Intrinsic heterogeneity of mossy cells mediates the differential crosstalk between the dorsal and ventral hippocampus

Glutamatergic mossy cells (MCs) are responsible for the associational and commissural connectivity of the dentate gyrus. MCs are widely distributed along the dorsoventral axis, but potential heterogeneity within MCs is scarcely explored. Here, we showed that MCs consist of two subpopulations which differ in their neuronal properties and functions. We discovered that MCs, depending on their dorsoventral location, extend distinct axonal projections in the molecular layers. Comparative transcriptional proling of dorsal and ventral MCs revealed different neurobiological characteristics in axon guidance and synapse assembly. Despite common activation by external stimuli, dorsal MCs, but not ventral MCs, provide net inhibitory control on granule cells across the longitudinal axis. Furthermore, dorsal MC inhibition, unlikely that of ventral MCs, increases behavioral anxiety and disables rapid contextual discrimination. Collectively, dorsoventral heterogeneity of MCs may provide a novel mechanism for functional differentiation as well as distinct association along the longitudinal extent of the hippocampus.


Introduction
The hippocampus plays critical roles in both cognitive and affective behaviors, and its dysfunctions are associated with a wide range of brain disorders 1, 2 . The long curved structure of the hippocampus extends along the dorsoventral (DV) axis in rodents and the posterior/anterior axis in humans 3 . Based on the traditional dichotomic view, the dorsal hippocampus has been associated predominantly with cognitive abilities, such as spatial and episodic memory, whereas the ventral side to a greater extent with anxiety-related behaviors 4,5 . DV differentiation of the hippocampus is causally linked to differential gene expression pro les in principal neurons as well as to neural connection within cortico-limbic structures 3,6 .
Mossy cells (MCs) are glutamatergic neurons in the dentate hilar region and are known to be vulnerable to neurotoxic insults such as ischemia, traumatic brain injury, and seizures 7 . Loss or dysregulation of MCs results in abnormalities in dentate gyrus (DG)-mediated functions including spatial memory formation, pattern separation, and mood regulation [8][9][10] . In the context of the DG circuitry, MCs receive dense excitatory inputs from nearby granule cells (GCs) through synaptic connections on their thorny excrescences 7 . On the other hand, MCs extend long-range axonal projections to the inner molecular layer (IML), across the DV axis of the ipsilateral-and contralateral DG 11,12 . This unique projection pattern suggests the possibility that MCs may integrate local GC inputs and propagate their signals across the long axis of the DG 7 . The neural circuits of MCs have the unique property of directly activating GCs 13 , or indirectly inhibiting them via GABAergic interneurons 9,14,15 . Electron microscopy studies showed that the axonal arborizations of MCs in the IML of the DG make direct synaptic connections with proximal dendrites of GCs and also of GABAergic interneurons 12,16 . Genetic ablation of whole MCs results in transient hyperexcitability of GCs 8 , whereas chemogenetic activation of MCs exert an excitatory effect on GCs 17 , suggesting considerable inconsistency in the net effect of MCs on GCs. Considering the functional heterogeneity of the hippocampus, it is plausible that MCs consist of multiple subpopulations with distinct properties and functions. However, the uniformity or heterogeneity of MCs along the DV axis remains unclear.
In this study, we found an intrinsic heterogeneity within the MC population. Our multidimensional analysis revealed that MCs are classi ed into at least two distinct subpopulations, dorsal MCs (dMCs) and ventral MCs (vMCs), which display signi cant differences in structural, molecular, and functional characteristics. This nding of drastic MC heterogeneity may contribute to a better understanding of the functional organization of the DG circuitry as well as of neural processing in the hippocampus.

DV heterogeneity within MCs in their axonal projection pattern
Based on the histological distribution of the classical MC marker calretinin (CRT), MCs are known to project to the IML along the longitudinal axis of the DG 18 . However, CRT expression in MCs is limited to ventral DG (vDG), and thus does not represent the structural features of MCs in the dorsal DG (dDG). Therefore, the axonal projections pattern of dMCs remains uncharacterized, mainly due to the absence of a valid tracing marker for dMCs. Here, we compared the axonal projection patterns of dMCs and vMCs in an animal model by labeling them with selective EGFP-and tdTomato expression, respectively (Fig. 1a). Fluorescence protein expression in adeno-associated virus (AAV)-injected Calcrl-Cre mice was highly speci c in MCs (dMCs: 95%±1.4, vMCs: 95%±1.7), but was undetectable in GCs and other interneurons such as parvalbumin-positive basket cells (PV + BCs) and hilar perforant path-associated cells ( Supplementary Fig. 1). When visualized in horizontal sections along the DV axis (Fig. 1b), tdTomatolabeled axonal bers of vMCs were found exclusively in the IML of the DG, which is consistent with the histological pattern of CRT expression 18 (Fig. 1c). In contrast, EGFP-labeled axons of dMCs were found mainly in the IML and to a lesser extent in the middle molecular layer (MML) of the dDG (Fig. 1c, d).
Strikingly, the relative density of dMC axons in these molecular layers was gradually shifted from IML to MML along the DV axis (Fig. 1c, d). Commissural axons from either dMCs or vMCs were found in the contralateral DG, and these layer-speci c projections were highly consistent in the contralateral DG, despite the reduced density of axonal bers as compared with that in the ipsilateral DG (Fig. 1e, f). Furthermore, CRT expression fully overlapped with tdTomato-labeled vMCs in the hilus as well as their axons in the IML, but not with EGFP-labeled dMCs or their axons ( Supplementary Fig. 2), con rming that EGFP-labeled axons along the DV axis originate from dMCs, but not from vMCs.
For holistic visualization of axonal projections of MCs in an intact hippocampus, we took advantage of the CLARITY technique. Our three-dimensional rendering images showed that dMCs and vMCs project to the bilateral DG through the dorsal and ventral hippocampal commissure, respectively (Supplementary Movie 1). Consistent with confocal images of hippocampal sections, a holistic view of MC projections displayed clear layer speci city along the DV axis of the DG (Fig. 1g). Furthermore, we noticed that longitudinal axons from dMCs and vMCs cover roughly up to 80% of the entire DG (Fig. 1g), highlighting the structural nature of MCs in the DG network. Taken together, dMCs and vMCs, display distinct projection patterns along the longitudinal axis of the DG (Fig. 1h).

Spatial distribution and cell quantities of two distinct MC subpopulations
The hippocampus is organized into multiple functional domains, which often exhibit sharply demarcated borders or linear gradient patterns along the DV axis 3 . Thus, we examined the spatial distribution of each MC subpopulation along the DV axis by double staining with GluR2/3 as a pan-MC marker and CRT as a vMC marker (Fig. 2a, b). dMCs (GluR2/3+|CRT-) were found in the dorsal one-third of the whole DG. They were well-segregated from vMCs (GluR2/3+|CRT+), with a narrow borderline (bregma: -2.1 mm, D-V) where the two populations intermingled (Fig. 2c). As a rough estimate, a total of around 6,122 ± 157 MCs (GluR2/3+) were counted per hippocampus (2487 ± 143 dMCs and 3635 ± 119 vMCs), indicating the presence of fewer dMCs than vMCs (Fig. 2d). Our proportion analysis showed that the dMC subpopulation was smaller than the vMC one (Fig. 2e). Taken together, these results suggest that MCs are divided into at least two subpopulations, which are spatially segregated with a narrow border line inbetween.
Transcriptional heterogeneity of MCs along the DV axis of the DG To investigate the molecular heterogeneity between dMCs and vMCs, we utilized the translating ribosome a nity puri cation (TRAP) method in combination with RNA-seq analysis. The Calcrl-Cre driver is highly cell-type speci c in AAV-injected adult animals, but not in double transgenic animals with oxed (ribosomal subunit L10a) EGFPL10a, most likely due to transient CA3 expression from the Cre-cassette during the earlier development stage. Alternatively, we adopted the Drd2-Cre driver to ensure higher EGFPL10a labeling speci city of the MC transcriptome (Drd2-Cre; oxed EGFPL10a) (Fig. 3a) 10 . In order to differentiate the transcriptome of each MC subpopulation, we divided the hippocampus into the dorsal one-third and the remaining ventral part (Fig. 3b), based on their distribution pattern along the DV axis ( Fig. 2). Our quantitative PCR analysis of each fraction veri ed that Gria3, encoding the pan-MC marker GluR3, was expressed in both dMCs and vMCs, whereas Calb2, encoding the vMC marker CRT, was expressed at a much higher level in vMCs (Fig. 3c), consistent with Allen Brain Atlas (ABA) in situ hybridization (ISH) data ( Supplementary Fig. 3a). Parallel RNA-seq analysis of both fractions identi ed differentially expressed genes (DEGs) in each MC subpopulation (Supplementary Table 1). A total of To categorize differential functional characteristics for each MC subpopulation based on transcriptomic features, we reviewed the annotations of all dMC and vMC DEGs in Mouse Genome Informatics database. Functional enrichment analysis revealed distinct biological processes and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway associated with each MC subpopulation. Notably, dMCs were enriched in multiple genes associated with axonal guidance, cell adhesion, synaptic assembly, and synaptic transmission (Fig. 3f, g). In contrast, vMCs expressed higher levels of genes associated with metabolic and proteasomal processes, and synaptic recycling (Fig. 3f, g). We further reclassi ed these gene families into the following functional groups: neuronal connectivity (axon guidance, cell adhesion, and synapse assembly), synaptic transmission (neuroactive ligand-receptor and ion channels), neuronal metabolism (glucose and fatty acid metabolism), and cellular processes (oxidative stress, proteasome, and exocytosis) (Fig. 3h, i). To single out individual genes in these functional categories, we ranked the top 20 genes based on absolute fold enrichment between MC subpopulations (  Table 2). Notably, distinct subsets of axon guidance and cell adhesion molecules genes, including Eph-ephrin and plexin-semaphorin family genes, were found to be differentially expressed in dMCs and vMCs ( Fig. 3j; Supplementary Fig. 3c). Altogether, our genomic approach revealed molecular heterogeneity within MCs, which represents the distinct neurobiological characteristic of each MC subpopulation.
Net inhibitory control of dentate GCs by dMCs but not by vMCs To examine whether external stimuli activate each MC subpopulation differentially, we exposed the animals to contextual fear conditioning (CFC) (Fig. 4a). The acute induction of c-Fos, a neural excitation marker was utilized to monitor the activity change of MCs along the DV axis. Our results suggest that dMCs and vMCs are both likely activated by contextual stimuli (Fig. 4b, c), regardless of their DV location. Next, we investigated how each MC subpopulation in uences GC activity along the longitudinal axis of the DG. Floxed control and Gi-coupled designer receptor exclusively activated by designer drugs (Gi-DREADD) AAVs were injected into either the dorsal (D-control and D-hM4Di) or ventral (V-control and V-hM4Di) hilus for chemogenetic inhibition of each MC subpopulation (Fig. 4d, e). Our previous electrophysiological recordings have shown strong inhibitory control of MC activity with the clozapine-Noxide (CNO)/Gi-DREADD system 10 . After CNO injection, the animals underwent CFC, and GC activation was assessed based on c-Fos immunoreactivity in a series of DV sections (Fig. 4f). Interestingly, acute inhibition of dMCs facilitates CFC-induced GCs activation throughout the DV extent of the DG, which was not reproduced by equivalent control of vMCs (Fig. 4g, i), indicating the selective role of dMCs in GC inhibition along the DV axis.
GCs receive main excitatory inputs from the entorhinal cortex, but their excitations are controlled by tight perisomatic inhibition from GABAergic PV + BCs, which plays a key role in several higher functions, such as feedforward, feedback inhibition, and pattern separation 19,20 . Therefore, we postulated that longitudinal projections of dMCs provide stimulatory inputs onto PV + BCs, which suppress the excitability of GCs across the DV axis of the DG. To test this possibility, we analyzed c-Fos expression in PV + BCs in animals with dMC inhibition. Notably, acute inhibition of dMCs markedly reduced c-Fos immunoreactivity of PV + BCs across the DV axis of the DG, but more strongly in the vDG than in the dDG ( Fig. 4j, l). These results suggest that dMCs, upon contextual stimulus, carry out net inhibitory control of GCs, indirectly through long-range activation of PV + BCs.

Selective involvement of dMCs in anxiety regulation and pattern separation
The dDG and vDG have been differentially implicated in cognitive and affective behaviors 5 . Our observation regarding the DV regulation of GC excitability by dMCs, led us to examine the putative behavioral outcome of dMC inhibition especially in term of anxiety and contextual memory. First, we examined anxiety-like behaviors after selective inhibition of either dMCs or vMCs (Fig. 5a, b). Acute chemogenetic inhibition of dMCs, but not of vMCs, signi cantly reduced total duration in the open arms of the elevated plus maze (Fig. 5c) without altering general locomotor activity. Consistently, our open eld test results showed that exploration activity in the center zone was drastically reduced in animals with chemogenetic inhibition of dMCs, but not of vMCs (Fig. 5d), suggesting selective involvement of dMCs in anxiety-like behaviors.
Next, we examined the selective implication of MC subpopulations in contextual fear memory. We silenced each MC subpopulation prior to fear exposure and tested fear memory retrieval in the same context ( Supplementary Fig. 4a). Strikingly, neither dMC-nor vMC-inhibition in uenced freezing behavior in the given context ( Supplementary Fig. 4b), suggesting that MCs are not involved in the encoding of fear memory. We further investigated the putative role of MCs in discriminating between different contexts. Animals were conditioned in context A with foot shock for 3 days ( Supplementary Fig. 5a), which was followed by a generalization period with alternative exposure to context A and B without foot shock for 2 days ( Supplementary Fig. 5b). We then performed contextual discrimination training (fear context A, safe context B) with selective inhibition of each MC subpopulation daily for 12 days (Fig. 5e, f). All animals displayed strong freezing behaviors in context A after fear conditioning ( Supplementary  Fig. 5a). During fear generalization, these animals also developed freezing behavior in context B, an unconditioned context. We observed no visible difference in freezing behaviors in each context between the comparison groups ( Supplementary Fig. 5b). Thereafter, animals were administrated CNO prior to daily contextual discrimination training (Fig. 5f). Dorsal and ventral control groups began to distinguish contexts A and B from block 2 (Fig. 5g, h). Pattern separation was deferred by selective inhibition of either MC subpopulation (Fig. 5g, h). Notably, dMC-and vMC-inhibition delayed contextual discrimination by additional 2 blocks (4 days) and 1 block (2 days), respectively, suggesting a more signi cant role of dMCs in this rapid pattern separation (Fig. 5g, h). Indeed, pattern separation at blocks 2 and 3 was normal in the vMC inhibition group, but was signi cantly impaired by dMC inhibition, despite complete pattern separation in both groups at block 5 (Fig. 5i, j). Collectively, these results suggest functional heterogeneity within MCs, especially in affective and cognitive behaviors.

Discussion
The hippocampus is functionally differentiated along the DV axis, and this differentiation has been ascribed to differences in afferent/efferent connectivity and gene expression in principal cell types (CA1, CA3, and DG) 3,6 . In this study, we found that MCs, depending on their DV location, differ signi cantly in axonal projection patterns. We clearly visualized distinct layer-speci c projections of dMCs and vMCs along the DV axis. Unlike vMCs projecting to the IML only, dMCs innervate either the IML or the MML depending on the DV position. Topographic projections in the hippocampus are governed by a complex interplay among axon guidance and cell adhesion molecules 21 . It is notable that dMCs express higher levels of multiple Eph receptor subtype genes (Epha4, Epha5, Epha7, Ephb1) and Sema5a, PlexnA4, and Robo3, which play a key role in axon guidance. In contrast, vMCs express multiple Eph ligand genes (Efna1 and Efnb2). Interestingly, entorhinal axons expressing Epha5 are repelled from the Eph ligandexpressing IML of the DG 22 . In addition, retinal ganglion cell axons expressing ephb1 avoid the midline at the efnb2-expressing optic chiasm, thereby establishing a partial decussation 23 . Topographically organized projections of MCs to the molecular layers are attributable to differential expression of axon guidance molecules in each MC subpopulation, as well as to the DV gradient of their cognate ligand/receptors in the molecular layers.
There is considerable controversy regarding whether glutamatergic MCs excite or inhibit GCs 8,9,13,24 . Here, we found that MCs, depending on their DV location, exert different net effects on GC excitability. Our data showed that the inhibition of dMCs, but not vMCs, resulted in GC hyperexcitability across the longitudinal axis. This is caused by the suppression of excitatory dMC input to widespread PV + BCs, followed by the loss of perisomatic inhibition onto local GCs. PV + BC dendrites extend across all molecular layers and receive the majority of excitatory synaptic inputs from the entorhinal cortex and MCs. It is of note that dendritic segments of PV + BCs display non-uniform cable properties, with a 10-fold higher membrane resistance at distal dendrites (MML and OML) than at proximal dendrites (IML), which likely accounts for the stronger synaptic inputs at distal dendrites 25 . Presumably, topological projections to the MML favor dMCs to provide strong excitatory inputs to PV + BCs along the DV axis, in comparison to vMCs projecting only to the IML. Furthermore, synaptic speci city and strength are largely determined by adhesion and synaptic assembly molecules expressed in pre-and post-synaptic neurons 26 . It is intriguing to speculate that differential gene expression may underlie the synaptic speci city of dMCs and vMCs onto their post-synaptic target cells. Interestingly, dMCs are enriched in a subset of synaptic regulators, including Epha7 and Nrg3, that are known to promote the excitatory synaptogenesis onto PV + BCs in the hippocampus 27,28 . Indeed, a recent study using electron microscopy showed that dMCs innervate PV + BC dendrites to a much higher extent than vMCs do 17 . It also reported that vMCs have a net excitatory effect on GCs in a novel environment, although we could not observe similar effect in our CFC paradigm. Altogether, these observations suggest that distinct projections from dMCs and vMCs may regulate GC excitability via a differential excitatory/inhibitory balance in given conditions. Future studies to determine relative dominance in post-synaptic targets of each MC subpopulation by using electrophysiology with layer-speci c manipulation of MC axons and further quantitative electron microscopy with axonal labeling of each subpopulation are warranted.
The DG rst receives multiple sensory inputs from the entorhinal cortex, which integrates diverse contextual and emotional information, as a gateway to the trisynaptic circuit 29 . Moreover, the dDG and vDG have been differentially associated with contextual representation and anxiety control, respectively 5,30 . In this regard, it is striking that dMC inhibition results in elevated anxiety-like behaviors, which seems contradictory to DV differentiation of DG functions, in that anxiety control has been associated with the ventral portion of the hippocampus. However, we showed that inhibition of dMCs results in ventral GC hyperexcitability, which has been causally linked to elevation of anxiety responses 30 . Consistent with this, direct manipulations of PV + BC activity in the DG altered anxiety-like behaviors 31,32 . Therefore, dMCs may regulate anxiety-like behaviors with longitudinal control of the DG network, primarily through PV + BCs. It is intriguing to speculate that dMC may integrate contextual information from the dDG and convey it to the vDG for driving contextual memory associated with emotional valence.
Here, we also found that MC inhibition had no effect on memory encoding, but signi cantly delayed contextual discrimination, which consistent with the established roles of the DG in contextual processing 33,34 . The sparse activation pattern of GCs is required for context-speci c representation, which is essential for pattern separation between similar contexts 29 . Adequate inhibition of extra neural ensembles beyond the extant engram pattern is critical for discriminating input patterns in the DG. Therefore, it is important that dMCs provide indirect inhibition onto GCs through PV + BCs. In fact, dysregulation in the high-frequency ring of PV + BCs impairs spatial pattern separation in the DG 31 . Supporting this idea, computational modeling suggests that inhibition of MC-to-BC connections resulted in the recruitment of signi cantly more GCs in response to simulated entorhinal input. Interestingly, direct excitatory MC-to-GC connections have insigni cant effects on the sparseness of GC activation and on pattern separation e cacy 35 . Thereby, dysregulation of the GC excitation pattern with dMCs inhibition is responsible for impaired pattern separation, possibly with increased uncorrelated noise in the GC ensemble. Meanwhile, vMCs were reported to convey predominantly excitatory inputs to dorsal GCs 17 , which may account for why vMC inhibition has rather marginal effect on rapid pattern separation. Collectively, the differential dominance between direct (MC-to-GC) and indirect (MC-to-BC) pathways may underlie the distinct function of each MC subpopulation during memory processing. To further explore how each MC subpopulation in uences the GC ensemble pattern, future work may require animal models in which the GC ensemble is traceable in real-time in the given contexts, while under simultaneous manipulation of each MC subpopulation.
Con icting with a classical dichotomic view of the hippocampal functions along the DV axis, there is growing evidence suggesting that contextual processing depends on not only the dorsal but also the ventral portions of the hippocampus. Lesioning or pharmacological manipulations of the ventral hippocampus often impair contextual processing that was traditionally ascribe to the dorsal one 36,37 .
Furthermore, dorsal and ventral young GCs contribute to contextual-dependent memory 38 . In addition, the vDG, but not the dDG, is critical for odor discrimination 39 , supporting differential contextual representation in the dDG and the vDG. MCs receive dense excitatory inputs from nearby GCs and propagate their excitatory signals across the long axis of the DG circuitry. This longitudinal projection of MCs may mediate functional crosstalk between the dDG and vDG, and thereby enable functional integration throughout whole DG networks along the DV axis. Our ndings highlight the within-cell-type heterogeneity of MCs, as dMCs and vMCs signi cantly differ in their net effects on GC excitability ( Supplementary Fig. 6). This dorsoventral heterogeneity of MCs may represent a signi cant difference in MC-mediated crosstalk between the dDG and vDG. Our ndings may provide a novel mechanism underlying the dorsoventral differentiation of the DG circuitry, as well as the functional association of the entire DG network to control cognitive and affective behaviors.

Dual uorescence labeling of dMCs and vMCs
The following AAV stocks to express a uorescence protein in Cre-dependent manner were used for our experiments: AAV1.CAG.FLEX.EGFP (PENN vector core), and AAV2.CAG.FLEX.tdTmato (UNC vector core).
Calcrl-Cre TG mice were deeply anesthetized with an intraperitoneal (i.p.) injection of Avertin (250 mg/kg) and placed into a stereotaxic apparatus (Angle Two™, Leica Biosystems, USA). Original AAV stocks were diluted in 1×10 12 GC/ml with AAV dilution solution (5% sorbitol in 1X PBS) and 500 nl were injected into either the dorsal hilus (dHil) (-2.1 mm AP, +/-1.4 mm ML, -1.95 mm DV) or ventral hilus (vHil) (-3.3 mm AP, +/-2.7 mm ML, -3.6 mm DV). Flow rate (200 nl/min) was controlled using a Nanopump controller (World Precision Instruments, Florida, USA). The needle was left in the target region after injection for another 5 min. Calcrl-Cre TG mice were allowed at least 3 weeks of rest before the next experimental stage. Any mice with abnormal recovery after stereotaxic surgery were euthanized and thus excluded from the analysis. All injections were veri ed histologically at the end of the experiments.

Fluorescence imaging of dMCs and vMCs
Brains were prepared for immunohistochemistry and uorescence imaging, as described above. Sections were scanned with a confocal microscope (LSM 780/800, Zeiss, Oberkochen, Germany) under the same imaging conditions. The uorescent intensity was quanti ed using ImageJ (NIH, Maryland, USA). The average of uorescent intensity was measured for the region of interest for each hippocampal subregion, including the hilus, the inner molecular layer (IML), the granule cell layer (GCL), the middle molecular layer (MML), and the outer molecular layer (OML). The linescans were 28 μm (40 pixels) wide.

Three-dimensional imaging of MC projections
Clearing of the intact brain: For imaging of the whole hippocampal region, the posterior two-third of a whole postmortem mouse brain was dissected and polymerized in a 1x PBS solution containing 1% acrylamide (acrylamide:bis-acrylamide=29:1) and 0.1% Azo-initiator (VA-044, Wako) overnight, followed by polymerization for 3 hours at 37 °C (X-Clarity polymerization system, Logos Biosystems, South Korea). Polymerized tissues were cleared in an X-Clarity™ tissue clearing system II (Logos Biosystems, South Korea) for 8 hours at a current of 2 A, temperature of 37.0 °C, and pump speed of 80 rpm. Cleared tissues were stored in a refractive index matching solution (50% sucrose, 20% urea) for imaging with light-sheet uorescence microscopy.
Three-dimensional (3D) uorescence imaging: For light-sheet uorescence microscopy, we used the Ultramicroscope II from LaVision BioTec (Bielefeld, Germany) equipped with Olympus (Tokyo, Japan) MVPLAPO 0.63x lens with dipping cap, NKT Photonics (Birkerød, Denmark) SuperK EXTREME EXW-12 white light laser, Andor Neo sCMOS camera (Thorlabs, New Jersey, USA) and a customized sample holder. Scans were made at 0.63 magni cation with a light sheet numerical aperture adjusted at 0.073. For EGFP and tdTomato uorescence proteins, excitation lters of 470/40, 560/25 and emission lters of 525/50, 620/60 were used. The scan step-size was set at 3 μm and both channels were obtained in two separate scans. For the image post-processing and 3D image rendering, serial tif image les were converted to an Imaris le format and analyzed with Imaris software (Bitplane, Cologne, Germany). Homogenates were cleared by centrifugation at 2000 x g for 10 min at 4 °C. NP-40 (EMD Biosciences, California, USA) and 1,2-diheptanoyl-sn-glycero-3-phosphocholine (DHPC; Avanti Polar Lipids, Alabama, USA) were added to the supernatant at a nal concentration of 1% and 30 mM, respectively, followed by incubation on ice for 5 min. The clari ed lysates were centrifuged for 10 min at 20,000 g to pellet insolubilized materials. Monoclonal anti-EGFP antibodies (50 µg each of the clones 19C8 and 19F7) were immobilized onto Dyna magnetic beads (Invitrogen, California, USA) via protein L. These EGFP-a nity beads were added to the supernatant, followed by incubation using an end-over-end rotator for 16 hours at 4 °C. Polysome-bound beads were washed 3 times in the high-salt washing buffer (20 mM HEPES [pH 7.4], 350 mM KCl, 5 mM MgCl2, 1% NP-40, 0.5 mM dithiothreitol, 100 µg/ml cycloheximide, and RNasefree water) with gentle agitation to resuspend the beads between washing steps. The beads were washed one more time with detergent-free washing buffer and immediately added to RLT buffer, followed by RNA puri cation with RNeasy Micro Kit (QIAGEN, California, USA) with in-column DNase digestion. The quantity and integrity of puri ed RNA was determined by using a Quant-iT RiboGreen RNA Assay Kit (ThermoFisher, Massachusetts, USA) and Fragment analyzer (AATI, USA), and only samples with RNA Quality Number (RQN) greater than eight (out of ten) were used for qPCR and RNA-seq. qPCR analysis: The isolated RNA from the TRAP assay was used to synthesize cDNA with iScript cDNA Synthesis Kit (BioRad, California, USA). 10 ng of cDNA was used for each Quantitative real-time PCR (qPCR) reaction and all samples were run in triplicate. qPCR analysis was carried out using SsoAdvance Universal SYBR Green Supermix (BioRad, California, USA) and AriaMx Real-Time PCR system (Agilent, Technologies, California, USA) following standard cycling conditions (95 °C for 30 s, then 40 cycles of 95 °C for 5 s and then 60 °C for 30 s). The relative quantitation of mRNA was calculated by the comparative Ct method after normalization to mouse Map2. The following primers were used: Calb2, 5′-TTTATGGAGGCTTGGCGGAA-3′ (forward) and 5′-TCATCATAGGGCCTGTTGGC-3′ (reverse); Gria3, 5′-CTCCAAGGACAAGACCAGTGC-3′ (forward) and 5′-GTTTGGACTCTGCCCGTGAT-3′ (reverse). Non-speci c ampli cation was excluded by con rming single melting curve patterns and ethidium bromide staining on 2% agarose gels. All statistical analyses were using Student's t test.

Quanti cation of dMCs and vMCs subpopulation
Data analysis: Full-length cDNA was generated from 500 pg of RNA with Ovation RNAseq V2 kit (NuGene, California, USA). cDNA quality and quantity had been checked on a Fragment Analyzer using DNA High sensitivity assay kit (AATI, USA) prior to sequencing library preparation. Sequencing libraries were constructed by a TruSeq RNA library Prep Kit v2 (Illumina, California, USA) and sequenced for 50 bp paired-end on Illumina HiSeq2500 using Rapid V2 sequencing chemistry (Illumina, California, USA). For each library, the number of detected genes is presented in Table S1. Reads were then aligned to the mouse reference genome (GRCm38) using STAR version 2.5.2b. Read counts per gene were calculated using HTseq-count version 0.7.2. Differential expression analysis on raw read counts was performed in R using the edgeR package. The package implements exact statistical methods based on generalized linear models. The particular feature of edgeR functionality is empirical Bayes methods that permit the estimation of gene-speci c biological variation, even for experiments with minimal levels of biological replication. The quasi-likelihood method is implemented for differential expression analyses of bulk RNAseq data. We rst identi ed 'expressed' genes as the genes with counts per million fragments mapped (CPM) larger than 1 under at least two of the biological replicates. Finally, we identi ed DEGs as the ones that have the adjusted p values < 0.05 and absolute log2-fold-changes > 0.58 (1.5-fold). GO and KEGG pathway analysis were performed in DAVID software using Fisher's exact test.
Gi-DREADD-dependent MC manipulation The AAV8.hSyn.DIO.hM4Di-mCherry (PENN vector core) virus was injected in either the dHil or vHil as described above. Control animals were injected with AAV8.hSyn.DIO.mCherry. All MC-speci c and regionspeci c expression of all injections were veri ed histologically at the end of experiments. The backmetabolized clozapine from CNO may lead to activation of endogenous receptors in non-DREADD animals, thus we applied low-dose CNO (less than 1 mg kg −1 ) to avoid potential side effects in a higher dose (10 mg kg −1 ) for our behavioral testing. CNO (Sigma-Aldrich, Missouri, USA) in 0.9% saline was injected i.p. as indicated in Fig. 4 and Fig. 5.

c-Fos Immunohistochemistry in GCs and PV+BCs
For counting c-Fos positive neurons in the DG, mice received CNO (1.0mg kg −1 ) i.p 30 min before contextual fear conditioning. Then, mice explored the conditioning chamber with a foot shock. After this, mice returned to their home cage in either 90 min or 30 min to examine GCs and PV+BCs activity, respectively. Brains were sectioned coronally and horizontally from each animal. The outline of GCL and SGZ for quanti cation of GCs and PV+BCs were manually drawn by well-de ned anatomical landmarks visualized with staining DAPI. The number of c-Fos positive GCs and PV+BCs were counted on three and two representative dorsal and ventral sections, respectively. The density of neurons in the DG was calculated dividing the number of cells by the total area.

Anxiety-related behavioral testing
All experiments were conducted during the light cycle (7 am to 7 pm). Mice were handled for 1 min on three consecutive days before testing and were randomly assigned to a blinded investigator. Mice were administrated CNO i.p. 120 min before the test was conducted Elevated plus maze test: the elevated plus maze consisted of two open arms and two enclosed arms (30 × 5 cm) with a center platform (5 × 5 cm). The entire apparatus was placed 50 cm above the oor. Mice were habituated in the testing room in their home cages for 60 min before the test was conducted. Each mouse was placed in the center platform and facing the enclosed arm. The time spent in the open arm was measured using an automated video-tracking system (EthoVision XT, Noldus, USA) for 5 min. Between sessions, the apparatus was cleaned using 70% ethanol.
Open eld test: locomotor activity was measured in an open-eld area (40 × 40 × 40 cm) using EthoVision XT for 30 min. Mice were transferred to the testing room and acclimated for 1 hour before the test. The center zone index was de ned as the center travel distance (20 × 20 × 20 cm) versus total travel distance.
After every session, the open eld was cleaned with 70% ethanol.

Contextual fear conditioning
The chamber for Context A (paired with shock) was a 15 × 15 × 20 cm chamber with a metal grid, white lighting, and background noise (provided by a fan) that was covered by a transparent acrylic lid. For context B (with no shock), the chamber had a checkered cylindrical wall, grid pattern, no background noise, and the odor of 0.25% benzaldehyde (in 100% ethanol) provided using a paper dipped in the solution and placed beneath the chamber during experiments. All mice were acclimated in the anteroom for 1 hour before the test. Mice were allowed to train the context for 180 sec and then received a foot shock (2 sec, 0.5 mA), followed by a post-shock period of 60 sec. Freezing scores were measured using Freezeframe 4 (Actimetrics Software, Evanston, IL, USA). The threshold was set at 20, and the freezing bout was set at 1 sec. The chamber and grid were cleaned with 70% ethanol between sessions.
Memory encoding test: mice were administrated CNO i.p. 120 min before training context. Mice were placed in context A, followed by a foot shock and a 60 sec post-shock period. After training, mice were returned to their home cage. 24 hours later, mice explored in the shock-associated context again to recall fear memory for 4 min. Freezing scores were quanti ed for 3 min using Freezeframe 4 (Actimetrics Software, Evanston, IL, USA).
Contextual discrimination: this procedure was based on a protocol described in detail previously 34 . For contextual fear acquisition, mice were trained in context A to associate fear memory for consecutive 3 days (days 1-3). Mice were allowed to explore context A for 180 sec, were administered a foot shock (2 sec, 0.5 mA), and were then returned to cage 60 sec later. For contextual generalization, mice were placed in context A or context B without shock for 240 sec and then were placed in the opposite context for 240 sec 1.5-2 hours later (counterbalanced order) on consecutive 2 days (days [4][5]. For contextual discrimination training, mice were administrated CNO i.p. 120 min before discrimination of contexts. They were placed in context B (not paired with shock) for 240 sec. following 1.5-2 hours context B, they were placed in context A (again paired with shock) for 180 sec, received a 2 sec 0.5 mA shock, left in the chamber for 60 sec following the shock. Mice were trained to discriminate in both contexts on each day for 12 days. The alternative training orders followed a BAAB -> ABBA pattern (day 6, B → A; day 7 A → B, day 8, A → B; day 9, B → A; day 10, A → B; day 11, B → A; day 12, B → A; day 13, A → B; days 6 through 17). Freezing scores were measured for an initial explored time of 180 sec. For contextual discrimination presentation, the freezing score was combined consecutive 2 days into the block, so that each block consisted of a freezing score of consecutive 2 days in both context A and B. The inhibition of each MC subpopulation and their control group was alternatively performed.

Data analysis and statistics
All data are represented as mean ± standard error of the mean (SEM). Statistical parameters and analysis performed can be found in the gure legends and Supplementary Table 3. Statistical analyses were performed using Prism 7.0a (GraphPad, La Jolla, CA). A comparison of two groups was analyzed by Student's t-test (paired and unpaired, two-tailed). Two-way ANOVA or repeated measures ANOVA for more than two groups were used to investigate main effects and Bonferroni-corrected post hoc comparisons.