Intact CA3 in the Hippocampus is Only Sufficient for Contextual Behavior Based on Well-Learned and Unaltered Visual Background

Transcription

1 Intact CA3 in the is Only Sufficient for Contextual Behavior Based on Well-Learned and Unaltered Visual Background Jae-Rong Ahn and Inah Lee* HIPPOCAMPUS 24: (2014) ABSTRACT: Computational models suggest that the dentate gyrus and CA3 subfields of the hippocampus are responsible for discrete memory representations using pattern separation and pattern completion when a modified external stimulus is recognized as an old memory or encoded as a new memory. Experimental evidence of such computational processes in the hippocampus has been obtained mostly from spatial navigational tasks, and little is known about the proposed computational functions of the hippocampal subfields in nonspatial memory tasks. We tested whether rats with major damage in the dentate gyrus induced by colchicine lesions could remember patterned visual scene stimuli presented on LCD screens in the background. Rats responded using a touchscreen to indicate the identity of the visual scene. Performance of the lesion group was normal when tested with familiar visual scenes that had been learned prior to surgery. Lesioned rats exhibited severe deficits in learning novel visual scenes, but eventually reached the same level of performance as controls. However, unlike in controls, novel scene-associated memories formed in the lesion group were unstable and easily disrupted when ambiguous versions of the novel scenes were presented intermixed with the original stimuli. Our findings confirm that the prior computational models can also be applied to the nonspatial memory domain and suggest that the dentate gyrus is not necessary for the retrieval of learned visual scene-associated behavioral responses but plays a crucial role in forming novel visual scene-dependent memory and recognizing altered or ambiguous visual scenes in the background. VC 2014 Wiley Periodicals, Inc. KEY WORDS: pattern separation; pattern completion; dentate gyrus; context; categorical memory; colchicine; scene memory; scene learning INTRODUCTION The hippocampus is important for remembering an environment in which an event took place. In particular, evidence suggests that the Department of Brain and Cognitive Sciences, Seoul National University, Seoul, Korea Grant sponsor: NIMH; Grant number: RO1 MH079971; Grant sponsor: WCU Program; Grant number: R ; Grant sponsor: Brain Research Program; Grant number: NRF-2013M3C7A and 2013R1A1A ; Grant sponsor: National Research Foundation of Korea (BK21 1 Program); Grant number: ; Grant sponsor: NIMH; Grant number: RO1 MH Research Resettlement Fund for the new faculty of Seoul National University. *Correspondence to: Inah Lee, Ph.D., Department of Brain and Cognitive Sciences, Seoul National University, Daehak-dong, Gwanak-gu, Seoul , Korea. inahlee@snu.ac.kr Accepted for publication 11 April DOI /hipo Published online 18 April 2014 in Wiley Online Library (wileyonlinelibrary.com). hippocampus plays a key role in activating discrete memory representations for similar environments that may or may not be different from each other (Gilbert et al., 1998; Tanila, 1999; Gilbert et al., 2001; Leutgeb et al., 2005; Gilbert and Kesner, 2006; Leutgeb and Leutgeb, 2007; McHugh et al., 2007; Bakker et al., 2008). Computational models have proposed that the hippocampus performs computations such as pattern completion and pattern separation, the former referring to forming and retrieving a common memory representation across slightly modified versions of the same environment and the latter referring to storing different memory representations for similar, yet different, environments. According to the computational models, the dentate gyrus (DG) of the hippocampus performs key computations for orthogonal memory representations mainly by reducing the amount of overlap as entorhinal cortical inputs enter the hippocampus (O Reilly and McClelland, 1994; Treves and Rolls, 1994). As a result, distinct neuronal ensembles may be activated in the CA3 subfield, and this may represent the underlying mechanism through which the hippocampus sends distinct memory representations as outputs to downstream structures even when it receives similar input patterns. These functions of the hippocampus can be critical when an animal encounters an environment that has been altered from its original form and must decide whether to consider the modified environment as new (i.e., orthogonalization) or old (i.e., generalization). The computational models have been supported by both behavioral and physiological experimental studies. For example, when rats were required to remember a certain location in an open space and later tested to discriminate the original location and an adjacent, yet different, location in the same environment, DG lesions severely disrupted the performance only when the two locations were close to each other, and not when they were far apart from each other (Gilbert et al., 2001). Similar results were reported when DG-lesioned rats performed an object-place paired associate task in which two identical objects should be discriminately treated depending on where they appeared in the same environment (Lee and Solivan, 2010). DG-lesioned rats also showed impairment in spatial memory tasks using a water maze (Xavier et al., 1999) and a radial arm maze (Walsh et al., 1986; McLamb et al., 1988). These results from VC 2014 WILEY PERIODICALS, INC.

2 1082 AHN AND LEE behavioral studies have been corroborated by physiological studies. In particular, granule cells in the DG exhibit very sparse discharge patterns and differentially fire for minimal changes in the geometric shape of the environment (Jung and McNaughton, 1993; Leutgeb et al., 2007). According to the computational models, the sparse firing properties of DG may ensure that a small and unique set of granule cells in DG (and also the pyramidal cells in CA3 connected to those granule cells) fire in association with a specific stimulus, increasing the chances that similar, yet slightly different, stimuli can be represented in orthogonal neuronal assemblies in the downstream structures (Treves and Rolls, 1992; O Reilly and McClelland, 1994). The aforementioned studies have some common aspects. For example, all behavioral studies focused on spatial functions of the hippocampus, whereby rats were required to navigate to different spatial locations to perform the tasks successfully (Walsh et al., 1986; McLamb et al., 1988; Xavier et al., 1999; Gilbert et al., 2001; Lee and Solivan, 2010). In the physiological experiments, only spatial or navigational information processing was examined in the DG (Jung and McNaughton, 1993; Leutgeb et al., 2007). However, the DG may also be important for remembering nonspatial and non-navigational information in an environment. For example, a contextual fearconditioning paradigm provides electric shock to an animal in association with a particular environmental background, often defined by background visual cues combined with other sensory cues present in a conditioning chamber. The animal is confined in a small conditioning chamber with transparent walls through which background cues in the room can be viewed, and is not required to navigate to any place. Thus, this is not a spatial memory task in a strict sense in comparison to typical spatial memory tasks that use radial mazes and water mazes. Nonetheless, rats with hippocampal lesions demonstrate severe deficits in remembering the background information that is associated with a reward or aversive stimulus (Kim and Fanselow, 1992; Phillips and LeDoux, 1992; Maren and Fanselow, 1997; Kim et al., 2012), and the DG appears to play a key role in this type of nonspatial memory (Nakashiba et al., 2012). DG lesions in rats also produce severe impairment in both acquiring and retrieving contextual memory (Lee and Kesner, 2004a). Rats with DG lesions demonstrate significant impairment in associating particular odors with background contexts (Morris et al., 2013). The DG (and CA3) in humans also shows differential activity levels for similar objects (pattern separation), whereas CA1 and other retrohippocampal areas exhibit similar activities (pattern completion; Stark et al., 2013). These studies collectively suggest that the role of the DG (and, in a broader sense, the hippocampus) are not limited to spatial pattern separation, but may also include nonspatial pattern separation. To investigate the role of the DG in processing nonspatial visual background more systematically, we used highly controlled visual scene stimuli displayed in the animal s background through LCD monitors. Although prior contextual studies used visual stimuli in the form of unspecified room cues in the animal s background (Kim and Fanselow, 1992; Phillips and LeDoux, 1992; Maren and Fanselow, 1997), intentionally or unintentionally, cues of different modalities (e.g., olfactory, auditory, and tactile cues) were often used together with the visual cues. We have previously shown that multiple cues are not necessary for testing hippocampal functions as long as a salient and surrounding visual scene is provided in the background via LCD screens (Kim et al., 2012). As nonspatial and visual inputs are provided to the hippocampus via the extrahippocampal cortical regions including the lateral entorhinal cortex and perirhinal cortex, we hypothesize that the DG is also critical for remembering nonspatial visual scene information. On the basis of computational models by Rolls and Treves (Treves and Rolls, 1992), we hypothesize that the roles of DG would be less critical for retrieving welllearned and familiar visual scene memories, but crucial when the familiar scenes are modified (and still need to be recognized) and when new visual scenes should be learned. MATERIALS AND METHODS Subjects Sixteen male rats (Long-Evans) weighing g were used in this study. Once obtained, all animals spent a week acclimation period in an animal colony. They were housed individually in Plexiglas cages in a temperature- and humiditycontrolled environment. Rats were kept on a 12-h light-dark cycle, and all behavioral experiments were conducted during the light phase of the cycle. For behavioral experiments, all animals were food restricted to 85% of their free-feeding weight with free access to water. All protocols were in compliance with the NIH guide for the Care and Use of Laboratory Animals and the Institutional Animal Care and Use Committee of Seoul National University. Behavioral Apparatus Behavioral experiments took place on an elevated linear track ( cm; 84 cm above the floor) with a start box ( cm) and a food tray attached at the ends of the track (Fig. 1A). An array of three LCD monitors was installed above the food tray. The center monitor (15") was equipped with a touchscreen panel (Elo Touch Systems, Menlo Park, CA) to register the animal s choice. A transparent acrylic panel with two rectangular openings was overlaid on the touchscreen panel to restrict the animal s choice response within dedicated areas on the touchscreen. On both sides of the central touchscreen, two peripheral LCD monitors (17") were present, and pattern visual stimuli (visual scene hereafter; e.g., a polka-dot pattern and a square pattern) were displayed on these peripheral monitors (Fig. 1B). Three optical sensors were installed along the linear track to record the animal s movement. Matlab (Psychtoolbox) was used to control the presentation of auditory

3 DENTATE GYRUS AND VISUAL PATTERN SEPARATION 1083 rat consumed food readily without any defecation or urination, they were transferred to a training room and familiarized to the environment for 1 day. Following the familiarization phase, rats learned to interact with the touchscreen monitor through three phases of shaping as follows. When the animal approached the center monitor, two identical rectangular images (response boxes, each cm) appeared along the horizontal meridian on the center monitor. The rat was trained to rear up and touch the response box using its front paw. In the first phase of shaping, rats were allowed to stay in front of the center monitor and were rewarded for touching either response box. Upon touching, the response box disappeared with a brief sound feedback (2 khz, 3 s, 83 db), and the rat was rewarded with a piece of cereal. In the next phase, the response box appeared in either the left or right side of the monitor to prevent the animal from developing a bias toward only one of the response boxes. In the initial part of the session, the rat was gently guided back to the start box after obtaining a food reward from the tray. As shaping proceeded, however, the rat returned to the start box voluntarily with food. In the third phase, the response box again appeared in either the left or right side of the monitor, but touching the side with no response box resulted in a low-frequency sound feedback (0.2 khz, 3 s, 83 db), and the rat was guided back to the start box with no correction allowed. This final phase of shaping ended when the animal performed six consecutive correct trials in a session. FIGURE 1. Behavioral paradigm. The experimental apparatus (A) consisted of an elevated linear track with a food tray attached at one end. Three LCD monitors were installed above the food tray, and the center monitor was equipped with a touchscreen that recorded the rat s touch responses. At the start of each trial, one of two visual scenes (polka-dot pattern or square pattern) appeared on the peripheral monitors (B) and two identical rectangular images (response boxes) appeared on the center monitor. Each scene was paired with only one of the response boxes, and rats were trained to touch the associated response box (1) for reward. A start box was positioned at the other end of the track, but is omitted for illustrative purposes. and visual stimuli, to register the animal s touch responses, and to monitor the activity of the optical sensors along the track. The apparatus was located in a sound-attenuating room and was surrounded by black curtains. The room was dimly lit by an overhead halogen light (0.2 lx) and loud speakers played white noise (80 db) throughout the behavioral session. Handling and Shaping The overall experimental schedule is shown in Figure 2. After a week of acclimation in the animal colony, rats were handled by an experimenter for 20-min per day for 5 days. As a part of the handling process, the rat was placed on a lab cart and was allowed to forage for multiple pieces of cereal (Cocoballs, Kellogg s) scattered on the surface of the cart. When the Behavioral Task Following the shaping procedure, rats were trained in a memory task. The rat was initially placed in the start box with a guillotine door and a trial started upon the opening of the start box door. The breakage of the optical sensor beam in front of the start box triggered the presentation of a visual scene through the two peripheral monitors, and two rectangular response boxes appeared in the center touchscreen at the same time (Fig. 1B). Each visual scene was paired with one of the response boxes, and the sequence of the presentation of the visual scenes was pseudorandomized. As in our prior study using a similar touchscreen paradigm (Kim et al., 2012), when the rat touched a response box, the visual scene disappeared with a sound feedback (2 khz, 3 s, 83 db for correct choices; 0.2 khz, 3 s, 83 db for incorrect choices). The sound feedback was given only after the rat made a choice response, and the procedures were used throughout all experimental conditions in the current study. A correct response was immediately rewarded with a piece of cereal (Cocoballs, Kellogg s) provided through the food tray below the center monitor. For incorrect choices, no reward was given and the animal was guided back to the start box. Each trial was separated by a 3-s intertrial interval, and a longer intertrial interval (7 s) was imposed after incorrect trials. One training session was composed of 50 trials, and animals were trained until they reached performance criterion (75% correct choice on both visual scene conditions for 2 consecutive days). On average, it took 10 sessions for the

4 1084 AHN AND LEE FIGURE 2. Schematic illustration of the behavioral training and testing schedules. Before surgery, animals were trained to criterion with the dot and square patterns. Following a 10-day recovery period from surgery, rats underwent three different testing phases. In the first phase, rats were tested with the same visual scenes (polka-dot and square scenes) for 8 consecutive days (retrieval). In the second phase, rats were required to learn new visual scenes (zebra and pebble) for 8 consecutive days (acquisition). In the last phase, ambiguous images of the zebra and pebble scenes were presented randomly intermixed with original zebra and pebble scenes (ambiguity). animals to reach this criterion, and no significant difference was found in the number of sessions to reach performance criterion between the two groups (t (14) , P ; t-test). Surgery Once the rats were trained to criterion, they were randomly assigned to either the lesion group (n 5 7) or the control group (n 5 9), and received neurotoxic lesions or sham lesions as follows. First, the rat was deeply anesthetized by intraperitoneal injection of sodium pentobarbital (Nembutal, 70 mg/kg) and fixed in a stereotaxic instrument (Kopf Instruments, Tujunga, CA). Anesthesia was maintained by isoflurane (0.5 2% isoflurane mixed with 100% O 2 ) throughout surgery. An incision was made along the midline of the scalp and the skull was exposed. The instrument was readjusted to place lambda and bregma on the same horizontal plane. Small burr holes were drilled, and colchicine (7 mg/ml, 0.5 ll/site at 10 ll/h) was injected in the dorsal DG at the following coordinates: 3.8 mm posterior to bregma, 61.8 mm lateral to midline, and 4.0 mm ventral from the skull surface. Colchicine is a potent neurotoxin that ablates most of granule cells in the DG with significantly less damage to the principal cell layers in CA1 and CA3 in the hippocampus (Walsh et al., 1986; Xavier et al., 1999). To test the feasibility of combining the neurotoxic lesion with a future electrophysiological recording study, we carried out a single injection per hemisphere as opposed to the conventional technique that involves multiple injections along the longitudinal axis of the DG. Injecting a relatively large amount of colchicine to the DG once per hemisphere resulted in bigger damage in the overlying CA1 subfield in our study (see below), although CA3 was left almost intact. Control animals underwent the same surgical procedures except that phosphate-buffered saline was injected into the DG. All injections were made using a 30-G cannula connected to a 10 ll Hamilton syringe operated by a microinjection pump (Cole-Parmer, Vernon Hills, IL). A 10 day recovery period was given after surgery before behavioral testing commenced. Post-Surgical Testing After recovery from surgery, two experiments were performed. In Experiment 1, rats were tested with the same visual scenes that had been used during the presurgical training (retrieval, Fig. 1B). In Experiment 2, animals were tested with novel visual scenes (acquisition). This was to assess whether DG lesions affected the capacity for learning novel visual scenes. Experiments 1 and 2 were each conducted for 8 consecutive days (16 days in total; Fig. 2). The day after the completion of Experiment 2, rats underwent a probe testing session in which the visual scenes used in Experiment 2 were randomly intermixed with modified (30% Gaussian blurred) versions of those stimuli (ambiguity). One animal from the control group did not perform the probe testing session due to injury. Histology After the completion of all experimental sessions, the rat inhaled an overdose of CO 2 and was perfused transcardially with phosphate-buffered saline followed by 4% formalin solution. The brain was extracted and soaked in 4% formalin-30% sucrose solution at 4 C until it completely sank. The brain was then gelatin-coated and cut in coronal sections at 40 mm on a freezing microtome (Thermo Fisher Scientific, Waltham, MA). Every third section was collected and mounted on a slide. The cut sections were dried overnight and stained with thionin (Sigma, St. Louis, MO) for verifying the extent of neurotoxin lesions. 3D Volumetry Digital photomicrographs of all Nissl-stained brain sections were taken using a light microscope at x1 magnification. A 3D volumetric analysis was conducted as described previously (Lee et al., 1999; Lee and Kesner, 2004a,b) to quantify the amount

5 DENTATE GYRUS AND VISUAL PATTERN SEPARATION 1085 FIGURE 3. Histology and quantification of neurotoxin lesions. (A) The extent of the lesion marked on serial brain slices taken from the rat brain atlas (Paxinos and Watson, 2007). The number on the right-hand side denotes the relative position from bregma (mm). Different colors represent different animals. (B C) Representative photomicrographs (x1) of Nissl-stained brain sections from a control (B) and lesioned group (C). The number on the right-hand side indicates the position from bregma (mm). (D) The amount of lesion was estimated based on 3D volumetry. DG and CA1 sustained considerable damage relative to the control (100%). Asterisks denote statistical significance (***P < ). of damage present in each hippocampal subfield. Specifically, using the 18 brain slices extending from 2 to 4 mm posterior to bregma, the boundaries of principal cell layers of hippocampal subfields including CA1, CA3, and DG were manually delineated using a professional pen tablet (Wacom, Intuos, Vancouver, WA). The traced cell layers were then colored in white (RGB value ) and sorted according to the different subfields. The extent of lesion was quantified by calculating the proportion of intact cell layers, measured in percent voxels, and compared between the control and lesion group (Voxwin, Voxar, U.K.). Animals with more than two missing brain slices (for technical errors during sectioning or collection of tissues) in the region of interest (n 5 2 in the control group) were not subjected to volumetry, and data from those animals were only used for behavioral analyses. Behavioral Data Analysis Performance was quantified as the percent correct trials in each session. A repeated-measures ANOVA was used to examine the effects of neurotoxic lesion on performance. A t-test corrected for multiple comparisons was used to reveal any difference in performance between the lesion and control group in individual conditions. The response latency from exiting the start box to touching the response box was also recorded in all sessions. Cumulative performance was calculated per session by summing all previous trials (0 for error and 1 for correct trials) and dividing the sum by the number of trials within the session. The cumulative average from the first and the last trials were not used. An F-test for variance was run on the first and last half of the cumulative performance to detect differences in performance variability between two groups. Unless otherwise noted, significant alpha levels for statistical tests were set at RESULTS Histological Verification of Neurotoxin Lesions Consistent with previous reports (Xavier et al., 1999, Lee and Kesner, 2004a,b; Lee and Solivan, 2010,b), colchicine injections into the dorsal DG ablated approximately 80% of the dorsal DG (spanning approximately 2 to 4 mm posterior to bregma;

6 1086 AHN AND LEE FIGURE 4. Post-surgical performance with learned visual scenes (retrieval). (A) Performance of control (filled circles) and lesioned (open circles) animals across 8 days of post-surgical testing in Experiment 1. Performance improved across testing days in both groups with marginal differences between the two groups. (B) The response latencies of both control and lesion groups were plotted as a function of testing days. Both groups decreased the response latencies similarly across sessions. All graphs show Mean 6 s.e.m. Paxinos and Watson, 2007) including both upper and lower blades (Fig. 3). With respect to the damage in CA1, however, the single-injection colchicine protocol used in the current study produced somewhat different results in comparison to the multiple-injection protocol used in prior studies (Xavier et al., 1999; Gilbert et al., 2001; Lee and Kesner, 2004a,b; Lee and Solivan, 2010). Specifically, previous studies showed that injecting colchicine damaged approximately 20% of the principal cell layer in CA1, and similar or lesser amounts of damage were observed in CA3 when small quantities (0.07 ml per site) of colchicine were injected at multiple locations (4 5 sites in the dorsal DG) along the anterior-posterior axis. In the current study, presumably because we carried out a single injection of a relatively large amount of colchicine (0.5 ll) per hemisphere, the neurotoxin damaged approximately half (55%) of CA1 (Fig. 3D). This is mostly likely due to an upward spread of the neurotoxin at the time of injection from the DG to the overlying CA1. However, the single-injection method yielded minimal damage in the CA3 cell layer (10%). Therefore, we find it more appropriate to call the lesions produced in this study as dorsal DG/CA1 lesions. The volume of pyramidal cell layer in CA3 was statistically not different between the lesion and the control groups (t (6) , P ; one sample t-test), whereas significant volume differences were observed in the other subfields (DG and CA1) between the lesion and control groups (P < ; Fig. 3D). DG/CA1 Lesions Minimally Affected the Retrieval of Learned Visual Scene Memory Rats were first tested with the familiar visual scenes (Fig. 1B) for 8 consecutive days. Control rats (n 5 9) started performing at around 75% and increased performance up to 95%. The lesioned rats (n 5 7) exhibited mild deficits initially, but gradually recovered the presurgical level of performance (75%) and then improved to the same level of performance as controls (Fig. 4A). A two-way repeatedmeasures ANOVA with day as a within-subject factor and group as a between-subjects factor revealed a trend for a difference in performance between the two groups (F (1,14) , P ). There was a significant effect of day (F (7,98) , P < 0.001). Both groups similarly relearned the task over time and there was no significant interaction between group and day (F (7,98) , P ). The response latency gradually decreased across days in both groups. The bigger latencies observed in both control and lesion groups (thus not likely to be associated with neurotoxic lesion per se) in the early phase of testing in the retrieval period might be attributed to the rest period following surgery. However, the latency subsequently decreased as the animals gained presurgical level of efficiency across days. According to a two-way repeated-measures ANOVA with day as a withinsubject factor and group as a between-subject factor, no significant difference was found between groups (F (1,14) , P ), but the day effect was significant (F (7,98) , P < ). The interaction was not significant (F (7,98) , P ). These findings suggest that the DG/CA1 circuits are not necessary for retrieving memory representation for well-learned visual scenes. Given our prior results showing that the dorsal hippocampus is necessary in this task, the results in turn suggest that, once acquired, familiar scene memory can be retrieved through CA3 even in the presence of significant damage to DG and CA1. DG/CA1 Lesions Severely Affected the Acquisition of Novel Visual Scene Memory After the rats were tested for the retrieval of familiar visual scene memory, we tested whether lesions in DG/CA1 affected the acquisition of a novel visual scene. In this task, rats were presented with more complex and natural visual stimuli, namely zebra-striped and pebble patterns (Fig. 5A). Control rats showed some saving in performance, showing approximately 65% correct performance on Day 1, which was significantlyabovechance(t (8) , P < 0.01, one sample t-test), whereas the lesion group performed near at the chance level

7 DENTATE GYRUS AND VISUAL PATTERN SEPARATION 1087 FIGURE 5. Post-surgical performance with novel visual scenes (acquisition). (A) New visual scenes used in Experiment 2. (B) Performance of control (filled circles) and lesioned (open circles) animals across 8 days of post-surgical testing in Experiment 2. The lesion group was slower at learning the novel visual scene than the control group. Day 9 indicates performance on standard trials in the ambiguity probe session. (C) The response latency of each group across acquisition days. In both groups, latency was maintained at similar levels across 8 days. There was no significant difference between the groups. (D) The response latency was significantly lower in the acquisition period compared to the retrieval period with no significant difference found between the two groups. All graphs show Mean 6 s.e.m. (t (6) , P ; Fig. 5B). Controls quickly learned the new visual scenes and showed performance of over 75% from Day 3 onward, but the lesion group showed sustained performance deficits (below 70%) for the first 7 days. However, lesions in DG/CA1 did not permanently impair the acquisitionofnovelvisualscenesbecause,onthelastday(day8), the performance of the lesion group sharply increased and reached the criterion level (t (6) , P , one sample t-test with the hypothesized mean at 75%; Fig. 5B). This sudden increase in performance on Day 8 was not due to random variability, as performance on these visual scenes further increased to 80% during the ambiguity probe session on the next day (see below). A two-way repeated-measures ANOVA showed significant effects of group (F (1,13) , P < 0.001) and day (F (7,91) , P < ), but no significant interaction between the two factors (F (7,91) , P ). The absence of interaction between group and day may be attributable to the delayed increase in performance in the lesion group. To address this issue, we defined Day 5 as a time point for dividing pre-asymptotic and post-asymptotic performance phases as the control group s performance showed no further improvement after Day 5 (Fig. 5B). The data from the pre-asymptotic phase (from Days 1 4) and the post-asymptotic phase (from Day 5 9) were then separately

8 1088 AHN AND LEE FIGURE 6. Post-surgical performance with ambiguous visual scenes (ambiguity). (A) Visual scenes used in the acquisition session were modified (30%-Gaussian blurred) to make ambiguous scenes. The ambiguous stimuli appeared randomly interspersed with original scenes within a single session. (B) Performance of control and lesioned animals in the ambiguity session. The performance level of the lesion group was significantly lower from the 75% performance criterion, whereas no significant difference was found in the control group s performance (sign test; *P < 0.05). (C) The response latency for each group for each ambiguity condition. There were no significant differences in the response latency. All graphs show Mean 6 s.e.m. analyzed with two-way repeated-measures ANOVAs. In the pre-asymptotic phase, there was a significant effect of group (F (1,14) , P < 0.001), day (F (3,42) , P < 0.001), but no significant interaction between the two factors (F (3,42) , P ). In contrast, a significant interaction was found in the post-asymptotic period (F (4,52) , P < 0.05) with a significant effect of group (F (1,13) , P < 0.01), and day (F (4,52) , P < 0.001) as well. The postasymptotic data were further subjected to a Bonferronicorrected t-test (alpha ), and significant performance differences between the two groups were observed from Days 5 7. We also examined whether the extent of DG or CA1 damage was correlated with the performance of the lesion group in the task, but failed to find significant correlation between the amount of lesion and performance for both CA1 (Spearman s Rho520.30, P ) and DG (Spearman s Rho , P ). The correlation between the combined lesion amount of CA1 and DG and performance was also not significant (Spearman s Rho , P ). The latency was constant throughout testing days with no significant effects found in any of the factors (two-way repeatedmeasures ANOVA, Fig. 5C). This indicates that the deficits in the lesion group were not driven by generic sensorimotor problems. The overall latency was largerintheretrievalsession compared to the acquisition session (F (1,14) , P < 0.01), but there were no significant group difference (F (1,14) , P ) and no significant interaction (F (1,14) , P ; two-way repeated-measures ANOVA, Fig. 5D). The mean response latency of the last day of retrieval (Fig. 4B) was not significantly different from that of the first day of acquisition (all P s > 0.05, two-way repeated-measures ANOVA), suggesting that the decrease in latency during the acquisition period reflects a procedural efficiency gained through repeated testing across sessions. These results strongly suggest that the CA3 network is not sufficient and that DG/CA1 circuits are critical for the rapid acquisition of a novel visual scene, even though DG/CA1 may not be critical for the retrieval of old memories.

9 DENTATE GYRUS AND VISUAL PATTERN SEPARATION 1089 FIGURE 7. Influence of ambiguous trials on standard trials. (A B) The cumulative performance for the ambiguity session plotted for the standard stimuli (STD, open circles) and the ambiguous stimuli (AMB; filled circles) for both groups. In the control group, AMB performance increased steadily to match STD performance. In the lesion group, performance plateaued at ~80% for STD and ~60% for AMB, and showed no changes across trials. (C) The mean difference between the two conditions, calculated by subtracting the performance for AMB from that for STD, plotted as a function of trial number. The difference gradually attenuated within the session in the control group (filled circles), tightly fitted to a linear function (R ), but no such trend was observed in the lesion group (open circles). DG/CA1 Lesions Severely Affected the Recognition of Ambiguous Visual Scenes It has been proposed that the DG-CA3 network plays a key role in recognizing a contextual environment modified from its original form (O Reilly and McClelland, 1994; Rolls and Kesner, 2006; Nakamura et al., 2013). To test whether DG/ CA1 lesions affected such computations in the hippocampus, we blurred the visual scene stimuli learned in the acquisition task by applying a Gaussian filter (30%) to those stimuli (Fig. 6A). The resulting visual scenes were ambiguous compared to the original stimuli and were presented intermixed with the original stimuli in pseudorandom order during the probe testing session. Rats received rewards when choosing the response box associated with the original visual scene for the ambiguous version. For example, the rat was rewarded when touching the left response box in response to the 30%-blurred zebra pattern if the correct response for the original zebra pattern was to touch the left response box. Each of the four visual scene stimuli (two original and two ambiguous stimuli) appeared 15 times, comprising 60 trials for the session. In the standard trials, in which the original visual scenes were presented, the control and lesion groups performed similarly with both groups exceeding the presurgical performance criteria (Fig. 6B). However, in the ambiguous trials, in which modified versions of the original stimuli were presented, the performance of the lesion group was markedly lower than that of the control group. A two-way repeated-measures ANOVA with visual scene ambiguity as a within-subject factor and group as a between-subjects factor revealed significant effects of group (F (1,13) , P < 0.05) and scene ambiguity (F (1,13) , P < 0.001), but no significant interaction (F (1,13) , P ). A Bonferroni-corrected t-test (alpha ) revealed a significant difference in performance between the two groups in the ambiguous trials (t (13) , P ;Cohen s d ), but not in the standard trials (t (13) , P ; Cohen s d ). A one-sample sign test also showed that the performance level for the ambiguous condition was significantly lower than 75% performance criterion only in the lesion group (P < 0.05). In other conditions, the performance was either significantly higher than criterion (e.g., STD in the control group, P < 0.01), or not significantly different from the criterion level (STD in the lesion, and AMB in the control group, P and 0.45, respectively; Fig. 6B). To examine how ambiguous stimuli influenced performance in more detail, we plotted cumulative performance for the standard and the ambiguous trials (Figs. 7A,B). In the control group, performance for the standard trials was stable at around 90% throughout the session and performance for the ambiguous trials gradually improved to approximately 80% (Fig. 7A). This suggests that the control group successfully resolved the scene ambiguity (perhaps by pattern completion) while maintaining the memory representations for the original visual scenes. In the lesion group, performance for the standard and ambiguous trials ran largely in parallel throughout the session with a more diverging trend toward the end of the session (Fig. 7B). Therefore, in contrast to the slowly rising performance for the ambiguous trials in the control group, there was almost no improvement in performance for the ambiguous trials in the lesion group. The mean difference between the two conditions (i.e., performance on standard trials minus performance on ambiguous trials) decreased linearly as the session

10 1090 AHN AND LEE FIGURE 8. The effect of ambiguity on standard scene memories. The cumulative performance on the last day of acquisition session (left) and performance for the standard condition (STD) in the ambiguity session (right) in the control (A) and lesion group (B). Note that the variability in performance increased following the introduction of the scene ambiguity on Day 9 (ambiguity session). Although the contextual ambiguity affected the initial performance for both groups, the effect was stronger in the lesion group (B). proceeded in the control group (solid line in Fig. 7C, P < ), but no significant decrease was present in the lesion group (dashed line in Fig. 7C; P , linear regression). Although performance on the standard trials was maintained at around 80% throughout the session, large variability in performance was observed in the early phase of the session, when the rat encountered the ambiguous visual scenes for the first time. An F-test for variance revealed that the variance of performance for the standard trials in the first half of the session was significantly larger in the lesion group than in the control group (F (104,119) , P < ), but this difference was not present in the second half of the session (F (104,112) , P ). The variance of performance for the ambiguous trials showed the opposite pattern. That is, it was similar between groups in the first half of the session (F (119,104) , P ) but significantly larger in the lesion group than in the control group in the second half of the session (F (104,112) , P < 0.05). The above analyses were further validated by O Brien test (Obrien, 1981) and Bartlett s test for variance, and similar results (significant differences in variances) were obtained (all P s < 0.05), suggesting strongly that the lesion group failed to show stable performance for the modified visual scenes in comparison to controls. There was no significant effect of group (F (1,13) , P ) or stimulus condition (F (1,13) , P ) on response latency, and no significant interaction between the two factors (F (1, 13) , P ; two-way repeated-measures ANOVA). We confirmed that the large variability in performance for the standard trials in the lesion group was induced by the introduction of ambiguous scene stimuli during the probe session and was not associated with incomplete learning inherited from the previous day (i.e., Day 8 of acquisition; Fig. 8B). Specifically, the variability in performance for the standard trials increased from the last day of acquisition (Day 8) to the probe session (Day 9) in both groups (Figs. 8A,B). An F-test for variance revealed that the variance significantly increased from the second half of the final acquisition session (Day 8) to the first half of the probe session (Day 9) in both groups (both P s < ), although the F-ratio was larger in the lesion group (F (119,104) 5 3.9) than in the control group (F (119,119) 5 3.4). In summary, the DG/CA1-lesioned rats were more vulnerable to modified visual scenes than the control group, who resolved the visual ambiguity rapidly and successfully. The control group experienced minimal deficit in performance for the

11 DENTATE GYRUS AND VISUAL PATTERN SEPARATION 1091 original visual scenes (standard trials) as they learned to discriminate the modified versions, but the lesion group exhibited severe perturbation in performance for the original visual scenes as they learned to discriminate the modified versions. Taken together, these results suggest that, in the absence of DG/CA1, CA3 alone cannot process visual scenes efficiently when there is a slight modification in the original visual scene. DISCUSSION In the current study, we tested whether the DG is essential for processing visual scene memory. Rats with severe damage to the DG (and some damage to CA1) exhibited almost normal performance when there was no change in the visual scene, but were markedly slower than the control rats when learning a novel visual scene. The lesion group did learn to make differential choices between the novel visual scenes; however, the performance was easily disrupted by slight modifications in the visual scene. These results confirm our previous findings that remembering a nonspatial visual scene in the background requires the hippocampus (Kim et al., 2012). The current findings also demonstrate that the functional integrity of hippocampal subfields (especially the DG) is critical when the hippocampus learns novel visual scenes and is also required to recognize an altered or ambiguous visual scene. The computational model by Rolls and Treves (Treves and Rolls, 1992; Treves and Rolls, 1994) suggests that the sparse mossy fiber inputs from the DG to CA3 are important and exert a powerful influence on determining which CA3 cell assemblies are activated during the formation of a new memory representation. The model suggests that another afferent system in CA3, namely, the perforant path inputs from the entorhinal cortex, exerts less influence during the acquisition stage of learning. The information carried by the perforant path is instead considered to enrich the contextual or episodic details of the memory content stored in CA3 so that degraded or partial cues may later retrieve the memory content (i.e., pattern completion). Therefore, the model puts more emphasis on the retrieval stage with respect to the functions of the perforant path system. The predictions of the model are supported by a behavioral study in which DG-lesioned animals were impaired in acquiring, but not in retrieving, spatial memory, and rats with electrolytic lesions in the perforant path in CA3 were impaired in retrieving, but not in acquiring, spatial memory in a purely spatial task (the Hebb Williams task; Lee and Kesner, 2004b). In the current study, DG-lesioned rats with some additional damage in CA1 were impaired in the acquisition, but not in the retrieval, of nonspatial memory of visual scene. The experimental apparatus comprised only a single linear track and the rat was simply required to touch one of the response boxes in the touchscreen to indicate its choice, meaning there was little, if any, spatial memory component to the current task. The similar pattern of results observed in the current and the previous study (Lee and Kesner, 2004b) strongly suggest that computations in the DG are indispensable for the hippocampus to quickly form a novel (nonspatial) contextual memory and to retrieve it after the original visual scene changes. It took longer for the rats with DG/CA1 lesions to learn the task when visual scenes were novel. That is, it took 7 days of training before the lesion group exceeded the 70% correct performance level, whereas it took only 1 day for the control group to reach this level of performance. However, the lesion group managed to learn the novel visual scenes with two additional days of training (no significant difference from the control group s performance for the standard stimuli on Day 9). Computational models suggest that the CA3 network serves as a site of new memory formation in the hippocampus in the absence of the DG; therefore, the slow learning in the lesion group might represent the inefficiency of the perforant path system in forming a new memory representation relative to the DG-based mossy fiber system. Although the perforant path fibers from the entorhinal cortex to CA1 and CA3 remain intact in DG-lesioned rats (Lee and Kesner, 2004b), they contain more synapses located remotely from the pyramidal cell bodies and have weaker synaptic strength than mossy fibers, and computational models hypothesize that the perforant path fibers may not be as powerful as mossy fibers in driving CA3 cellular activities. However, perforant path fibers contact more extensively with CA3 pyramidal cells than mossy fibers do, and this has led computational models to assign an effective partial cueing function to the perforant path system during memory retrieval that is particularly important for generalization during retrieval. Our results demonstrate that, in the absence of the DG, the perforant path-dependent CA3 network might acquire novel visual scene memory at the expense of learning speed. It is also possible that the remaining granule cells and their mossy fibers were responsible for the improvement in learning between Days 8 and 9 (i.e., the probe session) of Experiment 2 in the DG/CA1-lesioned rats. Other cortical areas such as the perirhinal cortex and postrhinal cortex might also underlie the inefficient, yet sufficient, learning of novel visual scenes in the lesion group. Regardless of the origin of the late learning observed in the lesion group, the non-dg-based learning was much slower and more susceptible to visual ambiguity than the DG-based learning. The lesion group had difficulty in dealing with the blurred visual scenes presented on Day 9 of Experiment 2. Computational models would suggest that the ability to choose the same response box when presented with the 30%-Gaussian blurred visual scenes may require pattern completion in CA3. The CA3 subfield is known for its autoassociative properties through its dense recurrent collateral connections (O Reilly and McClelland, 1994; Nakazawa et al., 2002; Rolls and Kesner, 2006), and is, therefore, considered a key structure in pattern completion in the hippocampus. At a glance, the lesion group s performance in our study may seem to conflict with the computational theory because CA3 was relatively intact in the lesion group, yet pattern completion for the ambiguous visual scene was impaired. However, it is important to consider that

12 1092 AHN AND LEE the absence of the DG may result in unfiltered (and possibly noisy) entorhinal inputs being passed to CA3 without proper inhibition or filtration, and this in turn may lead to faulty pattern completion (e.g., catastrophic interference) in the CA3 network. Furthermore, the absence of the DG may make it impossible for CA3 to hold well pattern-separated cell assemblies in association with different visual scenes during retrieval, and this may lead to abnormal disturbance of learned scene memories as well as newly modified ones. The highly variable performance for the standard (nonblurred) stimuli in the first half of the probe session in the lesioned rats also supports this possibility. Instead of activating the memory representations already formed in the hippocampus, the CA3 network without the DG might try to form new memory representations for the modified visual scenes in the absence of proper removal of interference, and this might produce instability in learned scene memory and inefficiency in generalizing the altered visual scenes. This suggests that pattern separation and pattern completion are two sides of the same coin, and the role of the DG in pattern separation could be important for efficient retrieval of scene memory when it comes to the retrieval of memory in the presence of stimulus ambiguity (Kheirbek et al., 2012). Colchicine has been used extensively for making lesions in the DG (Goldschmidt and Steward, 1980; Walsh et al., 1986; Emerich and Walsh, 1989; Xavier et al., 1999; Gilbert et al., 2001; Lee and Kesner, 2004b; Jerman et al., 2005; Lee and Solivan, 2010). A histological study showed that some degeneration of pyramidal cells in CA1 occurred within several days of colchicine injection into the DG, but there was almost no neuronal degeneration in CA3 (Jerman et al., 2005). In our study, we did not make multiple injections along the anteriorposterior axis of the hippocampus, but instead performed a single injection of a relatively large quantity of colchicine in each hemisphere. This single-injection method produced more CA1 damage in the hippocampus than the conventional multipleinjection method. Although it is unclear why colchicine produced more CA1 damage when being injected in a single dose of large quantity, it is important to note that there was minimal damage in CA3, as has been shown in prior studies. A dominant theory for explaining the hippocampal function in event memory formation and retrieval assigns an associative role to the hippocampus for combining nonspatial and spatial information (Moser and Moser, 1998; Hargreaves et al., 2005; Knierim et al., 2006; Kerr et al., 2007). The theory hypothesizes that nonspatial information is streamlined to the hippocampus via certain areas such as the perirhinal cortex and the lateral entorhinal cortex, whereas spatial information is conveyed via relatively separate areas including the postrhinal cortex and the medial entorhinal cortex. The spatial information-processing stream is presumably critical for spatial navigation and the nonspatial stream may be essential for remembering which items (e.g., objects, odors, sounds, etc.) were encountered in different places during the navigation. The theory, however, does not put forth a definitive statement regarding whether remembering nonspatial information alone (such as in the simple visual discrimination of scenes required in the current study) requires the hippocampus (Prusky et al., 2004; Kim et al., 2012; Yoon et al., 2012) or whether spatial memory (such as that required in a maze task) is prerequisite for processing nonspatial information in the hippocampus. The latter hypothesis may be supported by numerous experimental studies showing that the hippocampus is critical for paired associative memory for remembering items and their associated locations (Gilbert and Kesner, 2002; Day et al., 2003; Lee and Kim, 2010; Lee and Solivan, 2010; Kim et al., 2011; Yoon et al., 2012). With respect to whether nonspatial information processing alone requires the hippocampus as powerfully as the spatial memory task does, numerous behavioral studies on the roles of the hippocampus in a contextual fear-conditioning paradigm (Kim and Fanselow, 1992; Phillips and LeDoux, 1992; Maren and Fanselow, 1997; Lee and Kesner, 2004a) strongly suggest that navigation may not be always a necessary condition for recruiting the hippocampal circuits for memory. The results of the current study, along with prior studies (Otto and Eichenbaum, 1992; Wood et al., 1999; Sauvage et al., 2008; Kim et al., 2012; Yoon et al., 2012), also corroborate the idea by showing that a nonspatial task may critically recruit the hippocampus as long as the task requires the animals to associate a certain event with a distinct background (Hirsh, 1974; Lee and Lee, 2013). REFERENCES Bakker A, Kirwan CB, Miller M, Stark CE Pattern separation in the human hippocampal CA3 and dentate gyrus. Science 319: Day M, Langston R, Morris RG Glutamate-receptor-mediated encoding and retrieval of paired-associate learning. Nature 424: Emerich DF, Walsh TJ Selective working memory impairments following intradentate injection of colchicine: Attenuation of the behavioral but not the neuropathological effects by gangliosides GM1 and AGF2. Physiol Behav 45: Gilbert PE, Kesner RP Role of the rodent hippocampus in paired-associate learning involving associations between a stimulus and a spatial location. Behav Neurosci 116: Gilbert PE, Kesner RP The role of the dorsal CA3 hippocampal subregion in spatial working memory and pattern separation. Behav Brain Res 169: Gilbert PE, Kesner RP, DeCoteau WE Memory for spatial location: Role of the hippocampus in mediating spatial pattern separation. J Neurosci 18: Gilbert PE, Kesner RP, Lee I Dissociating hippocampal subregions: Double dissociation between dentate gyrus and CA1. 11: Goldschmidt RB, Steward O Time course of increases in retrograde labeling and increases in cell size of entorhinal cortex neurons sprouting in response to unilateral entorhinal lesions. J Comp Neurol 189: Hargreaves EL, Rao G, Lee I, Knierim JJ Major dissociation between medial and lateral entorhinal input to dorsal hippocampus. Science 308: Hirsh R The hippocampus and contextual retrieval of information from memory: A theory. Behav Biol 12: