Cortical glutamatergic and GABAergic inputs synergistically support learning-driven stability of hippocampal representations

Abstract

Stability and flexibility of neuronal ensembles are crucial brain functions, notably supporting learning and memory. Little is known about how these features of local circuits are influenced by long-range inputs allowing communication across brain regions. We show that lateral entorhinal cortex glutamatergic (LEC_GLU) and GABAergic (LEC_GABA) projections to CA3 recruit specific microcircuits that conjunctively provide stability to neuronal ensembles supporting learning. LEC_GLU drives excitation in CA3 but also substantial feedforward inhibition that prevents somatic and dendritic spikes. Conversely, LEC_GABA suppresses this local inhibition to disinhibit CA3 activity with compartment- and pathway-specificity by selectively boosting somatic output to integrated LEC_GLU and CA3 recurrent inputs. This synergy allows the stabilization of spatial representations relevant to learning, as both LEC_GLU and LEC_GABA control the remapping of CA3 place cells across contexts and over time. Our findings provide circuit mechanisms whereby long-range glutamatergic and GABAergic cortico-hippocampal inputs cooperatively shape experience-dependent spatial coding in CA3.

Graphical Abstract

One-Sentence Summary:

Long-range glutamatergic and GABAergic inputs conjunctively drive compartment- and pathway-specific activity to stabilize neural representations during learning.

Introduction

The brain encodes memories via the activation of specific ensembles of neurons to support behaviors critical for survival. Formation and recall of these representations require both stability and flexibility. Within the hippocampus, place cell ensembles can rapidly change their spatial tuning in response to differences in the environment (1–4). This phenomenon, known as remapping, allows for adaptive learned behaviors (5–8). Conversely, reliable navigation and memory recall from partial cues require that hippocampal place maps remain stable, withstanding some degree of environmental change (9–13). While we know that principal neurons in the hippocampus integrate multimodal inputs from the entorhinal cortex (EC) to form these internal representations of the environment, it remains unclear how the organization and function of the long-range circuitry support the dynamic interplay between stability and flexibility of representations. Here, we fill this gap by examining the functional interaction of long-range glutamatergic and GABAergic projections from the entorhinal cortex with the local hippocampal circuit using sub-compartment level circuit mapping and in vivo two-photon imaging coupled with goal-directed behavior and cell-type specific manipulations.

Aside from canonical glutamatergic inputs, EC sends direct long-range GABAergic projections to the hippocampus (14, 15). Recent studies have shown that long-range GABAergic connections play key roles in dendritic excitability, oscillatory synchronization, synaptic plasticity, context discrimination, and top-down control (14–19). Specifically, the lateral entorhinal cortex (LEC) encodes information about contextual features such as objects, odors, novelty, and salient rewards or punishments, to support goal-directed behaviors (14, 20–27). Perturbations of LEC activity disrupt learning (28–30) and impair rate remapping of hippocampal area CA3 (31), implicating LEC-CA3 circuit interactions in supporting the flexibility of neural representations. CA3 pyramidal neurons (PNs) receive stratified dendritic input from many different sources. These include (i) the direct inputs to distal dendrites from EC conveying multisensory contextual, spatial and cognitive information, (ii) the feedforward processed inputs from the dentate gyrus (DG) to proximal dendrites thought to provide orthogonalized information (32–42), and (iii) the feedback inputs from within CA3/2 through recurrent connections (RC) upon medial and basal dendrites implicated in attractor dynamics (32, 35, 43–48). Even with the convergence of multiple inputs, as leaky integrators, CA3 pyramidal neurons do not readily display spike output (49). How long-range glutamatergic and GABAergic inputs interact with local circuitry to support network-wide computations remains an open question. Mechanistically, how are the various inputs to CA3 gated to modulate output and information transfer?

To answer this, we examined LEC long-range glutamatergic (LEC_GLU) and GABAergic (LEC_GABA) inputs to CA3 and their interactions with local circuits at the single cell and sub-cellular level, with dendritic and somatic patch clamp electrophysiology and dual color optogenetics. We found a disinhibitory gating mechanism, where long-range LEC_GABA projections target local interneurons to differentially modulate the integration of LEC_GLU inputs with DG and RC inputs, thereby influencing pathway- and compartment-specific output and related computations in area CA3. To link this circuit organization to behaviorally relevant CA3 activity, we examined the independent influence of LEC_GLU and LEC_GABA inputs on neuronal ensemble dynamics during learning and recall. For this, we used in vivo two-photon calcium imaging during a goal-oriented learning behavioral task with cell type specific-chemogenetic manipulations. Lastly, based on our experimental findings, we present a model of how long-range LEC_GLU and LEC_GABA and local CA3 circuit interactions shape learning-driven stabilization in the recurrent network.

Results

LEC_GLU and LEC_GABA drive excitation and inhibition in CA3 pyramidal cells and interneurons

LEC sends glutamatergic (LEC_GLU) and GABAergic (LEC_GABA) projections to the hippocampus, but their synaptic targets in area CA3 are unknown. To map LEC inputs to CA3, we expressed CaMKII-driven-Chronos-GFP (50) in LEC_GLU and Syn-driven-flexed-ChrimsonR-tdTomato (50) in LEC_GABA by stereotaxic injection of viral vectors into LEC of GAD2-Cre mice (Fig 1A). LEC_GLU (GFP+) and LEC_GABA (TdTomato+) axons projected to the distal dendritic layer (stratum lacunosum moleculare, SLM) of area CA3 (Fig 1B). Most experiments used this dual-optogenetic approach with appropriate parameters (Fig S1) within the same animals, and similar innervation patterns were observed with independent mapping of ChR2-eYFP-expressing (51) LEC_GLU and LEC_GABA in separate animals (Fig S2A–B). LEC_GLU innervation was denser than LEC_GABA across the hippocampus (Fig S2C). To establish functional connectivity from LEC to CA3, we optogenetically stimulated LEC_GLU and LEC_GABA while performing patch-clamp recordings from CA3 neurons ex vivo (Fig 1A). LEC_GLU stimulation elicited excitatory (EPSCs) and inhibitory post-synaptic currents (IPSCs) in CA3 pyramidal neurons (PNs) (Fig 1C). Application of TTX & 4-AP (52) spared EPSCs but abolished IPSCs, revealing that LEC_GLU drives direct monosynaptic excitation and feedforward polysynaptic inhibition in CA3 PNs (Fig 1C, Fig S2D). Application of the AMPAR and NMDAR antagonists NBQX & APV abolished both LEC_GLU-driven EPSCs and IPSCs, confirming glutamatergic transmission (Fig 1C, Fig S2E) and feedforward inhibition. In contrast, LEC_GABA stimulation failed to elicit postsynaptic responses in CA3 PNs, as probed by somatic and dendritic patch-clamp recordings (Fig 1C–E, Fig S2F–G) as well as dendritic extracellular local field potentials (Fig S2H). However, we found that a subset of CA3 interneurons (INs) with somatic location in SLM received convergent input from both LEC_GLU and LEC_GABA, using TTX & 4-AP application to confirm monosynaptic connectivity (Fig 1F–G, Fig S2I). LEC_GLU-evoked EPSCs in CA3 INs were glutamatergic, as they were blocked by NBQX & APV (Fig 1F). Application of the GABAaR and GABAbR antagonists SR & CGP abolished LEC_GABA-evoked IPSCs in CA3 INs, confirming their GABAergic nature (Fig 1G, Fig S2J). While LEC_GLU targeted most CA3 PNs and INs (Fig 1E, 1H), LEC_GABA connected only a fraction of CA3 INs (Fig 1H) with somatic location in SLM, which we further characterized as VIP- and putatively CCK-expressing (Fig 1I, Fig S3). Of note, VIP- and CCK-coexpressing INs have been described in CA1 and cortex as a subpopulation of basket cells with somatic location in distal dendritic layers and perisomatic axonal projections (53–56), suggesting that LEC_GABA may selectively disinhibit CA3 PNs somatic compartment. Overall, these data show that LEC_GLU drives excitation in CA3 PNs and INs, as well as feedforward inhibition in PNs; whereas LEC_GABA targets a subset of CA3 INs but not PNs.

Fig 1. LEC_GLU and LEC_GABA connectivity to area CA3.

(A) AAV injection of CaMKII-driven and Syn-driven-flexed opsins into LEC of GAD2-Cre mice (left) allows expression of Chronos-GFP in LEC_GLU and ChrimsonR-tdTomato in LEC_GABA projections to the hippocampus (middle), enabling input-specific dual-optogenetic circuit mapping in area CA3 with ex vivo patch-clamp electrophysiology (right). (B) Sample injection site of AAVs targeting LEC_GLU (Chronos-GFP, green, left) and LEC_GABA (ChrimsonR-tdTomato, red, right) neurons and their projections to CA3 SLM (expanded views of demarcated area), with DAPI staining (blue). (C) Top left, sample traces of LEC_GLU-driven CA3 PN post-synaptic responses in ACSF (EPSC, green; IPSC, purple), TTX & 4-AP (grey), and NBQX & APV (orange). Bottom left, sample traces of CA3 PN somas under voltage-clamp at +10 mV upon LEC_GABA stimulation. Right, input-output curves of LEC_GLU-evoked CA3 PN EPSC (green) and IPSC (purple) amplitudes. (D) Top, sample traces of CA3 PN dendrites held at −50 mV under current-clamp upon LEC_GLU stimulation (top) and LEC_GABA stimulation (bottom), note the absence of hyperpolarization in response to LEC_GABA stimulation as opposed to feedforward inhibition-mediated hyperpolarization with LEC_GLU stimulation in the same samples. (E) Fraction of CA3 PN receiving LEC_GLU and LEC_GABA inputs (χ² test, p < 0.001). (F) Sample traces of LEC_GLU-evoked CA3 IN EPSCs in ACSF (green), TTX & 4-AP (grey), and NBQX & APV (orange). (G) Sample traces of LEC_GABA-evoked CA3 IN IPSCs in ACSF (red), TTX & 4-AP (grey), and SR & GCP (orange). (H) Fraction of CA3 IN receiving LEC_GLU and LEC_GABA inputs. (I) Fraction of Pvalb (navy blue), SST (yellow), VIP (purple) and CCK (cyan blue) INs receiving LEC_GABA input (χ² test, p < 0.001). Error bars represent SEM.

Inhibition gates the recruitment of CA3 PNs by LEC_GLU

CA3 PNs are leaky integrators requiring high levels of input to reach firing threshold (49), embedded in a complex microcircuit involving DG mossy fibers (MF) and CA2/3 recurrent collaterals (RC). Therefore, the influence of LEC_GLU-driven distal dendritic excitation and substantial feedforward inhibition on their input-output transformation is not trivial. To examine the contributions of LEC inputs to CA3 output, we monitored CA3 PN post-synaptic potentials (PSPs) in response to LEC_GLU optogenetic stimulation, as well as electrical stimulations of local DG (DCG-IV sensitive, Fig S4A) and RC inputs. Sampling input-output curves revealed that, although PSP amplitude increased with input strength, CA3 PNs did not fire action potentials in response to any given input single pulse stimulation (Fig 2A, Fig S4B). However, blocking GABA receptors allowed LEC_GLU input to evoke CA3 PN action potential firing (Fig 2A). This could stem from relieving CA3 PN from LEC_GLU-driven feedforward inhibition and recruiting additional excitation in the hippocampal network. To test the latter, we examined CA3 PN EPSC input-output curves in response to LEC_GLU stimulation in baseline (ACSF), inhibition-blocked (SR & CGP), and monosynaptic-only (TTX & 4-AP) conditions (Fig 2B, Fig S4C–E). Resolving the number of peaks in the EPSC waveforms, we found that increasing LEC_GLU input strength yielded multipeak EPSCs which further extended with GABAergic transmission blocked (Fig 2B), indicating that inhibition indeed curtails LEC_GLU-driven excitation onto CA3. Such multipeak EPSCs could be evoked with optogenetic stimulation of LEC_GLU above 3 % (0.21 mW/mm²) LED intensity in regular ACSF, thus suggesting that strong LEC_GLU drive can recruit polysynaptic excitation from DG and/or CA3/2. Conversely, monosynaptic LEC_GLU input could be probed below 3 % LED power threshold in our experimental conditions (Fig 2B). We confirmed this by testing low and high LEC_GLU stimulation intensity while probing extracellular population spikes in DG and CA3, summation of repeated LEC_GLU activation, and occlusion of CA3 PN output in response to combined LEC_GLU and MF or RC input stimulation (Fig S4–5, Text S1).

Fig 2. Input-output transformation of LEC_GLU and LEC_GABA projections to CA3.

(A) Left, input-output curve of LEC_GLU-evoked CA3 PN PSP amplitudes. Middle, sample traces of LEC_GLU-evoked CA3 PN PSPs and spikes in ACSF (green) and SR & CGP (orange). Right, time-course of LEC_GLU-evoked CA3 PN spike probability (spike probability in SR & CGP: n = 11, one-sample T test versus hypothetical mean of zero, p = 0.024). (B) Left, sample traces of LEC_GLU-evoked CA3 PN EPSCs in ACSF (green), SR & CGP (orange), and TTX & 4-AP (grey). Right, input-output curves of the number of peaks in the LEC_GLU-evoked CA3 PN EPSC waveform in ACSF, SR & CGP, and TTX & 4-AP (number of EPSC peaks: n = 11, two-way ANOVA, treatment, p < 0.001, power, p < 0.001, treatment x power, p < 0.001; number of EPSC peaks against LED power: n = 11, one-sample T tests versus hypothetical mean of one: 1 %, 0.07 mW/mm², p < 0.001, 2 %, 0.14 mW/mm², p = 0.167, 3 %, 0.21 mW/mm², p = 0.008). (C) Left, sample traces of CA3 IN action potential firing evoked by LEC_GLU (green) or LEC_GLU+LEC_GABA stimulation (magenta). Middle, sample raster plot (trials were interleaved during recordings but grouped here for visualization). Right, spike probability of CA3 IN evoked by LEC_GLU with or without LEC_GABA stimulation (n = 5, two-way ANOVA, input combination, p < 0.001, pulse #, p < 0.001, input combination x pulse #, p = 0.982). (D) Sample traces (left) and amplitude (right) of LEC_GLU-evoked CA3 PN IPSCs with (magenta) or without (green) LEC_GABA stimulation (n = 11, paired-T test, p = 0.042). (E) Sample traces (left), PSP peak (middle) and spike probability (right) of CA3 PN dendritic responses to repeated stimulation of LEC_GLU (green) or LEC_GLU+LEC_GABA (magenta) inputs at 10 Hz (PSP peak: n = 29, two-way ANOVA, input combination, p = 0.578, pulse #, p = 0.893, input combination x pulse #, p > 0.999; spike probability: n = 29, two-way ANOVA, input combination, p = 0.670, pulse #, p = 0.081, input combination x pulse #, p = 0.947). (F) Same as (E) but with CA3 PN somatic responses (PSP peak: n = 23, two-way ANOVA, input combination, p = 0.120, pulse #, p = 0.759, input combination x pulse #, p > 0.999; spike probability: n = 23, two-way ANOVA, input combination, p = 0.318, pulse #, p = 0.439, input combination x pulse #, p = 0.439). Error bars represent SEM.

LEC_GABA decreases feedforward inhibition in CA3 PNs

Given inhibition gates CA3 PN output in response to LEC_GLU input (Fig 2A), and since LEC_GABA selectively targets CA3 INs (Fig 1C–I), we asked whether LEC_GABA modulates feedforward inhibition. Indeed, LEC_GABA is known to disinhibit PN dendritic excitability in CA1 (14). Thus, we monitored CA3 SLM IN spike output to LEC_GLU +/− LEC_GABA stimulation in cell-attached mode, and found that LEC_GABA decreased CA3 IN action potential firing to LEC_GLU input (Fig 2C). This suggests that LEC_GABA likely disinhibits CA3 PNs. We tested this by recording LEC_GLU-evoked IPSCs in CA3 PNs with or without coincident activation of LEC_GABA, and found that LEC_GABA indeed decreased LEC_GLU-driven feedforward inhibition onto CA3 PNs (Fig 2D). Of note, although LEC_GABA-connected CA3 SLM INs received strong RC but weak DG excitation (Fig S6A), both DG- and RC-driven inhibition were unaffected by LEC_GABA stimulation (Fig S6B–C). These results suggest that LEC_GABA may disinhibit CA3 PNs by selectively reducing LEC_GLU-driven feedforward inhibition. To test whether LEC_GABA-mediated disinhibition promotes CA3 PN compartment-specific excitation, we recorded CA3 PN somas or dendrites and repeatedly activated LEC_GLU monosynaptic input (2 % light intensity, 0.14 mW/mm²) either alone or combined with LEC_GABA (Fig 2E–F). Surprisingly, we found no significant difference in PSP amplitude and spike probability in response to LEC_GLU with or without LEC_GABA in both somas and dendrites (Fig 2E–F). This suggests that, although LEC_GABA dampens LEC_GLU-driven feedforward inhibition, the disinhibitory effect does not contribute to increasing CA3 PN post-synaptic responses to LEC_GLU monosynaptic excitation alone.

LEC_GABA selectively boosts CA3 PN somatic output through input-specific disinhibition

As LEC_GLU may require excitatory input summation combined with relief from inhibition to drive CA3 PN output, we asked whether LEC_GABA-mediated disinhibition would promote integration of multiple inputs to yield CA3 PN output. For this, we tested if coincident activation of LEC_GABA with LEC_GLU monosynaptic input (2 % light intensity, 0.14 mW/mm²) and combinations of DG or RC inputs would supralinearly increase PSPs and spike output in CA3 PN soma or dendrites (Fig 3A). We compared PSP peak and spike probability evoked by repeatedly stimulating LEC_GLU, DG, and RC either alone or combined with one another, with or without additional LEC_GABA activation. We found that combined stimulations of LEC_GABA with LEC_GLU and RC together yielded more action potential firing in CA3 PN somas than expected by the linear summation of LEC_GLU plus RC (Fig 3B). In contrast, pairings of LEC_GLU and DG with added stimulation of LEC_GABA did not show such supralinear increase in somatic action potential firing compared to their linear sum (Fig 3C). Because non-linearities in input integration can stem from dendritic spikes that help depolarizations from distal inputs such as LEC_GLU reach the soma, we assessed post-synaptic responses to LEC_GABA pairing with LEC_GLU + DG or LEC_GLU + RC with dendritic recordings. Surprisingly, although LEC_GLU and RC pairing sometimes elicited dendritic spikes, additional stimulation of LEC_GABA did not significantly change dendritic spike probability (Fig 3D). Consistent with our somatic data, LEC_GLU and DG inputs did not sum efficiently in CA3 PN dendrites thus virtually never yielding dendritic spikes, and LEC_GABA had no further effect (Fig 3E). This suggests that LEC_GABA selectively boosts LEC_GLU integration with RC to disinhibit CA3 PN somatic output, highlighting a pathway- and compartment-specific disinhibitory mechanism.

Fig 3. Input- and compartment-specific disinhibition of CA3 PN output by LEC_GABA.

(A) Dual-optogenetic activation of LEC_GLU and LEC_GABA combined with electrical stimulation of local RC (left) and DG (right) allows probing of CA3 output in response to specific input integration. (B) Sample traces (top), PSP peak and spike probability (bottom left) of RC- (black), RC + LEC_GLU- (blue), and RC + LEC_GLU + LEC_GABA-evoked (orange) CA3 PN somatic responses, and delta spike probability (bottom right) between observed RC + LEC_GLU with or without LEC_GABA stimulation versus expected from linear sum of these inputs (difference between observed vs expected spike probability: n = 18, two-way ANOVA, input combination, p = 0.007, pulse #, p = 0.841, input combination x pulse #, p = 0.549). (C) Same as (B) with DG stimulation (brown) instead of RC (difference between observed vs expected spike probability: n = 18, two-way ANOVA, input combination, p = 0.115, pulse #, p = 0.993, input combination x pulse #, p = 0.258). (D) same as (B) but in CA3 PN dendrites (difference between observed vs expected spike probability, n = 17, two-way ANOVA, input combination, p = 0.173, pulse #, p = 0.111, input combination x pulse #, p = 0.925). (E) same as (C) but in CA3 PN dendrites (difference between observed vs expected spike probability, n = 17, two-way ANOVA, input combination, p = 0.319, pulse #, p = 0.409, input combination x pulse #, p = 0.409). Error bars represent SEM.

LEC_GLU and LEC_GABA inputs to CA3 contribute to goal-oriented learning

Since LEC_GLU and LEC_GABA synergistically drove CA3 activity ex vivo, we asked whether LEC inputs would have complementary roles in shaping behavioral learning and hippocampal representations in vivo. For this, we examined the contributions of LEC_GLU and LEC_GABA inputs to CA3 at the behavioral and neural coding levels during an operant goal-oriented learning (GOL) paradigm, where mice learned to associate distinct reward locations with contextually different environments. Mice were first trained to run head-fixed for water rewards on multi-textured belts (Fig 4A). Once habituated to randomly forage (Fig 4B), mice were subjected to the GOL task (57, 58), which required navigating to specific reward zones in three different contexts (A, A’, and B) defined by combinations of sensory cues with graded similarity. Environments A and A’ had high similarity as they retained the same tactile cues but differed in their olfactory and auditory cues. In contrast, A and B had low similarity as they differed in all tactile, olfactory, and auditory cues. Thus, mice needed to associate spatial and non-spatial cues from each environment to find the reward zone in each context. A two-sided Möbius belt design with distinct textures on each side allowed rapid switching between tactile cues without dismounting the mice (59, 60) to capture neural dynamics starting from the first laps in every environment.

Fig 4. Contributions of CA3-projecting LEC_GLU and LEC_GABA to goal-oriented learning.

(A) AAV injection of Syn-driven Cre-recombinase with retrograde tropism into CA3 and CaMKII-driven-flexed and Dlx-driven-flexed chemogenetic receptors into LEC of C57BL/6J mice allows expression of hM4D(Gi)-mCitrine in LEC_GLU and PSAM in LEC_GABA, enabling input-specific dual-chemogenetic perturbation of CA3-projecting LEC neurons in vivo. Further AAV injection of CaMKII-driven GCaMP6f into CA3 allows two-photon calcium imaging of CA3 PNs through a cranial window on head-fixed mice running for water rewards on multi-textured belts in different paradigms. (B) Fraction of licks spatial distribution in random foraging (RF) for control (black, n = 6), ΔLEC_GLU (green, n = 10) and ΔLEC_GABA (red, n = 9) mice injected with saline. (C) Sample injection site of AAVs targeting CA3-projecting LEC_GLU (hM4D(Gi)-mCitrine, green, left) and LEC_GABA (α-bungarotoxin-stained PSAM, red, right) neurons, with DAPI staining (blue). (D) Fraction of licks spatial distribution in GOL environments A (left), A’ (middle) and B (right) on day 1 (top), day 5 (middle), and expanded views of reward zones (bottom) for control (black, n = 6), ΔLEC_GLU (green, n = 10) and ΔLEC_GABA (red, n = 9) mice. (E) Fraction of licks in reward zones across days with ligand injections (control, n = 6, ΔLEC_GLU, n = 10, ΔLEC_GABA, n = 9, two-way ANOVA, treatment, p < 0.001, days, p < 0.001, treatment x days, p = 0.795, Tukey post hoc test, control vs ΔLEC_GLU, p < 0.001, control vs ΔLEC_GABA, p < 0.001, ΔLEC_GLU vs ΔLEC_GABA, p = 0.014). (F) Fraction of licks in reward zones across days with saline injections (control, n = 6, ΔLEC_GLU, n = 10, ΔLEC_GABA, n = 9, two-way ANOVA, treatment, p = 0.873, days, p < 0.001, treatment x days, p = 0.996). (G) Fraction of licks in reward zone with or without silencing (control, n = 6, ΔLEC_GLU, n = 10, ΔLEC_GABA, n = 9, one-way ANOVA, p < 0.001, Tukey post hoc test, Table 1). Error bars represent SEM.

Of note, LEC should be engaged in this task since it conveys multimodal sensory information to the hippocampus, especially odors (25, 26), tactile cues (35) and reward goals (14, 27, 61, 62). To test the role of LEC inputs to CA3 in this paradigm, we bilaterally injected a retrograde Cre virus into CA3 and Cre-dependent inhibitory chemogenetic constructs into LEC of C57BL/6J mice (Fig 4A, 4C). Most experiments targeted CA3-projecting LEC inputs with dual-chemogenetics by expressing CaMKII-driven-flexed-hM4D(Gi)-mCitrine (63) in LEC_GLU and Dlx-driven-flexed-PSAM (64) in LEC_GABA of the same animals. A separate subset of animals had silencing of either LEC_GLU (as above) or LEC_GABA with Dlx-driven-flexed-hM4D(Gi)-mCherry. Mice were injected intraperitoneally (ip) daily with the cognate ligands (CNO or PSEM) prior to running in environments A, A’, and B for 10 min sessions each. As a control, mice lacking expression of the chemogenetic receptors received randomized ip injections of CNO or PSEM (Fig S7A–B). These control mice readily increased their licking in the reward zones and had their performance improve daily (Fig 4D–E, S7C, S7E). In contrast, chemogenetic silencing of either LEC_GLU (ΔLEC_GLU) or LEC_GABA (ΔLEC_GABA) decreased performance throughout the task (Fig 4D–G, S7C, S7E). Such impairments indicate that both LEC_GLU and LEC_GABA inputs to CA3 contribute to GOL, although we note that CA3-projecting LEC inputs may collaterize to DG and that local circuit interactions in LEC enable crosstalk between LEC_GLU and LEC_GABA (Fig S8). Finally, silencing of LEC_GLU or LEC_GABA after mice had learnt the task had no effect on recall performance (Fig 4F–G, S7D, S7F), suggesting that LEC inputs are required for learning but not during memory recall once the task is successfully learnt.

LEC_GLU and LEC_GABA support context-dependent remapping of CA3 ensembles

To resolve the neuronal dynamics underlying LEC_GLU and LEC_GABA contributions to GOL, we virally expressed CaMKII-driven-GCaMP6f or GCaMP8f in CA3 PN to monitor their somatic and dendritic activity longitudinally with two-photon calcium imaging through a cranial window during GOL (Fig 4A, 5A–B, S9). We identified spatially-tuned CA3 PN somas and dendrites in environment A and compared their remapping in environments A’ and B using population vector (PV) and tuning curve (TC) correlations (Fig 5C–H). Control mice showed lower PV and TC correlations between A and B than A and A’ environments (Fig 5C, 5F–H), indicating more remapping from A to B than A to A’ which is consistent with B being more different from A than A’. In contrast, ΔLEC_GLU mice showed substantial remapping between A and A’ (Fig 5D, 5F–H), thereby attenuating the differences in spatial correlations from similar (A-A’) to different (A-B) environments. This increased remapping between the similar A and A’ environments was even more pronounced in ΔLEC_GABA mice (Fig 5E–H), reaching similar levels as between A and B. These observations indicate an instability of spatial representations with silencing of LEC_GLU and LEC_GABA. This was not due to basic differences in activity (Fig S10–11) or decreased encoding of spatial features, as the fraction of spatially tuned ROIs or the width of their place fields in ΔLEC_GLU and ΔLEC_GABA mice were not different from control (Fig S12). Rather, correlations of spatially tuned ROIs across similar environments were significantly lower in ΔLEC_GLU and ΔLEC_GABA mice than control (Fig 5F–H, S13). This suggests that both LEC_GLU and LEC_GABA inputs support context-dependent spatial coding by distinct stable CA3 ensembles. Imaging both soma and dendrites simultaneously at the population level allowed us to link the effects of LEC input silencing on behaviorally-driven in vivo neural dynamics with the ex vivo circuit mapping results showing compartment-specific disinhibition. Of note, analysis of dendritic tuning revealed higher levels of remapping compared to somas in control but not ΔLEC_GLU or ΔLEC_GABA conditions (Fig 5F–H, S13). This stronger effect of LEC input silencing on somas than dendrite may be linked to the preferential boosting of CA3 PN somatic activity in response to LEC_GLU + LEC_GABA + RC inputs (Fig 3B) through perisomatic disinhibition from local CA3 basket cells (Fig 1I, 2C).

Fig 5. Roles of LEC_GLU and LEC_GABA in CA3 context-dependent remapping.

(A) Sample CA3 field of view. (B) Sample ΔF/F traces from three ROIs (a, b, and c) in control (black), ΔLEC_GLU (green), ΔLEC_GABA (red) conditions. (C) Left, sample control normalized rate maps of spatially tuned CA3 somas in the A (top), A’ (middle) and B (bottom) environments sorted according to their place field location in the A session (darker colors indicate higher firing rates). Right, sample control spatial correlations reported as cumulative distribution of tuning curve (TC) correlation coefficients between A-A’ (dark), A’-B (medium) and A-B (light) sessions (top), and population vector (PV) correlation matrix of A-A’ (middle) and A-B (bottom) sessions (warmer colors indicate higher correlations within a −0.3:0.7 range). (D) Same as (C) but with ΔLEC_GLU. (E) Same as (C) but with ΔLEC_GABA. (F) Somatic (left) and dendritic (right) A-A’ PV correlation coefficients across days (control, n = 6, ΔLEC_GLU, n = 9, ΔLEC_GABA, n = 8, two-way ANOVAs, treatment comparisons on somas: treatment, p < 0.001, days, p = 0.776, treatment x days, p = 0.992, Tukey post hoc test, control vs ΔLEC_GLU, p < 0.001, control vs ΔLEC_GABA, p < 0.001, ΔLEC_GLU vs ΔLEC_GABA, p = 0.194; treatment comparisons on dendrites: treatment, p < 0.001, days, p = 0.907, treatment x days, p = 0.980, Tukey post hoc test, control vs ΔLEC_GLU, p = 0.002, control vs ΔLEC_GABA, p = 0.002, ΔLEC_GLU vs ΔLEC_GABA, p = 0.976; compartment comparisons on control: compartment, p < 0.001, days, p = 0.814, compartment x days, p = 0.992; compartment comparisons on ΔLEC_GLU: compartment, p = 0.186, days, p = 0.979, compartment x days, p = 0.815; compartment comparisons on ΔLEC_GABA: compartment, p = 0.848, days, p = 0.968, compartment x days, p = 0.874). (G) Somatic (left) and dendritic (right) A’-B PV correlation coefficients across days (control, n = 6, ΔLEC_GLU, n = 9, ΔLEC_GABA, n = 8, two-way ANOVAs, treatment comparisons on somas: treatment, p < 0.001, days, p = 0.609, treatment x days, p = 0.695, Tukey post hoc test, control vs ΔLEC_GLU, p < 0.001, control vs ΔLEC_GABA, p < 0.001, ΔLEC_GLU vs ΔLEC_GABA, p = 0.908; treatment comparisons on dendrites: treatment, p = 0.034, days, p = 0.562, treatment x days, p = 0.943, Tukey post hoc test, control vs ΔLEC_GLU, p = 0.103, control vs ΔLEC_GABA, p = 0.040, ΔLEC_GLU vs ΔLEC_GABA, p = 0.933; compartment comparisons on control: compartment, p < 0.001, days, p = 0.248, compartment x days, p = 0.983; compartment comparisons on ΔLEC_GLU: compartment, p = 0.939, days, p = 0.512, compartment x days, p = 0.688; compartment comparisons on ΔLEC_GABA: compartment, p = 0.979, days, p = 0.660, compartment x days, p = 0.863). (H) Somatic (left) and dendritic (right) A-B PV correlation coefficients across days (control, n = 6, ΔLEC_GLU, n = 9, ΔLEC_GABA, n = 8, two-way ANOVAs, treatment comparisons on somas: treatment, p < 0.001, days, p = 0.285, treatment x days, p = 0.838, Tukey post hoc test, control vs ΔLEC_GLU, p < 0.001, control vs ΔLEC_GABA, p = 0.002, ΔLEC_GLU vs ΔLEC_GABA, p = 0.756; treatment comparisons on dendrites: treatment, p = 0.572, days, p = 0.813, treatment x days, p = 0.825; compartment comparisons on control: compartment, p < 0.001, days, p = 0.296, compartment x days, p = 0.711; compartment comparisons on ΔLEC_GLU: compartment, p = 0.764, days, p = 0.530, compartment x days, p = 0.686; compartment comparisons on ΔLEC_GABA: compartment, p = 0.560, days, p = 0.736, compartment x days, p = 0.503). (F-H) Across session-pairs statistical comparisons: control, n = 6, two-way ANOVAs, PV somas: sessions, p < 0.001, days, p = 0.270, sessions x days, p = 0.988, Tukey post hoc test, A-A’ vs A’-B, p < 0.001, A-A’ vs A-B, p < 0.001, A-B vs A’-B, p = 0.639; PV dendrites: sessions, p < 0.001, days, p = 0.586, sessions x days, p = 0.950, Tukey post hoc test, A-A’ vs A’-B, p < 0.001, A-A’ vs A-B, p < 0.001, A-B vs A’-B, p = 0.306; ΔLEC_GLU, n = 9, two-way ANOVAs, PV somas: sessions, p = 0.012, days, p = 0.540, sessions x days, p = 0.862, Tukey post hoc test, A-A’ vs A’-B, p = 0.056, A-A’ vs A-B, p = 0.015, A-B vs A’-B, p = 0.785; PV dendrites: sessions, p = 0.538, days, p = 0.806, sessions x days, p = 0.899; ΔLEC_GABA, n = 8, two-way ANOVAs, PV somas: sessions, p = 0.462, days, p = 0.429, sessions x days, p = 0.961; PV dendrites: sessions, p = 0.300, days, p = 0.877, sessions x days, p = 0.929. Error bars represent SEM.

LEC_GLU and LEC_GABA drive stability of CA3 place maps associated with learning

Next, we examined the neural correlates of behavioral performance in the GOL task, and how they were influenced by LEC_GLU and LEC_GABA. Tracking place cell ensembles across days revealed a progressive increase in their correlation with day 5 over the course of learning in control mice (Fig 6A). This suggests that neural representations relevant to spatial navigation in the GOL task stabilize with learning. In contrast, cross-day correlations remained low in the ΔLEC_GLU and ΔLEC_GABA mice (Fig 6B–C), indicating that LEC inputs contribute to shaping CA3 spatial coding during learning. Specifically, somatic PV correlations between days 1–4 and day 5 in control were significantly higher than in ΔLEC_GABA, and trended to in ΔLEC_GLU (Fig 6D, S14, S15). Pairwise day-day comparisons also showed increased somatic PV correlations in control compared to ΔLEC_GLU and ΔLEC_GABA (Fig 6E, S14, S15). Restricting such analysis to place cells using TC correlations showed decreased differences between groups (Fig S14, S15), perhaps because of conserved tuning quality of individual cells (Fig S12). Altogether, these data implicate both LEC_GLU and LEC_GABA in stabilizing CA3 spatial representations associated with learning. Again, dendritic spatial ensembles were less stable than somatic ones in control but not ΔLEC_GLU or ΔLEC_GABA conditions (Fig 6D–E, S14, S15). Hence, it is possible that LEC inputs promote stability in CA3 in vivo by facilitating its recurrent activity through somatic disinhibition as seen in the LEC_GLU + LEC_GABA + RC recurrent condition ex vivo. Finally, we tested whether the decreased stability of CA3 somatic place cell ensembles observed with LEC manipulations would affect the precision of spatial coding. For this, we used the activity of the CA3 PN soma to decode position across sessions in each given context during learning. We found that silencing of either LEC_GLU or LEC_GABA impaired position decoding compared to control (Fig 7A, S16), consistent with reduced performance in GOL.

Fig 6. Involvement of LEC_GLU and LEC_GABA in learning-associated CA3 spatial coding.

(A) Left, sample control normalized rate maps of spatially tuned CA3 somas in the A’ environment across days 1–5 sorted according to their place field location from day 5 (darker colors indicate higher firing rates). Right, sample control spatial correlations across days 1–4 against day 5 reported as PV correlation matrix (warmer colors indicate higher correlations within a −0.3:0.7 range) and cumulative distribution of TC correlation coefficients (darker colors indicate increasing days relative to day 5). (B) Same as (A) but with ΔLEC_GLU. (C) Same as (A) but with ΔLEC_GABA. (D) Somatic (left) and dendritic (right) PV correlation coefficients across days relative to day 5 (control, n = 6, ΔLEC_GLU, n = 9, ΔLEC_GABA, n = 8, two-way ANOVAs, treatment comparisons on somas: treatment, p = 0.007, days, p = 0.255, treatment x days, p = 0.159, Tukey post hoc test, control vs ΔLEC_GLU, p = 0.067, control vs ΔLEC_GABA, p = 0.005, ΔLEC_GLU vs ΔLEC_GABA, p = 0.517; treatment comparisons on dendrites: treatment, p = 0.562, days, p = 0.055, treatment x days, p = 0.928; compartment comparisons on control: compartment, p = 0.002, days, p < 0.001, compartment x days, p = 0.766; compartment comparisons on ΔLEC_GLU: compartment, p = 0.496, days, p = 0.354, compartment x days, p = 0.890; compartment comparisons on ΔLEC_GABA: compartment, p = 0.277, days, p = 0.964, compartment x days, p = 0.251). (E) Somatic (left) and dendritic (right) PV correlation coefficients across days between pairs of days (control, n = 6, ΔLEC_GLU, n = 9, ΔLEC_GABA, n = 8, two-way ANOVAs, treatment comparisons on somas: treatment, p < 0.001, days, p = 0.801, treatment x days, p = 0.895, Tukey post hoc test, control vs ΔLEC_GLU, p = 0.0125, control vs ΔLEC_GABA, p < 0.001, ΔLEC_GLU vs ΔLEC_GABA, p = 0.075; treatment comparisons on dendrites: treatment, p = 0.050, days, p = 0.085, treatment x days, p = 0.833, Tukey post hoc test, control vs ΔLEC_GLU, p = 0.047, control vs ΔLEC_GABA, p = 0.734, ΔLEC_GLU vs ΔLEC_GABA, p = 0.221; compartment comparisons on control: compartment, p < 0.001, days, p = 0.048, compartment x days, p = 0.995; compartment comparisons on ΔLEC_GLU: compartment, p = 0.071, days, p = 0.424, compartment x days, p = 0.676; compartment comparisons on ΔLEC_GABA: compartment, p = 0.112, days, p = 0.784, compartment x days, p = 0.462). Error bars represent SEM.

Fig 7. Modeling LEC_GLU and LEC_GABA effects on CA3 ensemble stability during learning.

(A) Behavioral performance (fraction of correct licks) as a function of neural ensemble stability (PV correlation coefficient) and spatial information (decoding error) in control (black, left), ΔLEC_GLU (green, middle; control average shown in gray for reference), and ΔLEC_GABA (red, right; control average shown in gray for reference) conditions. (B) Ensemble formation and reactivation is modelled in a CA3 neuronal network with recurrent excitation (RC), and lateral feedback inhibition (FBI). LEC_GLU provides excitation to a subset of neurons, and LEC_GABA drives disinhibition by reducing feedforward inhibition (FFI). (C) Recall performance of the modelled neural ensemble. (D) Variability of activity in the modelled neuronal network. Error bars represent standard deviation.

Lastly, to test how instability of neuronal ensembles observed with ΔLEC_GLU and ΔLEC_GABA could underlie learning impairments, we simulated the formation of a neuronal assembly in a recurrent CA3 network (Fig 7B). The model consisted of excitatory neurons with recurrent connections plastic under a simple Hebbian learning rule. These neurons also received lateral inhibition as well as constant inputs from LEC_GLU and LEC_GABA. Since our optogenetic mapping work shows that LEC_GLU drives both excitation and feedforward inhibition, we assumed that LEC_GLU provides input to a subset of neurons that will form an assembly during the learning phase. Also based on our ex vivo findings that LEC_GABA drives disinhibition of CA3 recurrent activity, we modelled LEC_GABA as a general disinhibitory input to all neurons. At the end of the learning phase, we found that the recurrent weights within the assembly were strengthened. To test the learning performance, we simulated a recall protocol where we stimulated 70% of the assembly and recorded the activity of the remaining 30% of the assembly. In control conditions, we found that the recall performance was high. To mimic the experimental manipulations, we perturbed LEC_GLU or LEC_GABA inputs either during or after learning. For LEC_GLU, we did so by degrading the input to the neuronal subset forming the assembly, assuming that disrupted sensory information with ΔLEC_GLU would dilute the recruitment of specific postsynaptic CA3 PN ensembles. And for LEC_GABA, we modelled ΔLEC_GABA by reducing the constant disinhibitory input to the neuronal network. We found that both ΔLEC_GLU and ΔLEC_GABA perturbations impaired performance if applied during learning but not recall (Fig 7C). This was because, after learning, the strong recurrent weights within CA3 could promote pattern completion on their own. We note that, while both perturbations during learning similarly impaired recall performance, the variability of the representation was increased with ΔLEC_GLU but not ΔLEC_GABA (Fig 7D). This may reflect the differences in mechanisms modelled after our functional circuit mapping: ΔLEC_GABA was abstracted as a global reduction of activity in the recurrent network, thus making the threshold for potentiating connections harder to reach which resulted in a quantitatively weakened but qualitatively preserved assembly. In contrast, the net amounts of excitation and inhibition were conserved with ΔLEC_GLU. Instead, the distribution of neurons receiving input was broadened, yielding less specificity to the composition of the assembly.

Discussion

Flexibility and stability are critical features of neural operations supporting memory formation and recall. Despite being observed and modeled across species and in many brain regions, including the hippocampus (65, 66), cortex (67, 68) and subcortical areas (69, 70), little is known of the circuit mechanisms that influence flexibility vs stability of neuronal ensembles. Our study examined the specific roles of long-range glutamatergic and GABAergic circuit interactions between the entorhinal cortex and hippocampal area CA3 to resolve how they orchestrate activity supporting neuronal representation dynamics.

To capture flexibility and stability at the population level with cellular compartment-specific resolution, we leveraged an in vivo imaging preparation to monitor activity dynamics in CA3 PN dendrites and soma simultaneously during behavioral with targeted circuit manipulations. We mechanistically bolstered these observations by assessing fine-scale excitation-inhibition dynamics and input integration at the single neuron level with dendritic and somatic intracellular electrophysiology recordings paired with dual color optogenetics ex vivo. We found that LEC_GLU recruited a significant amount of feedforward inhibition, preventing CA3 PN spike output. In contrast, the LEC_GABA input targets a subset of CA3 INs, thereby suppressing the LEC_GLU-driven perisomatic feedforward inhibition. LEC_GABA acts as a pathway- and compartment-specific disinhibitory gate that selectively boosts the integration of LEC_GLU and RC inputs to yield CA3 PN somatic output. We found that silencing of either LEC_GLU or LEC_GABA inputs impaired GOL performance during encoding but not recall. This was accompanied by a destabilization of CA3 spatial representations, which now remapped even within same day sessions despite context similarity. This destabilization was also seen across days within a given context where LEC_GLU or LEC_GABA silencing induced greater remapping than controls. These activity changes were correlated with a drop in learning rate and decoding accuracy. Thus, using multidisciplinary approaches, we found that LEC_GLU and LEC_GABA inputs synergistically support stability in CA3 by recruiting pathway- and compartment-specific feedforward inhibition and disinhibition to promote recurrent circuit activity. This functionally endows CA3 with graded stabilization of its internal representations supporting learning of spatial-contextual associations.

The disinhibitory configuration of long-range GABAergic projections is emerging as a conserved circuit motif across the brain (71). We (14) and others (15) showed that in CA1, LEC and MEC long-range GABAergic projections target local interneurons in CA1, and seem to avoid PN soma. Our dendritic recordings directly show that LEC_GABA exclusively targets local GABAergic IN but not PN soma or dendrites, validating its disinhibitory nature in CA3. Our study highlights how disinhibitory inputs achieve functional versatility through their target circuit organization and spatio-temporal integration with other convergent inputs. In CA1, LEC_GLU recruits inhibitory CCK+ and disinhibitory VIP+ INs to selectively increase dendritic excitability (56), whereas LEC_GABA suppresses dendrite-targeting SR/SLM CCK+ INs that control timing-dependent dendritic spikes and associated CA3-CA1 heterosynaptic plasticity without affecting somatic output in CA1 PNs (14). In contrast, we found in CA3 that LEC_GABA input targets SLM VIP+ and CCK+ INs (Fig 1I, S3), which putatively include the VIP+/CCK+ soma-targeting basket cells (54, 72, 73). Thus, LEC_GABA predominantly mediates disinhibition of the CA3 PN soma rather than their dendrites. Consequently, our somatic vs dendritic recordings during co-activation of the LEC_GLU and RC inputs with the perisomatic compartment-specific disinhibition from LEC_GABA facilitate spike output in the soma but not dendrites of CA3 PNs (Fig 3B, 3D). This is compatible with our in vivo two-photon somatic vs dendritic imaging observations that baseline remapping differs between soma and dendrites, and that silencing LEC inputs shows a stronger effect on context-driven remapping (Fig 5F–H) and learning-driven stabilization of somatic spatial activity (Fig 6D–E). While propagated dendritic spikes and back propagating action potentials are bound to influence calcium and fast somato-dendritic coupling dynamics in vivo (59), which GCaMP cannot resolve, we could still capture compartment-specific differences across context-switches and longer time scales of learning.

Beyond resolving compartment-specific effects of LEC_GABA, our dual color optogenetics approach coupled with electrical activation of the local hippocampal inputs enabled us to probe non-linear integration of conjunctive inputs in a pathway-specific manner. The effect of the LEC_GABA disinhibition is also pathway specific, preferentially boosting LEC_GLU integrated with RC output over DG. Disinhibitory routing of LEC_GLU input through the CA3 recurrent network substantially amplifies its output, making the influence of the otherwise sparse LEC_GABA input powerful. In addition, strong activation of LEC_GLU may recruit the DG feedforward and CA3 feedback local circuits to allow for an activity- or state-dependent gain modulation of circuit output. For example, weak cortical activity may elicit local sodium spikes in distal dendrites (74), as we observe in our ex vivo dendritic recordings (~250 μm from soma). On the other hand, stronger cortical activity could engage conjunctive DG- and RC-triggered NMDA dendritic spikes (75, 76), which may propagate and elicit somatic spike. Conversely, coincident back propagating action potentials (74, 77) may depolarize dendrites, thus breaking compartmental barriers and globalizing signals.

Our finding that LEC_GLU and LEC_GABA inputs integrate in CA3 to boost RC activity predicts a role of LEC in the stabilization of hippocampal representations. Spatial representations formed and stored within the auto-associative CA3 subregion provided a rich substrate for testing this idea. CA3 place cells can remap in novel or morphed environments, as well as reactivate to complete representations from partial inputs, thus performing both pattern separation and completion computations (32, 33, 35, 40–44, 47, 48, 65, 66, 78–82). LEC neurons, rodents and humans alike, show weak spatial selectivity but strong context-dependent responses to a range of sensory and behaviorally salient stimuli including local cues, odors, objects and reward goals as well as learning rules or timing sequences (14, 20–27). However, the tuning of LEC neurons and their CA1 projections to navigational goals (62), their relevance to GOL behavior in CA1 (61) as well as rate remapping in CA3 (31), and our present findings of stabilizing CA3 place cell ensembles suggest a more direct role in hippocampal spatial coding functions.

The local and long-range circuit interactions we decipher in our study provide a critical functional link between CA3 neural dynamics and behavioral performance. Indeed, our in vivo imaging with GOL behavior shows that CA3-targeting LEC_GLU and LEC_GABA inputs similarly support graded representations of environments with different degrees of dissimilarity and their stabilization as a function of learning. The similar effect size seen with silencing LEC_GLU and LEC_GABA confirms the unidirectional action of both circuits in driving specific context-laden excitatory input (LEC_GLU) boosted by disinhibition (LEC_GABA) and amplified by the CA3 RC attractor network. It also raises the possibility of cross-talk at the level of LEC where glutamatergic projection neurons locally co-excite their GABAergic counterparts.

The role of long-range GABAergic projections goes well beyond elevating coincident excitation through disinhibition. By targeting local inhibitory neurons, which are particularly effective at pacing network oscillations (83), long-range GABAergic projections can have widespread effects to help synchronize distant brain areas (16), across several oscillation patterns (15, 18, 84–89). Long-range GABAergic neurons strongly respond to salient sensory stimuli regardless of the modality (14, 90), suggesting a role in encoding coincident multimodal inputs and novelty detection. Indeed, during encoding, silencing of LEC_GABA projections to CA1 (14) impairs learning of fear contexts and novel objects while silencing the CA3-projecting LEC_GABA neurons impairs goal-oriented spatio-contextual learning. This warrants future investigation into brain state-, neuromodulation- and task demand-dependent specific contributions of LEC_GLU vs LEC_GABA to shaping CA3 synaptic plasticity (91, 92) and non-linearities relevant for learning and memory recall. Overall, our study provides a sub-compartment and pathway specific circuit mechanism by which that LEC_GLU and LEC_GABA act together to mediate a tightly-controlled dialogue between excitation, inhibition and disinhibition, to regulate hippocampal place cell stability in service of context-dependent learning and memory formation.

Materials and Methods

Animals:

All experiments were conducted in accordance with the National Institute of Health guidelines and with the approval of the New York University School of Medicine Institutional Animal Care and Use Committee (IACUC). Mice were obtained from Jackson Laboratory and subsequent breeding was established in-house. Acute slice electrophysiology experiments used GAD2-Cre, Pvalb-Cre, SST-Cre, CCK-Cre, VIP-Cre, Ai9 and Ai14 transgenic mice from both sexes, 8 weeks- to 6 months-old, with a C57BL/6J genetic background. In vivo 2-photon calcium imaging experiments used C57BL/6J mice from both sexes, 3 to 6 months-old.

Virus preparation:

An adeno-associated virus encoding an inhibitory neuron promoter h56D (93), channelrhodopsin2 (H134R)-sfGFP, woodchuck hepatitis virus posttranscriptional regulatory element (WPRE) and SV40 polyadenylation signal was assembled using a modified helper-free system (Stratagene) as a serotype 2/7 (rep/cap genes) AAV and harvested and purified over sequential cesium chloride gradients as previously described (94). The codon-optimized channerhodopsin2 fusion protein included an EAGAVSGGVY linker between the protein domains and C-terminal Golgi and endoplasmic reticulum export signals (95, 96) to aid membrane expression.

Stereotaxic surgery:

Animals (4+ weeks-old) were anaesthetized with isofluorane (1.5–3 %, inhaled) and buprenorphine (0.1 mg/kg, injected intra-peritoneally). Mice were placed in a stereotaxic apparatus (Stoelting), a small incision (~0.5 cm length) was made in the skin to expose the skull, and the skull was levelled flat according to bregma and lambda. A small craniotomy (~0.5 mm diameter) was drilled in the skull above the injection sites and 276–414 nL of virus was injected (Drummond Scientific Nanoject II) into the brain at the following coordinates: anterior–posterior relative to bregma: 3.2 and 3.0 mm for LEC, 1.9 mm for CA3; medial-lateral relative to midline: 4.5 and 4.7 mm for LEC, 2.2 mm for CA3; dorsal-ventral relative to the surface of the brain: 2.5, 2.8 and 3.0 mm for LEC, 1.8, 2.0 and 2.2 mm for CA3. The injection pipette was slowly lowered into the brain up to 0.2 mm deeper than the deepest targeted coordinate, left at this location for 1 min, retracted to the actual coordinate and held there for 30 s prior to injection. 46–69 nL of virus was injected in 23 nL increments spaced by 15 s at each z coordinate, with an additional 2 min pause between z coordinates and a final 10 min incubation at the last (shallowest) z coordinate of each injection site before slowly retracting the injection pipette out of the brain. Mice were sutured, given neosporin topically on the incision site, injected with 1 mL of sterile saline sub-cutaneously, monitored until full recovery from anesthesia was observed, and given analgesic post-operative care (buprenorphine, 0.1 mg/kg, injected intra-peritoneally) for 3 days. Infection sites were confirmed to be specific post hoc for all experiments. The adeno-associated virus (AAV) AAV2/5:EF1α-DIO-hChR2(H134R)-eYFP-WPRE-HGHpA (Addgene #20298) was used at 7.7×10¹²-1×10¹³ vg/mL, AAV2/5:CaMKIIa-hChR2(H134R)-eYFP (Addgene #26969) was used at 1.5×10¹³ vg/mL, AAV2/5:CaMKII-Chronos-GFP (Neurophotonics #319) was used at 1.3×10¹³ vg/mL, AAV2/5:Syn-flex-ChrimsonR-tdTomato (Addgene #62723) was used at 4.74×10¹²-2×10¹³ vg/mL, AAV2/7:h56D-ChR2-sfGFP (kindly provided by Boris Zemelman) was used at 2×10¹³ vg/mL, AAV2/1:CaMKII-GCaMP6f-WPRE-SV40 (Addgene #100834) was used at 1×10¹³ vg/mL, AAV2/1:CaMKII-GCaMP8f-WPRE (Addgene #176750) was used at 1.7×10¹³ vg/mL, AAV2/5:CaMKII-DIO-hM4D(Gi)-IRES-mCitrine (Neurophotonics #810) was used at 1.3×10¹³ vg/mL, AAV2/5:hDlx-flex-PSAM (custom construct from Neurophotonics) was used at 6.8×10¹² vg/mL, AAV2/5:Dlx-DIO-hM4D(Gi)-mCherry (construct from Yingxi Lin laboratory, packaged by Boston Children’s Harvard Viral Core) was used at 5.04×10¹³ vg/mL, AAV2/9:Dlx-flex-TdTomato (Addgene #83894) was used at 7×10¹² vg/mL, AAVretro:Syn-Cre (Neurophotonics #1299) was used at 1.4×10¹³ vg/mL, AAV2/5:Dlx-Flex-ChR2-mCherry (custom construct from Neurophotonics) was used at 6.8×10¹² vg/mL, AAV2/9:CaMKII-DIO-Chronos-eGFP (Neurophotonics #1830) was used at 1×10¹³ vg/mL.

Electrophysiology:

Artificial cerebrospinal fluid (ACSF) and protective dissection ACSF (dACSF) were oxygenated with a 95 % O₂ and 5 % CO₂ mixture at all times. Mice were deeply anaesthetized with isoflurane (5 % for 5 min, inhaled) and perfused transcardially with ~20 mL of ice-cold NMDG-based dACSF (97) containing (in mM): NMDG 93, KCl 2.5, NaH₂PO₄ 1.25, NaHCO₃ 30, HEPES 20, glucose 25, thiourea 2, Na-ascorbate 5, Na-pyruvate 3, CaCl₂ 0.5, MgCl₂ 10. Brains were then rapidly removed, cortico-hippocampal complexes were dissected out and placed upright into a custom-made agar mold in ice-cold dACSF. 400 μm thick transverse slices were cut (vibratome, Leica VT1200S) at low speed (0.04 mm/s) and blade vibration amplitude (0.5 mm) in ice-cold dACSF. Slices were transferred to an immersed-type holding chamber and maintained in ACSF containing the following (in mM): NaCl 125, KCl 2.5, NaH₂PO₄ 1.25, NaHCO₃ 25, glucose 22.5, Na-ascorbate 1, Na-pyruvate 3, CaCl₂ 2, MgCl₂ 1. Slices were incubated at 32 °C for ~20 minutes and then maintained at room temperature for at least 30 minutes prior to recordings. Individual slices were transferred to a recording chamber perfused with ACSF at 3–5 mL/min (peristaltic pump, Watson Marlow) at 30 °C (in-line heater, Warner Instruments TC-324B). Tissue was visualized under an upright microscope (Zeiss Examiner A1 or Olympus BX51WI) equipped with DIC or Dodt gradient contrast at 5x-40x magnification with additional zoom optics 1–2.5x, and captured by a video camera (Hamamatsu ORCA-spark or ORCA-flash4.0). The headstages connected to the recording electrodes were mounted on motorized micromanipulators (Luigs & Neumann GmbH). Patch-clamp recordings were performed with potassium- or cesium-based intracellular solution containing the following (in mM): K- or Cs-methyl sulfonate 135, KCl 5, EGTA-KOH 0.1, HEPES 10, NaCl 2, MgATP 5, Na2GTP 0.4, Na2-phosphocreatine 10, and either Alexa 594 (50–100 μM) or biocytin (4 mg/mL). GΩ seal were formed and whole-cell recordings were obtained from CA3 pyramidal neurons soma (blind patch) or dendrites (blind patch) and interneurons (visual patch). Somatic patch pipette resistances were 2–5 MΩ, series resistances were 8–20 MΩ. Dendritic patch pipette resistances were 13–17 MΩ, series resistances were 40–60 MΩ. Series resistance was compensated 0–50 % to amount to an 8–10 MΩ actual resistance in voltage-clamp (somatic recordings). Bridge-balance was applied in current-clamp. Unless stated otherwise, the membrane potential was held at −70 mV in current-clamp. The liquid junction potential was < 10 mV and not corrected for. ChR2-, Chronos- or ChrimsonR-expressing LEC inputs were stimulated optically with 1–10 ms-long 470 or 625 nm light pulses of intensity 0–100 % (0.0–4.47 mW/mm²). Dual-color optogenetic experiments were performed with Chronos- and ChrimsonR-expressed in LEC_GLU and LEC_GABA, respectively stimulated with 1 ms 1–2 % (0.07–0.14 mW/mm²) 470 nm and 10 ms 100 % (4.47 mW/mm²) 625 nm light pulses that were empirically determined to allow spectral separation (intensity below cross-talk threshold of ChrimsonR activation by 470 nm light). Monosynaptic transmission from ChR2- or Chronos-expressing LEC_GLU was probed with 1 ms 1–2 % (0.1–0.14 mW/mm²) 470 nm light pulses that were empirically determined to elicit monophasic and TTX & 4-AP resistant EPSCs in CA3 PNs, as well as below population spike initiation threshold from LFP recordings in DG GC and CA3 SP layers. Conversely, poly-synaptic transmission from ChR2- or Chronos-expressing LEC_GLU was probed with 100% (3.76 mW/mm²) 470 nm light pulses that were empirically determined to elicit polyphasic and TTX & 4-AP sensitive EPSCs in CA3 PNs, as well as above population spike initiation threshold from LFP recordings in DG GC and CA3 SP layers. The medial CA3 recurrent input and proximal dentate gyrus mossy fiber input were stimulated electrically with ACSF-filled pipettes mounted on manual micromanipulators (Siskiyou) and placed in CA1 stratum radiatum (SR) to antidromatically activate CA3 and in the hilus near the border of the upper blade granule cell layer to directly activate DG axons, respectively. Post-synaptic responses were evoked with constant current stimulation units (Digitimer Ltd.) delivering 0.1 ms long 25–200 μA current pulses through monopolar electrodes. Pharmacological agents were added to ACSF at the following concentrations (in μM): 10 NBQX and 50 D-APV to block AMPA and NMDA receptors, 1–2 SR95531 and 2 CGP55845A to block GABAA and GABAB receptors, 10 clozapine N-oxide (CNO) to activate hM4D(Gi) DREADDs, 0.2–1 tetrodotoxin (TTX) to prevent sodic action potential generation, 100 4-aminopyridine (4-AP) to block KV1 potassium channels, 1–10 (2S,2'R,3'R)-2- (2',3'-dicarboxycyclopropyl)-glycine (DCG-IV) to inhibit glutamate release from mossy fiber terminals by activating mGluR2/3. Data was obtained using a Multiclamp 700B amplifier (Molecular Devices), sampled at 10 kHz, digitized using a Digidata 1550B AD/DA board (Molecular Devices), and acquired with the pClamp 10 software (Molecular Devices). Data analysis was performed in IgorPro (Wavemetrics) with custom-written code.

Two-photon calcium imaging:

All anesthesia, analgesia and pre-surgical procedures were performed as described above (see stereotaxic injections). C57BL/6J mice were first bilaterally injected with AAV2/1:CaMKII-GCaMP6f or AAV2/1:CaMKII-GCaMP8f and AAVretro:Syn-Cre in CA3, and AAV2/5:CaMKII-DIO-hM4D(Gi)-IRES-mCitrine and/or AAV2/5:hDlx-flex-PSAM or AAV2/5:Dlx-DIO-hM4D(Gi)-mCherry in LEC. Then, a unilateral circular 3 mm diameter craniotomy was made centered on the CA3 imaging coordinates (anterior–posterior relative to bregma: 1.4 mm; medial-lateral relative to midline: 1.6 mm). The skull fragment was removed, and a vacuum system was used to gently aspirate the overlying cortex and external capsule. Ice-cold ACSF was used to irrigate the area throughout the duration of the procedure. A cranial window (3 mm diameter, 1.7 mm length stainless steel cannula attached to 3 mm diameter glass coverslip) was then implanted over the area. The window was sealed to the skull (Vetbond), and a custom-designed 3D-printed plastic headpost was cemented over the skull. Mice were allowed to recover for 3–5 days before being placed under water restriction after which their weight was monitored daily to ensure it remained at least 80 % of baseline. Mice were head-fixed under the two-photon microscope on a treadmill belt (200 cm) and trained to run for 5 % sucrose water as described previously (57, 58). In vivo two-photon imaging was performed using a dual galvanometric and resonant laser scanning two-photon microscope (Ultima, Bruker), coupled to a tunable Ti:Sapphire laser (MaiTai eHP DeepSee, Spectraphysics) pulsed at a 80 MHz repetition rate and < 70 fs pulse width along with dispersion compensation. GCaMP fluorophore was excited at 920 nm, using a resonant scanning X-galvanometer (8 kHz) paired with a 6 mm standard scanning Y-galvanometer. The scanning system was mounted on a movable objective microscope (Ultima), equipped with an orbital nosepiece coupled to a 16 X, 0.8 NA, 3 mm water immersion objective (Nikon) and a piezo drive for angled imaging and ultrafast volumetric scanning. Imaging was performed at a scan speed of 29 fps, using 512 × 512 frame size (1.085 μm/pixel resolution). Fluorescence signal was detected using high-sensitivity GaAsP photomultiplier tubes (7422PA-40 PMTs, Hamamatsu). GCaMP-based calcium signals from CA3 PN somas and apical dendrites were imaged simultaneously on a single stable field of view in each mouse throughout the experiments. Mice were first imaged for a single 10 min-long session daily as they ran on the same textured belt (familiar) for water rewards delivered at random positions on the belt (random foraging, RF) for 7 days. On days 8–10, mice were subjected to a 10 min RF imaging session on the familiar belt and allowed to rest for 1 h before another 10 min session on the familiar belt immediately followed by a final 10 min session on a novel belt (a new belt was used for each day). Mice were injected with either saline, CNO (5 mg/kg, ip), or PSEM308 (3 mg/kg, ip) 30 min into the 1 h rest period between the familiar sessions. Thereafter, mice were trained in a series of goal-oriented learning (GOL) paradigms comprising different sessions defined by a combination of physical textures with olfactory and auditory cues (interleaved and pulsed for 1 s at 0.25 Hz) for 5–8 days. Each GOL block consisted of an A session with textures A, odor A (eg 10 % pentyl acetate in mineral oil), sound A (eg 4 kHz), and reward location A (eg 90–100 cm); an A’ session with textures A, odor B (eg 10 % (+)-α-pinene in mineral oil), sound B (eg 10 kHz), and reward location A’ (eg 150–160 cm); and a B session with textures B, odor B, sound B and reward location B (eg 30–40 cm). Textures, odors, sounds and reward locations assigned to A, A’ and B were changed for each GOL training. Mice were injected with either saline, CNO (5 mg/kg, ip), or PSEM308 (3 mg/kg, ip) 30 min before imaging. ΔLEC_GLU and ΔLEC_GABA mice were trained in one GOL with daily ip injection of the cognate ligand, and in another GOL with daily ip injections of saline. Control mice were trained in one GOL with daily ip injections of CNO or PSEM308, and in another GOL with daily ip injections of saline. As described previously (58), odor delivery and scavenging were achieved by routing odor-charged air through a nose cone under constant 1 L/min air flow and vacuum. Steady-state odor concentrations were verified using a photoionization detector (200B miniPID, Aurora Scientific). Environmental stimuli delivery and behavioral data recording were performed with custom-designed Arduino hardware. Imaging data was acquired with the PrairieView software (Bruker) and synchronized with the Arduino inputs and outputs. For each mouse, image sequences from all sessions were concatenated and processed by suite2p (98) for motion correction and ROI detection. ROI sets were thereafter manually curated and annotated using ImageJ (99). All subsequent analyses were performed in MATLAB (MathWorks) with custom-written code. For each ROI (spatial component), the activity trace (time component) was obtained by averaging across all pixels within the ROI at each time point (frame) and normalized to baseline fluorescence. Calcium transients were detected and demixed from overlapping ROIs using the fitness algorithm of the d-NMF package (59). Briefly, activity traces were detrended and z-scored to detect putative events as positive peaks overshooting a threshold of 2 standard deviations from baseline. Putative events were then screened using a correlation threshold of 0.27 between the spatial component of the ROI and the actual spatial footprint of the ROI in the movie frame at the time of the event. Events with offset within 200 ms of another event’s onset were merged, and events with rise time < 200 ms, or width < 200 ms were excluded from analysis. Subsequent analysis was performed on these curated events, with further behavioral restriction to run epochs as defined by a speed of the animal > 2 cm/s. Spatially-resolved activity rates were computed by smoothing and dividing the event counts by the occupancy within 5 cm-wide bins. For each session, spatial information content (100) was computed for each ROI as:

$$\mathit{\text{spatial}}\mathit{\text{information}}=\sum _{i=1}^N{P}_i{R}_i{\mathit{\text{log}}}_2(R_i)$$

where P_i is the normalized occupancy in the i^th spatial bin such that the sum across all P_i = 1, and R_i is the normalized value of the activity rate in the i^th spatial bin such that the sum across all R_i = 1. Statistical significance of spatial tuning was determined for each ROI within each session by shifting event times from −250 to 250 time bins and normalizing the spatial information content at zero shift to that of the 99^th percentile at non-zero shift. ROIs meeting criteria of a normalized spatial information content > 1 and a number of events > 4 were considered as significantly spatially tuned. Spatial tuning curves were parametrized by circularly smoothing rate maps with a sigma of 15 cm and fitting the resulting smoothed curve with multiple Von Mises functions increasing in number until the residuals fell below 25 % of the maximal value of the original rate map. Place field width was defined as the full width at half maximum of the fitted tuning curve component. Tuning curve (TC) correlations were computed as the correlation coefficient between tuning curves for each ROI active in both sessions and significantly tuned in at least one of the two sessions. Population vector (PV) correlations were computed by taking the correlation coefficient between spatial rate vectors constructed from ROIs active in both sessions, and averaging across all matching spatial bins (main diagonal of the PV correlation matrix). PV decoding of animal position was performed by taking the index of the maximal correlation value at each timepoint between spatial and time rate vectors constructed from tuned ROIs. Timepoints with no activity across all relevant ROIs were assigned the latest previously decoded position. Decoding error was computed as the angular distance between the real and the decoded positions.

Computational modelling:

We model a recurrent rate-based network in CA3 consisting of N=50 excitatory neurons. The total current $x_i$ in each neuron $i$ is modelled as

$$\tau \frac{d}{\mathit{\text{dt}}}{x}_i=-x_i+\sum _j{w}_{\mathit{\text{ij}}}{z}_j+I_E-I_I-I_{I\_\mathit{\text{lat}}}+I_{\mathit{\text{noise}}}$$

where $\tau =10[a.u.]$ is the neuronal time constant. The firing rate $z_i$ of the neuron $i$ is the current $x_i$ rectified at 0 with a saturation at 1. The neurons receive LEC_GLU current $I_E$, LEC_GABA current $I_I$, a noisy current $I_{\mathit{\text{noise}}}$ and lateral inhibition $I_{I\_\mathit{\text{lat}}}$. The noisy current is modelled as filtered noise, initialized at 0, with time constant $\tau \_\mathit{\text{noise}}=10[a.u.]$

$${\tau }_{\mathit{\text{noise}}}\frac{d}{\mathit{\text{dt}}}{I}_{\mathit{\text{noise}}}=-I_{\mathit{\text{noise}}}+\eta$$

where $\eta$ is drawn from a uniform distribution between −1 and 1. The lateral inhibition is modelled as

$$I_{I\_\mathit{\text{lat}}}=\gamma \sum _j{z}_j$$

with $\gamma =8.$ The recurrent weights $w_{\mathit{\text{ij}}}$ are plastic following a standard Hebbian learning rule

$$\Delta w_{\mathit{\text{ij}}}=\alpha {(z}_i{z}_j-\theta )$$

where $\alpha =1$ is the learning rate and $\theta =0.04$ is a threshold. There are no self-connections. All the incoming weights to a given neuron are normalized to have a L1 norm of 5. All the weights have a hard lower bound at 0 and an upper bound at 0.3. We initialized the weights at 0. We first simulate a learning phase and then a testing phase. During the learning phase, we stimulate an assembly of size 10 (neuron 1 to 10) with a $I_E$ of 0.4 and $I_I=0.$ We initialize the neural activity at 0 and we simulate the network for T=200 time steps, with a dt = 1. We then simulate a testing phase, where we probe our network with a pattern completion protocol. We stimulate the first 70% of the assembly and record the average activity of the remaining 30% of the neurons, averaged over the last 10 time steps. We simulate a total of 100 trials. To compute the variability of the representation, we compute the standard deviation of the neural activity, averaged across neurons. To mimic the silencing of LEC_GLU or LEC_GABA, we respectively perturb $I_E\mathit{\text{or}}{I}_I$ either during the learning phase or during the testing phase. To perturb LEC_GLU, we assume that the input information content is degraded. Thus, at each time point we stimulate a random subset of 4 out of 10 neurons within the assembly, and stimulate another 6 neurons as a random subset of the neurons outside the assembly, each with a current of 0.4. To perturb $I_I$, we set $I_I=0.17$ as LEC_GABA is disinhibitory.

Histology:

Mice were deeply anaesthetized with isoflurane (5 % for 5 min, inhaled) and perfused transcardially with ~20 mL of phosphate-buffered saline (PBS) followed by ~20 mL of 4 % paraformaldehyde (PFA) in PBS. Brains were removed and fixed in 4 % PFA in PBS at 4 °C for at least 24 h. Brains were washed in 0.3 M Glycine in PBS for 15 min followed by 3 × 15 min washes in PBS. 50–100 μm thick coronal slices were cut (vibratome, Leica VT1000S) and stored in PBS at 4 °C. Similarly, samples recovered from ex vivo electrophysiology (400 μm thick transverse slices) were fixed in 4 % PFA in PBS at 4 °C for at least 24 h, washed in 0.3 M Glycine in PBS for 15 min followed by 3 × 15 min washes in PBS, and stored in PBS at 4 °C. Tissue was permeabilized with 0.5 % Triton in PBS (PBST) for 2 × 20 min, blocked with 10 % NGS in 0.5 % PBST for 4 h, and incubated with primary antibodies in 3 % NGS 0.1 % PBST overnight at 4 °C. Tissue was then washed with 0.2 % PBST for 15 min followed by 3 × 30 min washes in PBS before being incubated with secondary antibodies and Streptavidin where applicable in 3 % NGS 0.1 % PBST for 24 h (100 μm slices) or 48 h (400 μm slices) at 4 °C. Lastly, slices were washed 5 × 15 min in PBS and mounted in Vectashield Hard Set Mounting Medium with DAPI (Vector Laboratories). Tissue stained for α-bungarotoxin underwent similar processing, except substituting PBS for Tris-buffered saline (TBS) pH-buffered at 7.4. Primary antibodies were rabbit anti-GFP (1:1000, ThermoFisher #A6455), chicken anti-GFP (1:1000, Abcam #13970). All secondary antibodies and dyes were purchased from ThermoFisher: Alexa Fluor 488-conjugated goat anti-rabbit (1:1000, #A11008), Alexa Fluor 488-conjugated donkey anti-chicken (1:1000, #A78948), Alexa Fluor 488-conjugated goat anti-chicken (1:1000, #A11039), Alexa Fluor 555-conjugated streptavidin (1:500, #S11227), Alexa Fluor 594-conjugated streptavidin (1:500, #S21381), Alexa Fluor 647-conjugated streptavidin (1:500, #S21374), Alexa Fluor 455-conjugated Neurotrace (1:200, #N21479); except Alexa Fluor 555-conjugated α-bungarotoxin (1:3000, Invitrogen) and Alexa Fluor 647-conjugated α-bungarotoxin (1:3000, Invitrogen). Samples were screened by epifluorescence imaging (Olympus VS200). Relevant samples were imaged with an inverted Zeiss Axio Observer Z1 confocal microscope using 10x (air, 0.3 NA), 20x (air, 0.8 NA) or 40x (oil, 1.3 NA) objectives (Zeiss) and 405, 488, 594, and 647 nm lasers for fluorophore excitation. Images were acquired as 1024 × 1024 pixels 16 bits Z-stacks with a 5 μm (10x, 20x) or 1 μm (40x) Z-step size and tiled in X and Y as needed to cover the samples, and thereafter stitched using the Zen Microscopy software (Zeiss). Further processing of confocal or epifluorescence images was done in ImageJ (99).

Statistics:

Results are reported ± SEM. Normality was tested with the Shapiro-Wilk and Jarque–Bera tests, and homoscedasticity was tested using Barlett’s test to choose between parametric and nonparametric statistical analysis. Statistical significance was assessed using the χ² test, Student’s T test, Mann-Whitney U test, paired Student’s T test, Wilcoxon signed-rank test, one-way ANOVA, Kruskal-Wallis ANOVA, repeated-measures ANOVA, Friedman ANOVA, Tukey post hoc test, Nemenyi post hoc test, Dunn-Holland-Wolfe post hoc test, or two-way ANOVA where appropriate. The symbols *, **, and *** denote P values <0.05, <0.01, and <0.001, respectively.

Materials availability:

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Jayeeta Basu (Jayeeta.Basu@nyulangone.org). This study used the AAV2/7:h56D-ChR2-sfGFP custom-designed viral reagent, courtesy of Boris Zemelman. Please contact the corresponding author for further details. All data reported in this paper will be shared by the lead contact upon request. All original code that has been deposited on github is available to reviewers and will be publicly available as of the date of publication. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Supplementary Material

Figs. S1 to S16

Text S1

Table 1.

Fraction of licks in reward zone p-values (Tukey post hoc test).

	control ligand	control saline	ΔLECgaba ligand	ΔLECgaba saline	ΔLECglu ligand
control saline	0.882
ΔLECgaba ligand	0.001	0.042
ΔLECgaba saline	0.590	0.999	0.032
ΔLECglu ligand	< 0.001	0.002	0.757	< 0.001
ΔLECglu saline	0.937	> 0.999	0.003	0.963	< 0.001

Data and materials availability:

All data are available in the main text or the supplementary materials.