Synaptic plasticity in the perforant pathway drives inhibitory reorganization enhancing dentate gyrus functionality

Summary

In the dentate gyrus (DG), tight inhibitory control confines the activity of granule cells (GCs)—a key characteristic for its proposed role as a pattern separator. Nonetheless, a conundrum arises concerning the balance between sparseness of GC firing and the activity level needed for efficient transmission downstream to CA3. Using in vivo electrophysiology, pharmacogenetics, and computational modeling, we identified a synaptic plasticity mechanism that decouples excitation from inhibition, enhancing information encoding and transmission without compromising pattern separation. Long-term synaptic potentiation of the perforant pathway, in addition to strengthening the glutamatergic synapses into GCs, depressed feedforward perisomatic inhibition. Computational modeling revealed a functional reorganization of the DG inhibitory network that decorrelated excitation/inhibition, enhanced GC burst firing, and improved temporal and spatial discrimination within the entorhinal-cortex-DG-CA3 network. Consistently, targeting parvalbumin + interneurons to reduce perisomatic inhibition during memory encoding improved pattern separation in freely behaving mice. Overall, our findings uncover a plasticity mechanism that boosts DG output while preserving its pattern separation function.

Graphical abstract

Highlights

• •Perforant-path LTP depresses feedforward inhibition onto granule cells
• •Perisomatic disinhibition rebalances activity in the DG inhibitory network
• •This state enhances GC bursting and DG-CA3 coupling, improving pattern transmission
• •Experimental PV cell inhibition improves hippocampal pattern separation

Behavioral neuroscience; Neuroscience

Introduction

The dentate gyrus (DG) is a hippocampal structure involved in pattern separation, a process critical for memory formation.¹ The DG transforms the dense multisensory information received from the entorhinal cortex (EC) into a sparse code that minimizes the overlap between the activated neuronal populations, facilitating the discrimination of similar input patterns. The low responsiveness of granule cells (GCs) in the DG is central to the mechanism for pattern separation. Although GCs receive large numbers of simultaneous inputs from the EC,² their low excitability and the tight inhibitory control^3,4 result in a low firing activity that filters the EC input.^2,5 Regulation of inhibitory activity in this circuit may, therefore, play a critical role, as a high inhibitory tone is required for sparse GC coding^6,7 but excessive inhibition would prevent effective activation of the downstream CA3 network.

At the microcircuit level, inhibitory synapses between DG interneurons support gamma-range oscillations through ultra-fast GABAergic transmission,⁸ while excitatory inputs onto interneurons are mediated by submillisecond AMPA receptor-driven currents that allow for precise spike timing.⁹ The excitation/inhibition (E/I) balance in the DG can be modulated by short- and long-term synaptic plasticity^10,11,12 and neuromodulation.¹³ The pioneering work of Bliss and Lomo,¹⁴ in which long-term potentiation (LTP) was first described, already documented a shift in the input-output transformation in GCs, so that the same excitatory synaptic input in response to perforant pathway (PP) stimulation produced a higher firing output after LTP. In LTP experiments combining electrophysiological recordings in the DG with whole-brain fMRI, the induction of LTP further unveiled a functional reorganization that extended beyond the local DG circuit to the entire hippocampus and other extrahippocampal structures such as the prefrontal cortex, perirhinal cortex, and nucleus accumbens.^15,16,17

In this work, we combined in vivo electrophysiological recordings with pharmacogenetic tools and computational modeling to investigate how synaptic plasticity in the PP sets the E/I balance in the DG that regulates the activity of GCs and their capacity to transmit and separate firing patterns.

Results

LTP in the PP depresses feedforward inhibition in GCs

We used layer-specific local field potential (LFP) decomposition with independent component analysis (ICA) to quantify the E/I balance in the DG. We used linear array recordings in anesthetized rats with stimulation electrodes implanted across the dorsal hippocampus and in the PP (Figure 1A, see supplemental information; Figure S1A).^15,16 Applying ICA^18,19,20 to field potentials evoked by subthreshold stimulation of the PP (e-LFP), we separated two main components from the DG that we named, based on their topographic localizations, e-PP and e-Hilar (Figures 1B and 1C) (see also Figures S1B and S1C). These components show dipoles in the current source density (CSD) depth profile in the mid-molecular layer and the hilus, respectively. Specifically, e-PP corresponds to the terminal fields of medial entorhinal cortex (MEC) layer II axons, exhibiting clear CSD sinks and excitatory postsynaptic potential (EPSP) in response to PP stimulation,²⁰ reflecting the excitatory activity in the dendrites of GCs elicited by PP inputs.

Figure 1.

Recording of excitatory and inhibitory components of the DG LFP

(A) Scheme of the hippocampus highlighting the DG in color (blue: molecular; red: granular; green: hilar layers) and illustrating the location of the recording electrode (coordinates respect to Bregma: −2 AP, +1.5 ML, −2 DV) and the stimulating electrode (coordinates respect to Bregma: −4.3 AP, +2.5 ML, +1.4 DV, −12° at the sagittal plane).

(B) Left: Evoked LFP recordings (e-LFP) across the DG in response to a PP stimulation (dashed lines) overlaid on the corresponding CSD map (color-coded). The upper trace (black) highlights the evoked potential in the center of the hilar region. Right: Decomposed ICs (virtual LFPs) overlaid on the corresponding CSDs are shown for the PP-IC (e-PP, middle panel) and Hilar-IC (e-Hilar, right panel). Upper traces highlight the corresponding e-PP (blue) and e-Hilar (green).

(C) Spontaneous activity in the LFP, PP-IC, and Hilus-IC corresponding to the experiment in (B). The black arrow indicates the evoked response shown in (B).

(D) Effect of gabazine (blue) and bicuculine (purple) on e-Hilar and e-PP amplitudes, normalized to the average value in each experiment. Gabazine: Mann-Whitney U test: U = 318, p < 0.0001; bicuculline: Mann-Whitney U test: U = 69, p < 0.0001. Bars represent the mean ± SEM.

(E) Effects of CGP (orange) over the pairing pulses on e-Hilar and e-PP amplitudes (measured as the ratio of the value after pairing to the value before pairing) in control conditions and in the presence of CGP. Paired t test: t(34) = 2.645, p = 0.0123. Bars represent the mean ± SEM.

(F) Comparison of the evoked potentials in the raw LFP and ICs in the control condition and after LTP induction. The arrow points to the evoked potential in the Hilar-IC that is strongly reduced after LTP.

(G) Upper panel: effects of LTP on the amplitude of the evoked IC’s (e-PP: paired t test: t(7) = 4.636, p = 0.0024; e-Hilar: paired t test: t(7) = 6.707, p = 0.0003). Lower panel: effect of the NMDA antagonisms with D-AP5 (e-PP: unpaired t test: t(10) = 3.766, p = 0.0037; e-Hilar: unpaired t test: t(10) = 2.293, p = 0.0448). Bars represent the mean ± SEM.

(H) LTP effects on the spontaneous IC’s power (upper panel, paired t test: t(11) = 2.9, p < 0.01), and their correlation (lower panel, paired t test: t(11) = 2.1, p < 0.05), normalized to the average value in each experiment. Bars represent the mean ± SEM.

(I) Comparison of spectral GC (mean ± SEM) from PP-IC to Hil-IC (upper panel) and from Hil-IC to PP-IC (lower panel) between the control (black) and LTP (red) conditions. The horizontal black line in the lower panel indicates the frequency band with significant differences between conditions (permutation test, see STAR Methods). ∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001.

DG, dentate gyrus; LFP, local field potential; PP, perforant pathway; ICs, independent components; CSD, current source density; LTP, long-term potentiation; GC, granule cell.

The e-Hilar exhibited active outward currents in response to PP stimulation, primarily in the GC layer (putative perisomatic inhibition). The latency of the evoked potentials was 5.1 ± 0.03 and 7.4 ± 0.05 ms (mean ± SEM, n = 8 animals) for the e-PP and e-Hilar, respectively, consistent with monosynaptic and disynaptic (feedforward) evoked potentials (Figure S1C). Microinjection of GABA_A antagonists (bicuculline or gabazine) into the DG selectively blocked the e-Hilar component without affecting e-PP (Figure 1D), confirming the inhibitory nature of the feedforward potential. Application of a GABA_B receptor antagonist, CGP, reduced the inhibition observed in a paired-pulse protocol (Figure 1E), further supporting the GABAergic nature of the e-Hilar component. Although this component likely reflects both feedforward and feedback GABAergic inputs during its ongoing resting state activity (Figures 1H and 1I), the feedforward nature of the evoked response shown in Figures 1B–1G is supported by its disynaptic latency (Figure S1C) and by its recruitment at PP stimulation intensities that are subthreshold for GC firing, and thus insufficient to trigger feedback inhibition.

We then investigated the effect of synaptic plasticity on the E/I balance. We induced LTP in the PP (see STAR Methods) and recorded both PP-evoked potentials and spontaneous activity. The comparison of the evoked subthreshold potentials in each generator before and after LTP induction revealed two distinct effects (Figure 1F). First, LTP produced the expected potentiation of the glutamatergic input on GCs (e-PP; Figure 1F, blue traces; see also Figure S2 for suprathreshold stimulus-response curves). Second, it produced a strong depression of the feedforward inhibition (e-Hilar; Figure 1F, green traces). Figure 1G (upper panel) shows the quantification of both effects. Importantly, the LTP protocol applied in the presence of the NMDA receptor antagonist D-AP5 microinjected in the DG (see STAR Methods) failed to potentiate the e-PP and largely prevented the e-Hilar reduction (Figure 1G, bottom panel). Finally, the bimodal effect of LTP on the e-PP and e-Hilar was confirmed with a second LTP protocol (theta-burst stimulation, Figure S3).

Using the spontaneous activity of the PP and Hilar components, we further measured the power of each component and the correlation between them (Figure 1H) and found that LTP increased the PP/Hilar power ratio and decreased their correlation, suggesting that LTP induction decoupled the excitatory and inhibitory activities. We performed a Granger causality analysis between the two signals and found a strong influence of the inhibitory activity over the excitatory activity in GCs (Figure 1I). This influence was decreased after LTP induction, specifically in the gamma frequency band, consistent with the inhibitory activity range received by GC from DG interneurons.² Overall, our results show a decrease in the feedforward inhibitory circuit in the EC→DG pathway.

LTP activates the hilar circuit responsible for the E/I imbalance

To investigate the mechanism responsible for the E/I imbalance, we built a neural circuit model of the DG (see STAR Methods). Figure 2 shows the model comprising five neuronal populations—MEC layer II (MEC LII) inputs, GCs, basket cells (BCs), hilar inhibitory interneurons (HILs), and mossy cells (MCs)—chosen for their critical roles in dentate feedforward and feedback circuits. MEC LII provides primary excitatory inputs to GC dendrites and local GABAergic interneurons, driving feedforward excitation and inhibition.²¹ GCs, the principal excitatory neurons, recruit interneurons and MCs via mossy fibers and local collaterals.²² BCs provide a strong inhibitory control over GC firing and receive direct inputs from MEC, GC, and MC collaterals.^22,23 HIL cells represent inhibitory interneurons, such as hilar commissural associational path (HICAP) or hilar PP-associated (HIPP) cells, that limit recurrent excitation.²² Finally, MCs mediate prominent feedback excitation onto GCs and interneurons.²¹ Other interneurons (e.g., HIPP, HICAP, and Molecular layer Perforant Path (MOPP) associated cells) were not explicitly modeled due to functional overlap already captured by the BC and HIL populations.^23,24

Figure 2.

Modeling results of the LTP-induced reorganization of the DG circuit

(A) Scheme of the DG circuit’s model. Five populations were considered: MEC, GCs, MCs, BCs, and HIL interneurons. Circles and triangles at the end of the connections indicate excitatory and inhibitory synapses, respectively.

(B) Power spectra of different cell populations in the model.

(C) Coherence between the granular LFP (average population voltage) and the excitatory (left) and inhibitory (right) postsynaptic currents in individual GCs.

(D and E) Amplitude ratio and correlation between the excitatory and inhibitory activities onto GCs. Different potential effects of LTP are tested and compared to the baseline (control) condition: concomitant potentiation of the MEC connection to GCs and BCs, depression of the BCs connection to GCs, and potentiation of the MEC connection to GCs alone. (D) Ratio (mean ± SEM): Mann-Whitney U test: control vs. LTP MEC-GC/MEC-BC: U = 44, p < 0.0001; control vs. LTD BC→GC: U = 77, p = 0.0006; control vs. LTP MEC→GC: U = 28, p < 0.0001. (E) Correlation (mean ± SEM): Unpaired t test: control vs. LTP MEC-GC/MEC-BC: t(38) = 6.404, p < 0.0001; control vs. LTD BC→GC: t(38) = 1.018, p = 0.3151; control vs. LTP MEC→GC: t(38) = 7.851, p < 0.0001.

(F) Firing rate of the different DG populations (mean ± SEM). GC: t(38) = 8.3, p < 0.0001; BC: t(38) = 12.6, p < 0.0001; HIL: t(38) = 12, p < 0.0001; MC: t(38) = 7.4, p < 0.0001.

(G) Granger causality between the inhibitory populations and GCs (mean ± SEM). GC→HIL: t(38) = 5.6, p < 0.0001; HIL→BC: t(38) = 8.6, p < 0.0001; BC→GC: t(38) = 10.3, p < 0.0001. ∗∗∗p < 0.001; ∗∗∗∗p < 0.0001.

DG, dentate gyrus; MEC, medial entorhinal cortex; GCs, granule cells; MCs, mossy cells; BCs, basket cells; HILs, hilar inhibitory interneurons; LFP, local field potential; LTP, long-term potentiation.

Neurons were modeled using the Izhikevich equations²⁵ and assumed to be connected by chemical synapses simulating the kinetics of the most common receptors participating in the circuit²⁶: NMDA, AMPA, and GABA_A (see Tables 1 and 2). The circuit parameters were tuned to reproduce the observed dynamics of the DG, in particular, the characteristic theta-gamma frequency component recorded in vivo in rats.² The MEC input mainly oscillates in the theta frequency range,^26,27,28 while the Hilus region does it in the gamma band,^29,30 as shown in Figure 2B. With the chosen parameter values (see Tables 3, 4, and 5), the model reproduced the coherence between the excitatory postsynaptic current (EPSC) and inhibitory postsynaptic current (IPSC) in GCs and the LFP simultaneously recorded in the DG (Figure 2C), as shown experimentally.² For simplicity and without losing generality, we modeled LTP as an instantaneous change in the synaptic weights in the PP inputs.

Table 1.

Izhikevich neuron parameters

Neuron type (rows)/Izhikevich parameter (Columns)	a	b	c	d
Fast spiking	0.1	0.26	−65	1
Regular spiking	0.02	0.2	−60	8
Burst spike	0.02	0.2	−50	2

Value of Izhikevich parameters (a, b, c, and d) obtained from a study²⁵ to generate the spiking dynamics of the main neurons used in the DG circuit model: fast spiking, regular spiking, and bursting neurons.

Table 2.

Parameters for synaptic current simulation in the DG model

Synapses	Ohmic conductance gsyn(nS)	Time constant τdecay(ms)	Reversal Potential Esyn(mV)
AMPA	0.5	5.6	0
NMDA	0.5	200	0
GABAA	2.0	5.6	−85

Value of the synaptic conductance (g_syn), time constant (τ_decay), and reversal potential (E_syn) for the AMPA, NMDA, and GABA_A synapses.

Table 3.

Parameters of connectivity for the DG computational model

From (row)/to (column)	Connectivity (%)
	GC	BC	HIL	MC
EC	5	2	–	–
GC	–	2	2	2
BC	12	10	–	20
HIL	–	60	30	20
MC	10	20	20	15

The connectivity of a circuit indicates the proportion of randomly selected neurons in the target population that receive input from a single presynaptic neuron in the projecting population.

Table 4.

Synaptic connection parameters for the DG computational model

From (row)/to (column)	Maximum Peak Kinetics (rpeak)
	GC	BC	HIL	MC
EC	AMPA: 0.05 NMDA: 0.05	AMPA: 0.01 NMDA: 0.01	–	–
GC	–	AMPA: 0.02 NMDA: 0.02	AMPA: 0.1 NMDA: 0.1	AMPA: 0.01 NMDA: 0.03
BC	0.4	0.2	–	0.3
HIL	–	0.11	0.3	0.4
MC	0.3	0.2	0.4	0.4

The maximum release probability represents the strength of the chemical synapse between the presynaptic neurons (rows) and the postsynaptic neurons (columns).

Table 5.

Parameters of Poisson noise

Parameters	Value
Mean rate (λ)	80 Hz
Maximum peak kinetics (rpeak)	AMPA: 0.1
Synaptic conductance	0.3 ns (for EC and GC population) 0.2 ns (for HIL, BC, and MC)

Parameters used for the synapse of the spike train generated by Poisson distribution, including the mean rate of the spike train, the maximum peak of released neurotransmitters, and the synaptic conductance of the synapse.

In the DG, feedforward inhibition is primarily carried out by perisomatically targeting parvalbumin (PV)-positive BC interneurons.^31,32 A depression of the PP-evoked feedforward inhibition, as observed experimentally, was reproduced in the model by reducing the strength of BC→GC synapses. This manipulation led to an increased E/I ratio (Figure 2D) but failed to reduce the correlation between excitatory and inhibitory components (Figure 2E), in contrast to in vivo findings (Figure 1H). A more parsimonious approach was to apply equal potentiation to MEC inputs onto both GCs and BCs. Under this condition, we observed reductions in not only the EPSC/IPSC amplitude ratio but also the correlation between excitatory and inhibitory currents (Figures 2D and 2E), consistent with experimental data. For completeness, we also simulated isolated potentiation of the MEC→GC connection and obtained similar results (Figures 2D and 2E). Additionally, we explored various combinations of LTP and LTD in the MEC-GC-BC circuit (Figure S4). Across all simulations, our results consistently underscore the importance of GC population activity in reproducing the feedforward disinhibition effect.

We next conducted a numerical analysis of the firing rate for each population in the model (Figure 2F) and calculated Granger causality to capture the directional flow of information within the circuit (Figure 2G). LTP in the PP increased GC output, thereby enhancing the recruitment of HILs, which subsequently heightened their inhibition over BCs. This reconfiguration of activity in the model resulted in decreased responsiveness of BCs to MEC inputs and reduced the influence of BCs over GC activity, providing a possible explanation of the experimental results. Interestingly, a similar phenomenon was recently observed in experiments in vitro, inducing perisomatic disinhibition over GCs through sustained activation of hilar somatostatin-expressing interneurons.¹³ Notably, in the model, both HILs and the MCs played pivotal roles in the circuit, as the removal of either of these populations eliminated the disinhibition associated with LTP (Figure S4).

Inhibition of the BC population recapitulates the LTP effect on DG E/I balance

Although depression of the specific BC→GC connection alone is not sufficient to fully reproduce the experimental results (Figure 2E), a reduction in BC activity is necessary for the disinhibitory mechanism triggered by LTP. To test whether overall decreased BC activity is sufficient, we conducted a new simulation in which BC activity was reduced to 30%, without altering any other model parameters. As expected, this manipulation led to a reduction in the IPSC in GCs within the gamma frequency range, while EPSC remained unchanged (Figure 3A). Importantly, this selective reduction in BC activity also decreased the coherence between EPSCs and IPSCs in the gamma range in GCs (Figure 3B), thereby reproducing the experimental findings. To experimentally validate this model prediction, we performed in vivo electrophysiological recordings in PV-Cre transgenic mice, which enabled selective manipulation of BC activity in the DG.

Figure 3.

Role of BCs (PV + cells) in the DG E/I balance

Modeling results.

(A) Power spectrum of the EPSC and IPSC in GCs (left and right panels, respectively) in control (black) conditions and during a 70% reduction in BC activity (yellow).

(B) Coherence between EPSC and IPSC in GCs in the same conditions as in (A) (permutation test, see STAR Methods).

Experimental results: (C) Scheme of the AAVs DREADDs bilateral injection (left) and anatomical validation of the DG injection site and viral expression (right). Nuclear staining with DAPI (blue) and DREADD expression (red). Scale bars: 400 μm.

(D) Top: scheme of the DG and the position of the recording electrode (from bregma: −2 AP, +1.5 ML, −2 DV) and the stimulating electrode in the PP (from bregma: −4.3 AP, +2.5 ML, +1.4 DV, 12° angle). Middle: scheme of the experimental design. Bottom: representative traces of evoked population spikes (PSs) in the DG pre- (black) and post- (yellow) CNO administration. The CNO-induced effect is quantified in the right histogram (n = 3). The PS amplitude increased post-CNO injection in mice with DREADD expression in PV + cells (PV inhibition, yellow) in comparison to mice without DREADD expression (sham, black).

(E) Experimentally obtained power spectrum of the PP-IC and Hilar-IC (left and right panels, respectively) before (black) and after (yellow) the CNO injection (permutation test, see STAR Methods).

(F) Coherence between PP-IC and Hilar-IC in the same conditions as in (E) (permutation test, see STAR Methods).

BCs, basket cells; EPSC, excitatory postsynaptic current; IPSC, inhibitory postsynaptic current; GCs, granule cells; DG, dentate gyrus; PP, perforant pathway; CNO, clozapine-N-oxide; PV, parvalbumin.

Cre-dependent adeno-associated viruses (AAV5-hSyn-DIO-hM4D(Gi)-mCherry) injected in the DG of PV-Cre mice were used to selectively express a pharmacogenetic tool, DREADD (Designer Receptors Exclusively Activated by Designer Drugs), in PV + cells. Upon administration of the synthetic ligand clozapine-N-oxide (CNO; 1 mg/kg, i.p.), the activity of PV + cells was reduced (see STAR Methods; Figure S5). PV-Cre sham animals received a corresponding AAV expressing only mCherry. The anatomical location of the injection, which primarily targeted the DG with minor spread to the area CA3c, was confirmed histologically at the end of the experiment (Figure 3C). The functional effect of the manipulation was validated both in vitro (Figure S4) and during the in vivo experiment as an enhanced response of GCs to PP stimulation (Figure 3D). In these conditions, inhibition of PV cells resulted in a selective reduction of the gamma frequencies measured in the spontaneous Hilar-IC activity, with no additional changes in the PP-IC power spectrum (Figure 3E). Notably, coherence between the two components decreased specifically in the gamma frequency band (Figure 3F), closely mirroring the model predictions (Figures 3A and 3B). Overall, these results support a critical role of the PV + BCs in the regulation of DG E/I balance operated by LTP.

Perisomatic disinhibition improves DG encoding and pattern separation transfer to CA3

We reason that a mechanism regulating the E/I balance based on the dynamic reconfiguration of inhibitory cell populations decreasing feedforward inhibition will have important functional consequences. To address this question, we built a multicompartmental GC model, following the work of Schmidt-Hieber et al.,³³ containing three dendritic branches and the soma (a total of 27 compartments), and considering GC excitatory inputs at medial and proximal dendritic compartments from the MEC and MCs, respectively (Figure 4A), inhibitory inputs from HIPP cells synapsing distal dendritic segments, and HICAP interneurons innervating proximal segments. BCs were modeled as single-compartment units that project to the soma of the GC and received excitatory inputs from MEC, MCs, and GCs and inhibitory inputs from the HIL population (Figure 4A). Our GC model exhibited the capacity to generate both regular spiking and bursting regimes (Figure S5), consistent with experimental findings.^34,35 By incorporating various temporal and spatial patterns through the MEC input, we calculated the resulting GC output.

Figure 4.

Functional role of inhibitory perisomatic vs. dendritic inhibition over the GCs

(A) Scheme of the detailed multicompartmental model of an individual GC coupled to different inhibitory and excitatory sources. Circles and triangles indicate inhibitory and excitatory synapses, respectively. MEC, MC, and HIL sources represent inputs from the MEC, MC, and HIL populations (see Figure 3A). HICAP and HIPP sources represent inhibitory inputs into the distal and proximal dendrites.

(B) Consistency and firing frequency of the GC output.

(C) Single spike vs. burst firing probability of GCs, as a function of the firing rate of basket cells (BCs).

(D) Scheme of the temporal pattern encoded into the GC population. Four patterns of 2 s duration each were generated and injected into the GC via the MEC input. The process was repeated for 30 s.

(E) Same as (C) but for the temporal pattern input; baseline: original configuration of the inhibitory inputs; DD, dendritic depression (dendritic inhibitory inputs over GCs were reduced); PD, perisomatic depression (perisomatic inhibitory inputs were reduced).

(F) MI between the MEC input and the GC output.

(G) Scheme of the spatial PST; 1,000 independent GCs were considered.

(H–K) PSE (H), number of activated neurons by the pattern (I), probability of generating bursts (J), and the pattern separation transmission (K) for a pair of input patterns as a function of the percentage of overlap between patterns. All measurements are the mean ± SEM over fifty random pairs of input patterns. ∗∗∗∗p < 0.0001.

GCs, granule cells; MEC, medial entorhinal cortex; MCs, mossy cells; HILs, hilar inhibitory interneurons; HICAP, hilar commissural associational path; HIPP, hilar PP-associated; MI, mutual information; PST, pattern separation task; PSE, pattern separation efficiency.

We first investigated the contribution of perisomatic inhibition to the GC input/output transformation. We found that the consistency, measured as the correlation between two GC outputs in response to the same MEC input, increased with a decrease in BC activity. In these conditions, the GC firing rate and the probability of burst firing (PB) also increased (Figures 4B and 4C). Next, we studied the temporal coding capacity of GCs. We generated four concurrent temporal patterns of afferent MEC inputs (see STAR Methods; Figure 4D) and analyzed the response patterns of GCs and the mutual information (MI) between the MEC input and GC output. Our findings highlighted the crucial role of perisomatic disinhibition in the emergence of bursting activity within temporal patterns (Figure 4E). Regarding the temporal coding efficiency, we found that disinhibition increased MI, with disinhibition in the perisomatic compartment being more efficient than that in the dendritic compartment (almost tripling MI, Figure 4F).

An increase in the consistency and MI in GCs might indicate enhanced temporal encoding. However, a function like pattern separation might suffer from the increased excitation of GCs and the possible reduction in sparseness.^1,36,37,38 To investigate the impact of disinhibition on pattern separation, we extended the model by considering 1,000 independent GCs and used two spatial MEC input configurations that generated GC activity patterns with different degrees of overlap (Figure 4G). We started by evaluating the pattern separation efficiency (PSE), defined as the ratio of the overlap between two patterns and the active population (see STAR Methods). Reducing inhibition in any compartment in the model worsens PSE, as expected from an increase in GC firing and the loss of sparseness (Figure 4H). However, we reason that PSE performance does not necessarily imply a robust transmission of the separated patterns. In fact, the best PSE performance obtained under baseline inhibitory conditions is achieved at the cost of activating a very low number of GCs (compare Figures 4H and 4I). This occurred due to the tight inhibitory control, which resulted in a drastic loss of MI between the MEC input and GC output (Figure 4F). A balance is required between sparse firing (to filter the dense MEC input) and sufficient mass action to impact the dynamics of the downstream CA3 network. Therefore, a better indicator should consider not only the PSE but also the number of active neurons (Figure 4I) and the probability of burst firing (Figure 4J). The latter is known to act as a detonator with a strong impact on CA3 pyramidal cell firing and synaptic plasticity.³⁹ To investigate the role of the inhibitory reconfiguration on information transmission and pattern separation, we defined the pattern separation transmission (PST) index as a function of the three previous variables (see STAR Methods).

Figure 4K shows that the PST is largely enhanced by reducing perisomatic inhibition. Dendritic disinhibition also enhanced PST, but to a smaller degree, as it produced a smaller increase in spiking activity and, most importantly, did not contribute to burst firing. Finally, considering the complete depression of inhibition in the perisomatic-dendritic axis in the model, the PST is reduced to 0 due to the massive recruitment of GCs leading to overlapped patterns. Therefore, the reduction in perisomatic inhibition measured experimentally after LTP induction might be balancing firing sparseness and bursting to support pattern separation while optimizing information transmission between the MEC and CA3.

Experimentally reducing PV + cell activity in the DG during memory encoding enhances subsequent pattern separation

To directly test this model-based hypothesis, we used the novel object location (NOL) protocol and PV-Cre mice with DREADDs to reduce BC activity in the DG. In this behavioral task, animals must discriminate a change in the position of two objects within a familiar environment (Figure 5A). During the encoding phase, following habituation, animals explore both objects equally. Subsequently, one of the objects is displaced by a given distance. If the animal detects the spatial change, it preferentially explores the relocated object. As shown in Figure 5B, for displacements that are normally indiscriminable under baseline conditions (animals injected with saline; equal exploration of objects = 0.5 discrimination index), the same animals (injected with CNO in a counterbalanced design; see STAR Methods) reliably discriminated the change when PV + cell activity in the DG was reduced during encoding, thereby demonstrating enhanced pattern separation. Moreover, when objects were progressively displaced in 5 cm steps, animals under baseline conditions (saline-injected) were also able to discriminate the spatial change once the separation was large enough (15 cm)—that is, when the overlap between patterns was reduced (Figure 5C; see STAR Methods). The complete sequence of object displacements and associated discrimination indexes are shown in Figure S7. Male and female mice were used in these experiments, and no effect of sex was found. Overall, these experiments provide strong validation of the model’s hypothesis.

Figure 5.

Reduced perisomatic inhibition enhances spatial pattern separation in mice

(A) Schematic of the spatial pattern separation task protocol, illustrating object displacements along the sequence of trials (t0 corresponds to the encoding phase and t1-t4 to the testing phase). Red dots indicate the positions of the objects within the arena, with faded red shadows representing their locations in previous trials. Red arrows denote object displacements of 5 cm (short arrows) or 10 cm (long arrow, first displacement only). The lower panel depicts the experimental design across two weeks (n = 8 mice), showing counterbalanced CNO (PV inhibition) and vehicle (control) injections.

(B) Comparison (two-tailed paired t test) of individual discrimination indices (DIs) for the first displacement (t1) with (CNO-injected) and without (saline-injected) PV activity inhibition during the encoding phase (t0) (t(7) = 2.53, p = 0.039). Dots represent individual values, and bars indicate the mean.

(C) Effect of PV inhibition on the distribution of discrimination thresholds. The shift in the distribution indicates that the mice were able to detect object displacements at shorter distances when encoding occurred under PV inhibition (CNO-injected) compared to control conditions (saline-injected). Two-tailed exact Wilcoxon matched-pairs signed-rank test (W = −21, n = 8 pairs, p = 0.031). Data are presented as the mean ± SEM. ∗p < 0.05.

Discussion

In this work, we found, by combining in vivo experimental data and modeling results, that synaptic plasticity in the PP, the main input from the entorhinal cortex to the hippocampus, decouples excitation and inhibition in the DG, facilitating information encoding and transmission. The key finding was an unexpected reduction in feedforward inhibition onto GCs driven by synaptic potentiation in the PP. Our experiments revealed that this effect predominantly occurred in the inhibitory perisomatic compartment of GCs.

Our results showed that LTP regulates both excitatory and inhibitory inputs into GCs. In addition to the well-known potentiation of the glutamatergic input from the EC, the disynaptic feedforward inhibition was depressed, a change that altered the E/I balance and thus the input/output transformations in GCs in favor of greater responsiveness. The pioneering work by Bliss and Lomo on synaptic plasticity and the discovery of LTP,¹⁴ already showed that the firing activity in the population was larger than would correspond to the sole increase in synaptic activity after LTP induction. Thus, the so-called EPSP-to-spike potentiation could be, in part, the result of a reduced feedforward inhibition as the one shown here. This phenomenon has been also observed in the CA1 region, where studies have demonstrated that synaptic plasticity-induced reductions in inhibitory input can amplify the EPSP-to-spike ratio.⁴⁰ These findings in CA1 highlight a potentially conserved mechanism across hippocampal regions.

In a recent study, Hainmueller et al.⁴¹ used two-photon calcium imaging to show that PV + neurons in the DG decrease their activity, while dendrite-targeting hilar interneurons (somatostatin + neurons) increase their activity, during novelty exploration. Our results are consistent with these findings and provide a circuit-level explanation for this shift in inhibitory balance, which produces a state of perisomatic disinhibition that facilitates the encoding of new information and enhances pattern separation. Notably, the reduction in perisomatic inhibition reported by Hainmueller⁴¹ occurs under novelty conditions—precisely when a network state that favors the integration of new, behaviorally relevant information into existing memory representations is required. Under these conditions, potentiation of glutamatergic inputs from the entorhinal cortex, together with inhibitory rebalancing, is expected to promote more effective communication within the hippocampal formation. In this regard, our findings may also account for previous fMRI experiments showing facilitated communication within the hippocampus and from the hippocampus to cortical and subcortical mesolimbic structures after LTP induction.^15,16,17 Overall, our results point to a DG mechanism operated by synaptic plasticity and controlled by BCs that regulates communication in memory-related networks.

Computational modeling offered mechanistic explanations. A first computational model of the DG, calibrated to match experimental data, successfully reproduced key experimental findings and suggested that increased excitatory drive from the PP following LTP induces a functional reorganization of the local hilar network. Specifically, the model demonstrates that potentiation of the excitatory input from the EC, combined with intrinsic DG connectivity, leads to enhanced recruitment of HILs via GC activity, a process further amplified by MCs. As these interneurons become more active, they exert increased inhibitory control over BCs, resulting in a net reduction of feedforward inhibition onto GCs and thereby establishing a new set point in the E/I balance within the DG. These findings are consistent with experimental observations, including the effects reported under cholinergic modulation of the DG.¹³

A second computational model, with more anatomical detail, showed the distinct value of perisomatic vs. dendritic inhibition and highlighted the functional relevance of the LTP-induced reorganization of the inhibitory balance. Perisomatic disinhibition improved transmission from the MEC to CA3 (increased mutual information), enhanced reliability (consistency) of input/output transformations, and allowed GCs to better transmit temporal and spatial patterns to CA3 while preserving their pattern separation ability.

These mechanistic insights advance our understanding of GC firing dynamics. GCs fire at remarkably low rates,^25,42 likely due to their intrinsically low excitability combined with strong inhibitory regulation from local circuits.^3,4 Nevertheless, when GCs fire, they do it in bursts with a higher-than-expected probability compared with other principal cells.² Although the mechanism that switches from regular to burst firing in GCs has not been elucidated experimentally, our computational results suggest that perisomatic disinhibition could be a key determinant. The possibility to switch between regular and burst firing is functionally relevant because the latter acts as detonators enhancing communication between DG and CA3.^{43,44,45,46,47} The CA3 response to GC bursts is different in pyramidal neurons than in interneurons.^48,49 While the response of the former is facilitated, that to the latter is unchanged or depressed, resulting in an increased E/I ratio in CA3. Regulation of burst firing in GCs may represent a mechanism to boost information transmission.

MCs are pivotal excitatory neurons within the DG that significantly contribute to the regulation of GC excitability and the overall E/I balance. MCs synapse with GABAergic interneurons, mediating feedforward inhibition onto GCs, thus playing a dual role in modulating excitatory and inhibitory signals within the DG circuitry.⁵⁰ Our computational findings align with these observations, indicating that MCs are essential for maintaining the proper E/I balance that underlies synaptic plasticity and reliable information transfer. By activating local inhibitory interneurons, MCs modulate GC excitability, contributing to the regulation of their firing thresholds. Following LTP, the functional state of the inhibitory network—shaped through interactions with MCs—establishes a higher firing set point for GCs, thereby promoting burst firing. Thus, MCs may play a key role in maintaining both the stability and adaptability of the hippocampal network function.

The cognitive consequences of the above mechanism might be reflected in the ability of individuals to differentiate between similar input patterns. DG has been repeatedly linked to pattern separation, for which sparse firing of GCs is considered the key.^7,51,52 The proposed role of pattern separation in the DG is to generate sufficiently distinct representations so that the attractor dynamics of the CA3 network do not merge similar inputs into a single pattern. It has been shown that when rodents are exposed to very similar environments, their ability to discriminate between them depends on the integrity of the DG.^53,54,55 Our results show that decreased inhibition in the DG may increase the overlap between the patterns represented by GCs and compromise pattern separation locally. However, we argue that, to be an efficient pattern separator, the orthogonalized output of the DG needs to reach the CA3 region with sufficient strength to push local dynamics away from established attractors. Our simulations support this notion by demonstrating that when pattern separation relies solely on the firing of a limited number of GCs, their transmission to CA3 is hindered. Conversely, some level of overlap in the DG can be compensated by the firing patterns represented by a higher proportion of GCs, especially those exhibiting bursting activity, thus ensuring effective transmission to CA3. In addition, the unique abundance of BC-mediated lateral inhibition in the DG (BC → GC), as opposed to reciprocal inhibition (GC ↔ BC) typical of the neocortex, would help maintain sparsity despite increased GC activity.⁶ Importantly, our ad hoc experiment demonstrated that decreasing PV inhibitory activity in the DG during memory encoding facilitates subsequent pattern separation during recall. Overall, our results indicate that perisomatic disinhibition is well suited to balance sparsity and effective transmission, facilitating pattern separation.

Taken together, we proved, by experimental and computational means, that synaptic potentiation in the PP connecting the entorhinal cortex with the DG not only facilitates excitatory glutamatergic activity in GCs but also specifically decreases perisomatic feedforward inhibition involving BCs. Our modeling work suggests important functional consequences, such us balancing firing sparseness and bursting, enhancing consistency of GC outputs, and, as a consequence, supporting pattern separation by improving information transmission between the MEC and CA3.

Limitations of the study

While our study provides significant new insights into the mechanisms of synaptic plasticity and inhibitory reorganization in the DG, several limitations should be acknowledged. First, as with all computational models, our simulations necessarily simplify the underlying biological complexity. We adopted a minimalistic and well-constrained modeling approach to capture the essential dynamics of the DG circuit, but some cellular diversity and connectivity details were not included. For example, in the circuit depicted in Figure 2A, we focused on the predominant neuronal populations in the hilar circuitry described in the literature. We modeled a single population of hilar interneurons (SOM + cells encompassing HICAPs), which project near the soma and dendritic territories, but did not explicitly include HIPP cells, which target distal dendrites not represented in the model. Despite these simplifications, we believe that the proposed model is well justified and sufficiently robust to capture the key mechanisms under study. Indeed, the model successfully reproduces the experimental power spectra and coherence functions reported by Pernía-Andrade et al.²

Similarly, in the circuit shown in Figure 4A, we explicitly modeled only the GC and BC populations, while representing other populations as synthetic inputs. This simplification accounted for the balance of inputs described in Figure 2, but it also represents a limitation that could be addressed in future, more detailed implementations. Nonetheless, this reductionist strategy enabled us to isolate and investigate specific mechanisms within a controlled and interpretable framework.

Second, our experimental paradigm used electrical induction of LTP in male rats as a proxy for natural learning processes. Although electrical LTP is a widely accepted and highly reproducible model for studying synaptic plasticity, it does not fully recapitulate the complexity of physiological learning and memory formation. Nevertheless, the use of LTP provides a powerful tool to dissect fundamental principles underlying circuit reorganization.

Third, the in vivo recordings were performed under anesthesia, which may alter some aspects of network dynamics. Moreover, field potential recordings were fundamental in guiding our investigation, but they also present certain limitations. In particular, CSD analysis in our study could not discriminate between interneuron subtypes that reduce their firing in response to LTP; it could only localize inhibition to the perisomatic region. Together, the combined use of computational and experimental approaches in this study offers complementary perspectives, allowing us to test hypotheses that would be otherwise inaccessible and providing a robust foundation for future investigations in more naturalistic conditions.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Santiago Canals (scanals@umh.es).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •Data: All original data required to reproduce the main results have been deposited at Zenodo (rat data: https://zenodo.org/records/17990190; mice data: https://zenodo.org/records/18015214) and are publicly available as of the date of publication. The link is listed in the key resources table.
• •Code: All required code to reproduce the computational models have been deposited at Zenodo (https://zenodo.org/records/17975096) and are publicly available as of the date of publication. The link is listed in the key resources table.
• •Additional information: Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

The authors acknowledge funding from the Spanish Ministerio de Ciencia e Innovación, Agencia Estatal de Investigación (PID2021-128158NB-C21, PID2021-128158NB-C22, PID2024-162400OB-C21, PID2024-162400OB-C22 , and MICIU/AEI/10.13039/501100011033), Programs for Centers of Excellence in R&D Severo Ochoa (CEX2021-001165-S MICIU/AEI/10.13039/501100011033) and María de Maeztu (CEX2021-001164-M MICIU/AEI/10.13039/501100011033). C.E. was funded by the Conselleria d’Innovació, Recerca I Turisme of the Government of the Balearic Islands and the European Social Fund (FPI/1900/2016).

Author contributions

Conceptualization, C.E., C.R.M., and S.C.; methodology, C.E., L.P.-C., J.M.C., E.A.-S., E.P.-M., and R.G.-H.; software, C.E.; validation, C.E., L.P.-C., J.M.C., and E.A.-S.; formal analysis, C.E., V.J.L.-M., J.M.C., and E.P.-M.; investigation, C.E., L.P.-C., J.M.C., E.P.-M., and R.G.-H.; resources, C.R.M. and S.C.; data curation, C.E., L.P.-C., and J.M.C.; writing – original draft, C.E., C.R.M., and S.C.; writing – review & editing, all authors; visualization, C.E., E.A.-S., and E.P.-M.; supervision, C.R.M. and S.C.; project administration, C.R.M. and S.C.; funding acquisition, C.R.M. and S.C. All authors have read and approved the final version of the manuscript.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used ChatGPT-5.1 for English editing. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rabbit monoclonal anti-Snail	Cell Signaling Technology	Cat#3879S; RRID:AB_2255011
Mouse monoclonal anti-Tubulin (clone DM1A)	Sigma-Aldrich	Cat#T9026; RRID:AB_477593
Bacterial and virus strains
pAAV-hSyn-DIO-hM3D(Gi)-mCherry	ETH Zurich	RRID:Addgene_44361
Chemicals, peptides, and recombinant proteins
Clozapine-N-Oxide (CNO)	ENZO Life Science	BML-NS105-0025
Urethane	Sigma Aldrich	U2500; CAS: 51-79-6
Gabazine	Tocris Bioscience	1262; CAS: 104104-50-9
CGP 52432	Tocris Bioscience	1246; CAS: 139667-74-6
D(−)-2-Amino-5-phosphonopentanoic acid (AP5)	Sigma Aldrich	A8054; CAS: 79055-68-8
Deposited data
Rat electrophysiological data Mice electrophysiological data	This paper This paper	https://zenodo.org/records/17990190 https://zenodo.org/records/18015214
Experimental models: Organisms/strains
Sprague-Dawley rats	Janvier	RRID:RGD_38676310
Mice: 129-Pvalbtm1(cre)Arbr/J	The Jackson Laboratory	RRID:IMSR_JAX:008069
Software and algorithms
Intel oneAPI Fortran Compiler	Intel	https://www.intel.com/content/www/us/en/developer/tools/oneapi/fortran-compiler.html
BRIAN2 Neural Simulator	Stimberg et al.56	DOI: https://doi.org/10.7554/eLife.47314; https://brian2.readthedocs.io/en/stable/
Computational models code.	This paper	https://zenodo.org/records/17975096
Spike2 Acquisition & Analysis Software	Cambridge Electronic Design (CED)	RRID:SCR_019152; https://ced.co.uk/products/spike2
EEGLAB Toolbox	Delorme and Makeig57	RRID:SCR_007292; https://sccn.ucsd.edu/eeglab/
ICA-based LFP decomposition methods	Makarov et al.18; Herreras et al.58	DOI: https://doi.org/10.1007/s10827-009-0206-y; DOI: https://doi.org/10.1093/cercor/bht022
MVGB Granger Causality Toolbox	Barnett and Seth59	DOI: https://doi.org/10.1016/j.jneumeth.2013.10.018
MC_Stimulus	Multi Channel Systems (MCS GmbH)	https://www.multichannelsystems.com
MC-Rack software	Multi Channel Systems (MCS GmbH)	RRID:SCR_014955https://www.multichannelsystems.com/software/mc-rack
Prisma 10	GraphPad Software	RRID:SCR_002798https://www.graphpad.com
Clampfit	Molecular Devices	RRID:SCR_011323https://www.moleculardevices.com/products/axon-patch-clamp-system/clampfi
Axo pClamp	Molecular Devices	RRID:SCR_011323https://www.moleculardevices.com/products/axon-patch-clamp-system/pclamp
MATLAB	MathWorks	RRID:SCR_001622; https://www.mathworks.com/

Experimental model and study participant details

Animals

Adult male Sprague-Dawley rats were used for in vivo LTP experiments, with a weight of 250-300 g. Adult male and female C57BL/6J mice originally from the line 129-Pvalb^tm1(cre)Arbr/J (Jackson Laboratories, RRID: IMSR_JAX:008069) were used for in vitro and in vivo pharmacogenetic experiments with DREADDs. All animals’ procedures were approved by the Animal Care and Use Committee of the Instituto de Neurociencias de Alicante (Alicante, Spain) and comply with the Spanish law (53/2013) and European regulations (EU directive 2010/63/EU).

Virus

Cre-Dependent AAVs DREADDs (AAV5-hSyn-DIO-hM4D(Gi)-mCherry; ETH Zurich) was selected to inhibit PV interneurons firing after Clozapine-N-Oxide (CNO) injection. DREADDs is based on the mutation of a G-protein receptor that cannot be activated by its usual endogenous ligand,⁶⁰ but by an exogenous agonist CNO (ref: BML-NS105-0025, ENZO Life Science Inc., New York, USA) which is inert otherwise.^61,62 CNO effect takes circa 25 minutes to get the peak and is sustained for several hours in vivo.^62,63,64 Correct functional efficacy of DREADDs was confirmed electrophysiologically prior to the recordings (Figure 3D). Coordinates for targeting the injections in the hilus of the DG, from Bregma, were -2mm AP, ± 1.4 mm LM, +2mm DV. We injected 0.5 μl of viral vectors per hemisphere using a micropipette attached to a pump infusion Nanoliter 2010 Injector (WPI) coupled to the stereotaxic frame (Sutter Instruments Company, California, USA).

Method details

Experimental and computational procedures

In vivo electrophysiology in rats

Rats were anaesthetized with 1.2–1.5 g/kg of urethane (Sigma-Aldrich, Missouri, USA) injected intraperitoneally. Supplemental doses (10% of the initial dose) were applied when required. After confirming the absence of reflexes, animals were placed in a stereotaxic frame (Narishige, Tokyo, Japan). During the experiment, the temperature was kept at 37°C, and blood oxygen saturation, heart and breathing rate monitored. After a subcutaneous dose injection of 8 mg/kg of local anaesthetic (Bupivacaine, Braun Medical SA, Barcelona, Spain), the scalp and periosteum were separated. The skull was opened with a manual drill (2 mm diameter, Fine Surgery Tools, USA) over the dorsal hippocampus (coordinates with respect to Bregma: A-P -3.5 mm, M-L 2.6 mm, 3.2-3.5 mm ventral to the dural surface) and the medial Perforant Pathway (PP; coordinates with respect to lambda: A-P 0 mm, M-L 4.1 mm, 2.3–2.7 mm ventral to the dural surface, with an angle of 15° at the sagittal plane directing the tip to rostral).⁶⁵ The dura was carefully punctured at the craniotomies with a needle, making the smallest hole possible to facilitate the penetration of both electrodes.

For orthodromic stimulation of the dentate gyrus, a tungsten bipolar electrode (10-15 kΩ, 325 μm diameter, World Precision Instruments, Florida, USA) was positioned in the medial PP. A multielectrode silicon probe (32 recording sites, 100 μm inter-site distance, 413 μm² electrode area, Neuronexus Technologies, Michigan, USA) was placed at the dorsal hippocampus to record the LFP. An Ag/AgCl wire (World Precision Instruments, Florida, USA) electrode was placed in contact with the skin bounded surgery area and used as ground. We found the accurate position of both electrodes using as a reference the control evoked potentials at the dentate gyrus⁶⁶ obtaining maximal population spike in the dentate gyrus. The brain electrophysiological signals were filtered (high-pass 0.1 Hz), amplified and digitalized (20 kHz acquisition rate) (Multi Channel Systems, Reutlingen, Germany), and stored for posterior analysis.

In vivo electrophysiology in mice

Mice were anaesthetized (1.4 g/kg urethane, i.p.: Sigma Aldrich, Madrid, Spain) and placed over a heat pad in a stereotaxic frame (Narishigue, Tokyo, Japan). All physiological parameters were continuously monitored. Following general surgery procedures, two 1.8 mm Ø trepan were done in the skull with a milling cutter (ref: FST 18004-18, Fine Science Tools, FST, Heidelberg, Germany) attached to a cordless microdrill (Stoelting Co., Illinois, USA), in the coordinates needed to introduce the electrodes (see below). One bipolar stimulating electrode (10-15 kΩ, 325 μm Ø, WPI, ref. TM53CCNON) was gently targeted to perforant pathway (from bregma: -4.3 AP, +2.5 ML, +1.4 DV, 12° angle); and one recording probe (single shank, 50 μm contact spacing, 32 channels; NeuroNexus, Technologies, Michigan, USA) was placed in the hippocampus (from bregma: -2 AP, +1.5 ML, -2 DV), in all CA1 and dentate gyrus layers. The final position was optimized based on the online recordings using the typical Population Spike (PS) evoked potential in the dentate gyrus after perforant pathway stimulation.⁶⁶ Stimulating electrode was connected to a pulse generator and current source (STG2004, Multichannel Systems, Reutlingen, Germany), controlled by MC_Stimulus software (Multichannel Systems). Electrophysiological data from recording probes were amplified, digitalized and filtered (0.1-3 kHz) using MC-Rack software (Multichannel Systems), and analyzed off-line.

Once the electrodes’ position was optimised, the tissue was allowed to rest for at least 30 minutes prior to the recordings’ start. Its stabilization was confirmed online by checking the stability of the population spike evoked potential in dentate gyrus after perforant pathway stimulation. Two sets of electrophysiological data were obtained:

• 1.Evoked population spikes (PS): PSs (32 kHz sampling rate, 100 ms window) were recorded during perforant pathway stimulation before and after CNO injection, using a fixed intensity adjusted to evoke 80% of maximal PS amplitude. One pulse was delivered every 30 seconds for 5 minutes. Evoked PSs were analyzed using Spike2 software (Cambridge Electronic, Cambridge, UK) and averaged per animal before and after CNO injection. These differences confirmed proper DREADD activation (Figure 3D)
• 2.Continuous spontaneous activity: Continuous activity across CA1 and dentate gyrus was recorded for 5 minutes (20 kHz sampling rate). All recordings were performed in the same animals before and 1 h after CNO injection (1 mg/kg, i.p.).

In vivo pharmacological experiments

A borosilicate glass pipette (World Precision Instruments, Florida, USA) was used to deliver pharmacological agents (dissolved in artificial cerebrospinal fluid, aCSF) into the dentate gyrus (DG). The pipettes were equipped with Ag/AgCl electrodes to guide accurate implantation, which was confirmed by stimulation of the perforant path (PP). The pipette was inserted into the hilus, positioned in close proximity to the recording electrode, and guided to the hilar layer of the DG using evoked potential recordings. To ensure precise placement near the recording probe, the pipette tip was bent at an angle of approximately 90°. Drug delivery was performed via air pressure pulses using a custom-built picospritzer, except for gabazine, which was administered via microiontophoresis. Control experiments using aCSF confirmed the absence of undesired electrophysiological changes due to volume injection under these conditions.

To inhibit GABA_A receptors, we used gabazine 1 mM (SR95531 hydrobromide, GABA_A-type receptor antagonist; Tocris Bioscience, Bristol, UK) and bicuculline 100 μM (bicuculline methiodide, GABA_A-type receptor antagonist; Sigma-Aldrich, Missouri, USA). To inhibit GABA_B receptors, we used CGP 1 mM (CGP 52432, GABA_B-type receptor antagonist; Tocris Bioscience). To inhibit NMDA receptors, we used AP5 30 mM (D(−)-2-amino-5-phosphonopentanoic acid; Sigma-Aldrich, Missouri, USA).

The effect of the administered drugs was evaluated by measuring the evoked potential in the DG in response to single-pulse stimulation of the PP at intensities subthreshold for eliciting population firing. This stimulation protocol facilitates the separation of pathway-specific evoked potentials.^18,19 The only exception was CGP, whose effect was quantified by analyzing the reversal of the paired-pulse protocol, in which the first pulse is suprathreshold and the second (delivered 150 ms later) is subthreshold. In this paradigm, the first pulse conditions the response to the second by inducing GABA_B-mediated presynaptic inhibition of GABAergic transmission.^67,68,69 This protocol was used to demonstrate the GABAergic nature of the evoked hilar potential (e-Hilar).

In vivo LTP protocol

In rats, LTP was regularly induced using a high-frequency stimulation protocol of the PP.⁷⁰ This tetanic stimulation consisted of six trains of pulses (400 Hz, lasting 20 ms), delivered at a 10 s interval, and repeated six times at an interval of 2 minutes. In addition, a theta-burst stimulation protocol,⁷¹ consisting of five trains of pulses (100 Hz, lasting 30 ms) delivered every 150 ms and repeated three times at an interval of 1 min, was tested, yielding the same results (Figure S5).

To evaluate the synaptic potentiation, we measured the population spike, (PS, defined as the amplitude from the precedent positive crest to the negative peak in the hilar evoked LFP), the excitatory postsynaptic potential (EPSP, defined as the maximal negative slope of the falling potential in the molecular layer evoked LFP), and the PS latency (defined as the delay between the stimulation artefact and the PS) at different stimulation intensities (the called input-output curve). We collected Input-Output curves before and after (30-60 minutes) the tetanizing protocol.

In vitro experiment

Mice were sacrificed and perfused through the ascending aorta with ice-cold N-Methyl D-glucamine (NMDG) buffer solution [concentrations (in mM): 92 N-Methyl D-Glucamine, 30 NaHCO₃, 1.25 Na₂PO₄, 2.5 KCl, 25 glucose, 0.5 CaCl₂, 10 MgCl₂, 20 NaHEPES, 5 Na Ascorbate, 3 Na Pyruvate, 2 Thiourea at pH 7.4; Sigma-Aldrich]. After perfusion, skull was carefully peeled to remove the whole brain. Then, injected right hemisphere were glued frontally on the slicing platform device. Tissue was sectioned into 300 μm thick 45 degrees coronal slices cutting from dorsal to ventral in ice-cold NMDG solution with the vibratome (VT1200S, Leica Biosystems Nussloch, Nussloch, Germany). Slices were then immediately placed into 37°C temperature, oxygenated (95% oxygen, 5% CO2) NMDG glass for 15 minutes. Afterwards, slices were transferred to another oxygenated room temperature glass with HEPES holding, artificialcerebrospinal solution (ACSF): [concentrations (inmM) 92 mM NaCl, 2.5 mM KCl, 1.25 mM NaH₂PO₄, 30 mM NaHCO₃, 20 mM HEPES, 25 mM glucose, 2 mM thiourea, 5 mM Na-ascorbate, 3 mM Na-pyruvate, 2 mM CaCl₂·4H₂O and 2 mM MgCl₂]. Here, slices were kept for at least 90 minutes until transferred to a submersion-type recording chamber perfused with (ACSF) solution [concentrations (inmM): 26 NaHCO₃, 124 NaCl, 3 KCl, 1.25 NaH₂PO₄, 20 glucose, 1 MgCl₂, 2 CaCl, 5 NaAscorbate, 3 NaPyruvate, 2 Thiourea pH 7.4; 300-310 mOsm: Sigma-Aldrich]. Slices were visualized with a Leica DM6000_FS-ModSys infrared microscope (Leica Microsystems, Wetzlar, Germany). During recordings (∼ 60min), slices were continuously perfused with room temperature oxygenated ACSF solution (2–2.5 ml/min).

Parvalbumin (PV+) neurons from the dentate gyrus were visually identified using a fluorescent light source and filter N2.1 with stimulation in green and emission in red. For assessment of electrophysiological properties and postsynaptic analysis, standard whole-cell patch-clamp recordings were performed. Electrodes were fabricated on a horizontal Flaming and Brown micropipette puller (model P-87, Sutter Instruments, Novato, CA, USA) and filled with filtered potassium gluconate-based internal solution for current clamp experiments [concentrations (in mM): 145 Potassium gluconate, 5 NaCl, 0.5 EGTA, 4 MgATP,0.3 NaGTP, 10 NaHEPES, 1% biocytin. At a final pH of 7.2 and 280 mOsm. Sigma-Aldrich]. For voltage clamp recordings of Excitatory Postsynaptic currents (EPSC), internal solution is [(in mM): 121.5 CsMeSO₃, 0.1 EGTA, 4 MgCl₂, 13.5 CsCl, 2 MgATP, 10 NaHEPES, 10 Na₂ phosphocreatine, 0.3 Na GTP. At a final pH of 7.2 and 279 mOsm. Sigma-Aldrich]. In ACSF, electrodes had resistances of 5–7 MΩ. Clozapine N-oxide from ref: BML-NS105-0025, ENZO Life Science Inc., New York, USA, 2 μM were added to ACSF for perfusion after recording in control conditions. Data were acquired with MultiClamp 700A (Axon Instruments, Union City, CA), filtered online at 4–5 kHz, and digitized with a digidata 1440A on a personal computer at a sampling rate of 10 kHz. Access resistance was monitored for the duration of each experiment.

Axo pClamp 10.7 software (Molecular Devices, San Jose, CA, USA) was used to analyze either the passive, active or postsynaptic events. Action potentials (current clamp mode): A series of 500 ms hyperpolarizing and depolarizing current pulses (5 pA steps, from -60 to +20 pA). IPSCs (voltage clamp mode): Spontaneous IPSCs (sIPSCs) were recorded for 20 min at holding potentials of 0 mV. Recordings was lowpass filtered at a 1000 Hz cutoff, using the clampfit software.

Behavioral experiments

Spatial memory was assessed using the Novel Object Location (NOL) paradigm. After habituation to handling, mice explored an empty square arena (50 × 50 × 30 cm) containing spatial cues under dim illumination (∼23 lux at the center) during two 5-min sessions.

Twenty-four hours later, animals underwent the familiarization (encoding) phase, time t0, during which two identical objects were placed in opposite corners (13.5 cm from the walls). Exploration continued until each mouse accumulated 20 s of total object exploration or reached a 10-min limit. Mice exploring less than 8 s in total were excluded according to predefined criteria.

After another 24 hours, mice were reintroduced to the arena for the test (retrieval) phase, with one object displaced (13.5 cm from the walls and 10 cm from the other object; time t1). Exploration was recorded under identical conditions. Following a 1-minute interval, the displaced object was moved an additional 5 cm (t2), reaching a total distance of 15 cm from the stationary object. This procedure was repeated twice more (t3 and t4) until the moved object reached a final distance of 25 cm from the static object.

Each animal completed both treatment conditions (CNO and vehicle) in a within-subjects design. Injections were administered 90 min before familiarization (encoding), and sessions were separated by one week. The discrimination index (DI) was calculated as the proportion of time spent exploring the displaced object relative to total object exploration.

To assess whether individual animals reliably discriminated the object displacement and at which trial this occurred, we calculated a detection threshold from the discrimination index (DI) measured during the encoding trial (t0), when the objects are presented for the first time. Thus, the DI distribution at t₀ represents baseline variability and provides an appropriate reference for identifying values that exceed chance-level exploration.

For each experimental condition, we calculated the mean and standard deviation (SD) of DI values at t₀ across subjects. An animal was classified as having detected the spatial change at a given test position (t1, t2, t3 and t4) when its DI exceeded two standard deviations above the corresponding t₀ mean This cutoff corresponds approximately to the upper tail of the baseline DI distribution and allows us to exclude DI values that fall within the variability expected at encoding.

Computational model of the dentate gyrus circuit

We used individual spiking neurons governed by Izhikevich equations to build the circuit.²⁵ By modifying the Izhikevich parameters (a, b, c, and d) different spiking regimes can be generated (see Table 1). Fast-spiking parameters were used for inhibitory interneurons and excitatory MCs. Under baseline conditions, firing rates in mossy cells can be slower. However, physiological studies show that mossy cells can readily fire above 50 Hz when strongly excited or disinhibited. In our model, applying LTP to GC—which directly excite MCs—combined with reduced inhibition from BCs leads to enhanced MC excitability, consistent with experimental findings.^72,73,74,75 Regular spiking neurons were used for the excitatory cells of the EC population and 35% of the granular population. However, 65% of GCs were represented by bursting neurons.²

We scaled the DG circuit 1:100 to match the anatomical number of GCs in adult rats, and then simulated 100 BCs, 100 Hil interneurons, and 300 MCs.⁷⁶ To reduce the computational cost, we only included the active cells in the granular group, which represented 5% of the considered population, corresponding to 500 cells in the scaled system.⁷⁷ Additionally, we simulated the EC population, which consisted of 200 neurons, with 20% inhibitory interneurons and 80% excitatory cells, to generate an input theta into the DG circuit.³³ The total number of cells in the simulation was 1200, allowing for a low computational cost necessary for the fitting process method.

Synaptic model

The synaptic current was modeled as an ohmic conductance (g_syn) multiplied by the ion channel kinetics (r(t)) and the driving force resulting from the difference in voltage between the membrane and reversal potential (E_syn) (Equation 1;⁷⁸). We assumed that the opened ionic channel is instantaneous, simulating only the decay exponential of the closing channel.

$$I_{syn}=g_{syn}r\left(t\right)\left(V_t-E_{syn}\right)\text{,}$$ (Equation 1)

$$\frac{dr}{dt}=-\frac{r}{{\tau }_{decay}}+\sum _k\delta \left(t{-t}_k\right)\text{.}$$

After each presynaptic spike at time t_k, the variable r instantaneously increases (simulating neurotransmitter release). Between spikes, r(t) decays exponentially with time constant τ_decay. Each synapse is thus characterized by the specific time constant (τ_decay), ohmic conductance (g_syn)and reversal potential (E_syn), whose values are indicated in Table 2.

Fitting procedure

We tuned the connectivity and the maximum peak of ion channel kinetics based on previous descriptions,^33,79,80 for each synapse to fit the model (see Tables 3 and 4). We relied on three experimental observations: the theta frequency range of the MEC region,^26,27,28 the gamma frequency range of the Hilus region,^29,30 and the excitatory and inhibitory inputs characterized by the theta and gamma frequency components, respectively, from the EC and Hilar regions into the granular population, matching the in-vivo intracellular recordings Jonas’ group.² The modifications of both parameters were made manually for each simulation, and we tested the three experimental observations with the power spectrum of the LFP and coherence between EPSC and IPSC of individual GC with the LFP of the granular population. The model also included a stochastic contribution where each neuron received a glutamatergic noise with a Poisson distribution. This approach mimics the random arrival of action potentials from upstream neuronal populations by generating a train of spikes, with inter-spike intervals drawn from an exponential distribution characterized by a mean rate (λ). The parameters of the Poisson inputs are summarized in Table 5.

Multi-compartmental model of granule cells

Model structure

The simulated circuit for the functional model included granule cells (GCs) and basket cells (BCs), which were adapted from published scripts.³³ BCs were implemented using a somatic one-compartment adaptive exponential integrate-and-fire (AdEx) model,⁸¹ which was fitted to produce a fast-spiking regime. In contrast, GCs were represented with three dendritic branches comprising a total of 21 passive compartments, each modeled using the LIF framework.⁸² In addition, a modified LIF model⁸² was applied to the somatic compartment, incorporating a second differential equation with an adaptive variable to capture action potential generation. Other DG cells were included in the model as synaptic inputs to the simulated neurons.

Synaptic mechanisms

The model includes both glutamatergic and GABAergic connections, with AMPA and NMDA receptors mediating the former and GABA_A receptors mediating the latter. The synaptic conductance was modelled to include both the rise and decay exponential phases that characterize neurotransmitter transmission. In the case of NMDA receptors, their conductance was also modulated by the voltage of the postsynaptic neuron, as these receptors are blocked by a positively charged magnesium ion that is sensitive to voltage changes. This effect was modelled using a sigmoidal function⁸³ that multiplied the synaptic conductance.

Input patterns and network connectivity

The GC received GABAergic inputs from HIPP and HICAP interneurons onto the distal and proximal dendrites, respectively, and perisomatic innervation from BCs. Glutamatergic inputs were focused on the proximal dendrite due to commissural AMPA synapses and the medial dendrite representing AMPA and NMDA connections from the MEC. These two glutamatergic sources affect the BCs as well, with an inhibitory contribution from Hilar interneurons. The dendritic inputs, except the MEC, were Poisson spike trains with a gamma firing rate. There were one HIPP and HICAP interneuron and five MCs per GC, maintaining the anatomical distribution of the DG. The BCs received the same MC input. The GABAergic input was constituted by several Poisson spike trains with a gamma firing rate that was modified to change the inhibitory input from the Hilar region. The MEC input in both neurons, GC and BC, was a regular input of 4 neurons with a theta frequency of 8 Hz.

Pattern separation analysis

The functional model focused on studying the information processing of an individual GC model and the pattern separation function in a granular population model. For the individual granule cell, a regular synchronized signal of 4 neurons with a theta frequency of 8 Hz was used as input from MEC in GC and BC. The consistency was evaluated by measuring the spike correlation of the granular output between each non-repetitive combination pair of interactions across 50 iterations. In the spatial codification simulation, each MEC neuron was assigned a firing probability with a theta frequency, generating a repeating pattern of two seconds for 30 seconds.

Mutual information (MI)

Mutual Information (MI)⁸⁴ was measured for two discrete binary variables - the spike activity of input from MEC and the GC output, represented by 1-spike and 0-no spike.

$$MI\left(X;Y\right)=\sum _{xϵ\left(0,1\right)}\sum _{yϵ\left(0,1\right)}P\left(x,y\right)\mathit{log}\left(\frac{P\left(x,y\right)}{P\left(x\right)P\left(y\right)}\right)$$ (Equation 2)

where, x and y are the discrete variables, P(x,y) the joint probability distribution, and P(x) and P(y) are the marginal probabilities.

Pattern Separation Efficiency (PSE)

For the Pattern Separation function, the circuit was scaled up to include 1000 GCs, with the number of neurons in each population determined based on the proportion found in the granular layer: 200 MEC neurons, 50 BCs, 40 MCs, 20 HIPPs, and 20 HICAPs. In this case, the MEC input consisted of a neural pattern of synchronized neurons, constituting 40% of the total MEC population and repeated at a theta frequency for 30 seconds. For further analysis, we selected the neurons that were activated in more than 80% of the input patterns. The Pattern Separation Efficiency (PSE) was calculated by subtracting the ratio of common neurons between a pair of two patterns from the maximum value of 1 (PSE_MAX).

$$PSE=PS{E}_{\mathit{MAX}}-\frac{C{N}_{ij}\left(A{ctN}_j+A{ctN}_i\right)}{2\ast Act{N}_j\ast Act{N}_i}\text{,}$$ (Equation 3)

where CN_ij is the number of common neurons in the pair of patterns i and j, and ActN is the number of active neurons in a given pattern.

Active neurons (ActN)

For the measurement of active neurons (ActN), we compute it as the ratio between the neurons activated during the input pattern (N_activated) and the total number of neurons (N_population). We consider a neuron as active if it is activated at least in the 80% of the input patterns generated in the 30 seconds of stimulus. The value showed in Figure 4 is the mean of AN between each pair of patterns.

$$ActN=\frac{N_{activated}}{N_{population}}$$ (Equation 4)

Probability of bursting (PB)

The probability of bursting (PB) is the ratio between number of burst (B) and total number of spikes, the sum of burst (B) and single (S) spikes, of an activated neuron. We count a burst if the neuron generates consecutive spikes with and inter-spike frequency equal or higher than 40Hz. The value showed in Figure 4 is the mean of PB between each pair of patterns.

$$PB=\frac{B}{S+B}$$ (Equation 5)

Pattern Separation Transmission (PST)

We also defined a Pattern Separation Transmission (PST) index as:

$$PST=PSE<PB>f\left(<ActN>\right)\text{.}$$ (Equation 6)

Where <X> means the mean value of X for the overlapped patterns. For the function f(ActN)(ActN≔<ActN> for simplicity) we considered two limits’ cases yielding zero transmission: AN=0 and population overexcitation threshold. A simple equation that allowed us to represent this effect is of the form:

$$f\left(ActN\right)=\{\begin{array}{c}ActN\left(\mu -ActN\right)forActN\in [0,\mu ]\\ 0forActN>\mu \end{array}$$ (Equation 7)

where the parameter μ controls the overexcitation of the system. Its value was determined by assuming that the number of active neurons (ActN) that maximizes f(ActN), i.e., that provides maximum information transmission, corresponds to the experimentally estimated number of active granule cells ActN=0.05 (5 %)^77,85 i.e., μ≅0.1.

Computational simulations

The DG circuit model was compiled and executed in FORTRAN90, and the functional model was simulated in BRIAN2⁵⁶ (Python 3.1), both computed in the Nuredduna computer system of the Institute of Cross-disciplinary Physics and Complex Sytems (IFISC) with an integration interval of 0.01 ms.

Quantification and statistical analysis

Data analysis

All analysis done in this work was performed with the software MATLAB (The MathWorks Inc., Massachusetts, USA) and Spike2 (Cambridge Electronic Design Ltd., Cambridge, UK).

Experimental data analysis

The in vivo experiments in rats analyzed LFP signals with minimal cortical activity and stable theta-band oscillation (4–12 Hz) in CA1. The Independent Components (IC) were obtained using the runica algorithm of the ICA method, which is part of the EEGLAB MATLAB toolbox.⁵⁷ The method has been developed and tested using brain signals and numerical models^18,19 and has been validated in previous reports.^{58,86,87,88,89,90,91,92}

The cross-correlation between spontaneous signals was calculated using a temporal window of 500 ms, and the power of the signal was computed as the square of the signal. The sub-threshold amplitude is the difference between the maximum potential peak and the baseline voltage in the evoked potential data, and the latency was measured as the time between the stimulation artefact and the peak of the evoked potential.

Computational data analysis

Granger Causality is a statistical method used to identify the directionality of causation between two time series by comparing the ability of past values of one series to predict future values of the other series, with and without the inclusion of past values of the second series. In the computational circuit, the excitatory and inhibitory currents over the GCs were normalized and analyzed as spontaneous in vivo recordings.

The total signal was discretized in temporal windows with high overlap between them to improve the accuracy of the analysis for coherence analysis. The MVGC Multivariate Granger Causality MATLAB Toolbox⁵⁹ provides a set of tools for Granger causality analysis of multivariate time series. In the computational circuit described, the Granger causality was used to determine the directionality of causation between the average voltage of each population of cells.

Statistics

After identifying and removing outlier values, normality distribution of data was checked using the D’Agostino–Pearson and Shapiro–Wilk tests. Depending on the distribution of the samples (Gaussian or non-Gaussian), different statistical tests were used.

For comparisons of one or two samples (one when compared to a hypothetical value), unpaired and paired t-tests were used for Gaussian samples, and Mann–Whitney and Wilcoxon signed-rank tests were used for non-Gaussian samples, respectively.

When comparing three or more samples, one-way ANOVA and Kruskal–Wallis tests were used for unpaired Gaussian and non-Gaussian samples, respectively, while repeated-measures ANOVA and Friedman tests were used for paired Gaussian and non-Gaussian samples, respectively.

For comparing two variables and their interaction, repeated-measures two-way ANOVA was used for paired Gaussian samples. All statistical analysis was conducted using Prism 10 software (GraphPad Software, Inc., California, USA) and MATLAB (The MathWorks Inc., Massachusetts, USA).

To test whether the Granger Causality (Figure 1), power spectrum or coherence (Figure 3) were different between conditions (control vs. LTP or control vs. CNO), we used a permutation test followed by a cluster-based correction for multiple comparisons.

For the Granger Causality example, we first estimated the Granger Causality for each animal and computed a t-test between “control” and “LTP” conditions across subjects. We kept the t-values of those points with a p-value lower than 0.05 as uncorrected significance values.

To correct for multiple comparisons, we selected clusters of consecutive frequency-points with a significant p-value. Each cluster had an associated t-value, corresponding to the sum of all the t-values within it. Then, we randomly permuted the label associated with each Granger Causality result (i.e., we randomly labeled them as “control” or “LTP”) and followed the same procedure to select the cluster with highest absolute t-value.

We repeated this permutation test 100 times, obtaining 100 permuted t-values. Finally, we tested whether the original t-value was higher than at least 95 of the permuted t-values, corresponding to a p-value of 0.05.⁹³ All statistical details for each experiment, including the statistical tests used, exact n values, definition of n (animals, recordings, cells, simulations), test statistics (t, F, U, χ² values together with degrees of freedom), exact p values, effect sizes when applicable, and the definition of center and dispersion (mean ± SEM unless otherwise stated), are provided in the corresponding figure legends. Asterisks in the figures indicate statistical significance (∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.001).

Published: February 2, 2026