingly, however, the network was able to discriminate more than a single temporal interval and, in fact, could discriminate a sequence of four pulses (giving three interpulse intervals). Unlike the experimental results presented in this dissertation, however, the interstimulus interval in this model was a fraction of 1 second, far shorter than the timescale observed in experimental studies of working memory (Fuster and Alexander, 1971). State-dependent networks have also been investigated experimentally. (Buonomano et al., 1997) demonstrated that CA3 neurons in hippocampal slices could respond selectively to the first or second of a pair of input pulses separated by 100ms. Input pulses were delivered to the dentate gyrus to activate fibers of the perforant pathway. When an associative long-term potentiation protocol in CA1 neurons was combined with the perforant path pair of input pulses, the authors were able to potentiate different populations of CA3 neurons. In this 120

127 manner the protocol was able to demonstrate that different populations of CA3 neurons are able to represent the context of incoming inputs. Context emerges as the degree of facilitation induced by an incoming input, which in turn depends on whether or not the input stimulus was preceded by an input pulse 100 ms earlier (Buonomano et al., 1997). Thus, synaptic plasticity mechanism within the hippocampal neural network alone can provide a substrate for representing the temporal context of incoming information at short (~100 ms) time intervals. A subsequent model of discriminating temporal information patterns suggested that short-term plasticity combined with dynamic changes in the balance of excitatory-inhibitory interactions can enable a network to discriminate a wide range of temporal intervals (Buonomano, 2000). In such a network, the variability in temporal tuning relies on the variability of synaptic tuning. Thus a network with a robust population code can encode temporal information for a wide range of intervals. The same network was also able to encode simple temporal sequences. However, discrimination of temporal sequences required prior training. Similar results encoding temporal sequences have been obtained in vitro, especially after training with many repetitions of specific stimuli (Johnson et al., 2010). In vivo, Nikolic and colleagues (Nikolic et al., 2007) have demonstrated cortical network activity that could reflect the identity of the first stimulus in a sequence of visual stimuli even after that stimulus was no longer present, suggesting that state-dependent networks may underlie the fading memory of information. 121

128 A proposed mechanism for information storage in the dentate hilus While both the individual properties of neurons and the organizational structure of the networks to which they belong clearly influence how neural circuits process information, it is unclear what mechanisms underlie the ability of a local circuit to represent both individual stimuli and sequences of stimuli over 10s of seconds. I have discussed previous experimental and computational studies suggesting how a local circuit may be able to process such information, and while they shed insight on potential mechanisms for encoding information in this study, they are not without caveats. In brief, no study to date has discussed how a circuit could be organized to maintain information with all the features presented in this dissertation: neural activity at long time scales (>10 s), selectivity to individual stimuli, selectivity to the order or sequential stimuli, and invariance to the temporal interval of those sequences. Therefore a critical question in understanding this work is the extent to which the known properties of the dentate gyrus can account for the data presented herein. In the following, I propose a hypothetical mechanism, based on experimentally observed properties of the dentate gyrus, that can account for the results presented in this dissertation (Hyde and Strowbridge, 2012). As previously discussed, input to the dentate gyrus arrives from the entorhinal cortex via the perforant path. In response to subthreshold stimulation, semilunar granule cells exhibit sustained depolarized plateau potential on a time scale of 10s of seconds (Larimer and Strowbridge, 2010). In response to the same subthreshold stimulation, however, some granule cells are greatly inhibited 122

129 (Larimer and Strowbridge, 2010). This functional inhibition of granule cells may be due to feedforward or feedback inhibition arising from dentate gyrus interneurons (Larimer and Strowbridge, 2010). As previously discussed, there are a variety of interneurons in the dentate gyrus, including MOPP cells in the molecular layer, HIPP cells in the hilus, and basket cells throughout out the granule cell layer (Houser, 2007). Although the function of these and other interneurons in the inhibitory control of granule cells has been proposed by a number of studies (Coulter and Carlson, 2007; Houser, 2007), semilunar granule cells have been only recently re-discovered (Williams et al., 2007), and thus distinctions between granule cells and semilunar granule cells as the postsynaptic targets of interneurons in the dentate gyrus are scarce. Both granule cells and semilunar granule cells are polarized projection neurons with dendritic arbors in the molecular layer of the dentate gyrus, and both monosynaptically excite hilar interneurons and mossy cells (Williams et al., 2007). In addition, both express the Prox1 marker, suggesting that they share a common neural progenitor (Gupta et al., 2012). However, in contrast to granule cells, semilunar granule cells extend a much wider dendritic arbor in the molecular layer, exhibit less spike-frequency adaptation, and exhibit a lower input resistance (Williams et al., 2007). Thus, it is conceivable that inhibitory control of semilunar granule cells may be distinct from that of granule cells. Indeed, a very recent study indicated that semilunar granule cells demonstrate inherent differences in inhibition from granule cells (Gupta et al., 2012), which may 123

130 account for a distinct contribution of semilunar granule cells in experimental models of brain injury and epilepsy (Gupta et al., 2012). Because some granule cells are effectively silenced during SGC plateau depolarizations, contributions of GC activity to information processing in the dentate hilus will not be considered further. Subthreshold activation of SGCs drives persistent synaptic barrages onto both mossy cells and hilar interneurons (Larimer and Strowbridge, 2010), and while it is known that SGCs form monosynaptic connections with both these cell types, currently no reports exist that can address more detailed questions regarding distinctions between hilar interneurons and mossy cells during synaptic barrages. Thus, whether hilar interneurons and mossy cells receive barrages with same frequency, amplitude, and kinetics, are currently open questions. Presumably such differences could arise from the strength, number, and divergence of synaptic connections between SGCs and hilar cells. In a landmark study, (Acsády et al., 1998) combined in vivo intracellular labeling of granule cells, immunocytochemistry, and electron microscopy to reveal important differences in connectivity between granule cells and their postsynaptic targets. They demonstrated that a single granule cell forms large complex mossy synapses (Claiborne et al., 1986) on approximately 7-12 mossy cells. In contrast, 10 times more connections, consisting of filopodial extensions of mossy terminals and small en passant boutons (Claiborne et al., 1986), were found between granule cells and GABAergic interneurons (Acsády et al., 1998). Thus, while granule cells form 124

131 divergent connections with hilar cells, that divergence is much higher for interneurons than for mossy cells. Because there have been no attempts to address similar questions regarding the divergence of semilunar granule cells onto hilar neurons, it is unclear if interneurons and mossy cells differentially contribute to processing information during synaptic barrages. As already mentioned, granule cells and semilunar granule cells, although different, share a number of genetic and morphological properties. Therefore it may not be unreasonable to predict on the basis of what is known for granule cells that a single semilunar granule cell displays a similar contrast in its divergence onto hilar interneurons and mossy cells (Fig. 3-2). Mossy cells are the predominant cell type in the hilus (Amaral, 1978), accounting for approximately 64% of the total population of hilar neurons (Buckmaster and Jongen-Rêlo, 1999), so interneurons can only account for roughly one-third of the hilar cell population. Thus, assuming a high rate of connection between semilunar granule cells and hilar interneurons, a single interneuron must be innervated by a sizeable fraction of the population of semilunar granule cells. By a similar argument, the larger size of the mossy cell population and the lower connection probability between a single semilunar granule cell and the cells of that population argue that a single mossy cell must be innervated by a much smaller fraction of the population of semilunar granule cells. This potential asymmetry in connectivity between SGCs and mossy cell and hilar interneurons has important consequences for how hilar neurons may process information in individual stimuli as well as in sequences (Fig. 3-2). 125

132 Before considering those consequences, another critical element in the hilar circuit must be clarified. Intracellular quadruple recordings in brain slices of the dentate gyrus have indicated that there is not only a very low probability of connection between two mossy cells (0.5%), but that there is also a low connection probability between two hilar interneurons (5.9%) (Larimer and Strowbridge, 2008). By contrast, there is a higher probability of connections from hilar interneurons to mossy cells (14.9%), and a somewhat lower probability for connections from mossy cells to hilar interneurons (6.1%). However, more than half (57%) of these latter connections are part of reciprocal connections back to mossy cells (Larimer and Strowbridge, 2008). Thus hilar interneurons exhibit a high degree of divergence to mossy cells. In response to subthreshold activation of PP fibers, SGCs drive persistent synaptic barrages onto hilar neurons. The long duration of SGC plateau potentials are due to the presence of intrinsic conductances that allow them to fire APs persistently, and the duration of these plateau potentials is correlated to the duration of persistent EPSPs seen in hilar neurons (Larimer and Strowbridge, 2010). In this dissertation I have shown that the synaptic barrages recorded in three mossy cells can reliably encode the identity of four distinct stimuli driving subpopulations of semilunar granule cells. By applying to semilunar granule cells the same connectivity profile of granule cells, one would expect that activating SGCs by a single stimulus (e.g. A) would result in persistent activation of a smaller fraction of mossy cells than hilar interneurons. Indeed, the high degree of divergence of SGC-HI connections and the smaller size of the hilar interneuron 126

133 population would suggest that a sizeable fraction of hilar interneurons would receive persistent barrages in response to a single stimulus. In response to activation by a different stimulus (e.g. C), the same high divergence suggests that a similar population of hilar interneurons would fire persistently. However, in response to stimulus C, approximately the same number of mossy cells would be active in response to stimulus A, but owing to the sparse SGC-MC connectivity, the population of A-responsive mossy cells and the population of C-responsive mossy cells would exhibit far less overlap than that between the A-responsive HI and the C-responsive HI populations. This result might not be expected if SGCs targeted their mossy cells in a topographical arrangement such that nearby SGC populations innervated more similar subsets of hilar neurons, but we found no correlation between the physical distance between the stimulating electrodes and the accuracy of classifying inputs (Fig 2-6d). While sparse connectivity between SGCs and MCs may be sufficient to account for the ability of the hilar cell network to encode distinct stimuli, HI may provide an important role in enhancing the capacity of the network to encode multiple stimuli. Because both stimulus A and stimulus C may activate highly overlapping populations of interneurons, the divergence of interneuron projections to mossy cells would suggest that mossy cells receive broad inhibition disynaptically from SGCs. Such broad inhibition could effectively attenuate excitatory potentials in mossy cells by slow modulation or shunting (Doiron et al., 2001; Mitchell and Silver, 2003; Prescott and De Koninck, 2003). Therefore, the already narrow range of MC input specificity established by sparse 127

134 SGC excitation could be further refined by a subtractive offset or divisive modulation of the input-output relationship (Silver, 2010). As in the visual system, divergent inhibition could effectively enhance the contrast between similar inputs, allowing the hilar cell network to maintain a greater capacity of stimulus patterns than possible by networks of purely excitatory neurons. This asymmetry in the divergence of SGC connections to mossy cells and hilar interneurons becomes even more functionally relevant as the network responds to sequences of stimuli. In order to understand this effect it is perhaps easier to consider the effect of two stimuli on a single mossy cell. For the sake of argument, we assume first that there is no inhibition and that, presented in isolation, stimulus A generates a high frequency train of barrages while stimulus B generates a lower frequency train. Presented in the sequence A-B (forward), this cell would respond with high frequency barrages for a period of seconds after stimulus A, followed by a small increase after stimulus B. In the reverse sequence, the cell would respond with a small increase above its baseline activity followed by a large increase in barrage frequency following the second stimulus. Since most persistent barrages recorded remained elevated for several seconds immediately following the stimulus (Fig 2-7), one would expect that the sum of the frequency of barrages to each stimulus would be independent of the stimulus order. That is, the sum of the barrage frequencies after the first stimulus (e.g. A in the forward sequence or B in the reverse sequence) and after the second stimulus (e.g. B in the forward sequence or A in the reverse sequence) should be roughly equivalent. 128

135 However, with divergent inhibition this would not necessarily be so. Assuming stimulus A and stimulus B contact overlapping populations of hilar interneurons, a sequence of two stimuli would evoke long plateau potentials in SGC subpopulations that would result in prolonged activation of these interneurons. However, after the second stimulus, irrespective of sequence order, hilar interneurons would be driven by nearly twice the excitatory drive from SGCs. In turn, their divergent projections to mossy cells would result in much stronger MC inhibition following the second stimulus. In this scenario, a weak stimulus B that, in isolation and without inhibition, produces a low frequency train of barrages, could be greatly attenuated when presented in the forward sequence by this strong inhibitory drive. The degree to which the response to that same stimulus B is attenuated in the reverse sequence could be much less since there is roughly half the inhibitory drive at that point. However, the degree to which the higher frequency A response is attenuated by the same strong inhibitory drive in the reverse sequence could be different as shunting inhibitory conductance can exhibit nonlinear dependencies on the membrane potential (Pavlov et al., 2009). Thus the sum of synaptic barrages evoked by the two stimuli in the forward sequence may not be equal to the sum of the barrages evoked by the stimulus presented in reverse order. The role of divergent inhibition could manifest not only as a sequence-specific effect in the detectable amplitude or frequency of excitatory synaptic potentials, but also as a modulation of the input-output relationship of the neuron. By modulating the frequency of APs, divergent inhibition could yield noncommutative representations of stimuli 129

136 both at the dendritic input processing stage as well as at the subsequent integration and output processing stage (Silver, 2010). Thus hilar neurons would be able to encode the order of the sequence, or more generally, the context of a stimulus based on its prior inputs. While this local circuit mechanism may be able to encode context of stimuli, there are several caveats. In this dissertation, I have shown that hilar cells can encode the sequence of stimuli presented at various time intervals (4, 5, and 8 seconds), but that at longer time intervals (120 seconds) sequence order cannot be maintained (Figs. 2-10, 2-13 and 2-14). There are two possible reasons for this finding. One is that short intervals (4-8 seconds) are well within the decay time constant of plateau depolarizations (Larimer and Strowbridge, 2010), so divergent inhibitory drive onto mossy cells should be unchanged at these time scales. However, at longer interstimulus intervals, inhibitory drive from a preceding stimulus would not be additive with a subsequent stimulus. Thus the effect of shunting inhibition would not depend on the order of the sequence. Another possibility is that persistent synaptic barrages from a stimulus may result in some intracellular process (e.g. Ca 2+ -induced calcium release) that could enable short-term changes in synaptic efficacy. As in a statedependent network, the response to a subsequent stimulus would be biased by this previous input (Buonomano and Maass, 2009). Although such short-term changes have usually been considered to operate at shorter time scales (e.g. < 1000ms), it may be possible that these short-term changes could nevertheless influence the temporal course or amplitude of barrages in subtle ways. 130

137 The proposed mechanism of divergent inhibition driven by sustained plateau potentials in semilunar granule cells obviously requires several assumptions about inhibitory network activity. The high degree of divergence in the synaptic connections of granule cells to hilar interneurons may not hold for the connections between semilunar granule cells and hilar interneurons. Moreover, in order to quantify the contribution of hilar interneurons to mossy cell inhibition during persistent barrages, one needs to know not only the number of connections between a semilunar granule cell and hilar interneurons, but also the physiological properties of those synapses, especially synaptic strength and degree of facilitation/depression. The same caveat applies to synaptic connections between semilunar granule cells and mossy cells, as the degree of sparse excitation may constrain the number of stimulus patterns that the hilar cell network can discriminate. Additionally, the number of inputs that a network can be observed to maintain is limited not only by sparseness, but also by experimental limitations in recording the activity of the entire hilar cell network. Thus, two stimulus inputs that produce similar patterns of persistent barrages on three mossy cells may in fact be driving more dissimilar patterns of activity in the network that are unseen by the experimental observer. Population imaging of the hilar cell network, therefore, is required to address further questions of sparseness and capacity. In addition, the effective shunting of inhibitory conductances can vary greatly depending on the location of inhibitory synapses along the somatodendritic compartment (Blomfield, 1974; Koch et al., 1983). Finally, 131

138 whether inhibitory conductances exhibit a nonlinear dependence on membrane potential can depend on a complex interplay of several factors, including Na + and AHP conductances, input resistance, synaptic background noise, and geometric constraints (Silver, 2010). Such factors may influence the response of mossy cells to hilar interneuron inhibition. 132

139 Figure 3-1 Feedforward and feedback neural networks a, Example of a multilayer neural network consisting purely of feedforward excitatory connections. The neural network consists of an input layer ( Input ), a hidden layer ( Hidden ) and an output layer ( Output ). b, Example of a multilayer neural network consisting of feedforward and several feedback excitatory projections. 133

140 a b 134

141 Figure 3-2 A hypothetical mechanism for maintaining contextual information in the dentate gyrus A proposed model of network connectivity in the dentate gyrus includes connections between semilunar granule cells (SGC), hilar interneurons (HI), and mossy cells (MC). SGCs form highly divergent excitatory connections to hilar interneurons (barbed black arrows), which in turn form highly divergent inhibitory connections onto mossy cells (blunted blue arrows). By contrast, SGCs form sparse excitatory connections onto mossy cells (barbed red arrows). For clarity, MC-HI connections are not shown. 135

142 136