Sleep deprivation impairs ventral hippocampus-medial prefrontal connectivity and task learning in rats

Summary

Sleep is essential for memory consolidation, and sleep loss has long been associated with cognitive impairment. However, the circuit-level mechanisms through which sleep deprivation disrupts hippocampal-prefrontal communication during learning remain poorly understood. Here, we combined a T-maze learning task with simultaneous local field potential (LFP) recordings in the ventral hippocampus (vHPC) and medial prefrontal cortex (mPFC) of rats. Using dynamic causal modeling (DCM), we quantified directional connectivity during task acquisition under normal and sleep-deprived conditions. We found that learning strengthened forward connectivity from vHPC to mPFC while weakening backward connectivity from mPFC to vHPC. Sleep deprivation impaired this forward connectivity, resulting in slower learning and reduced task accuracy. These findings characterize the altered circuit dynamics by which sleep loss disrupts hippocampal-prefrontal circuit dynamics, providing insight into how impaired cross-regional communication may underlie cognitive deficits observed in sleep-related psychiatric and neurodegenerative disorders.

Graphical abstract

Highlights

• •Dynamic causal modeling revealed directed vHPC-mPFC interactions during T-maze learning
• •Learning enhanced forward vHPC-to-mPFC connectivity while weakening backward coupling
• •Sleep deprivation selectively disrupted forward hippocampal drive and delayed task acquisition
• •These findings identify a circuit mechanism linking sleep loss to cognitive dysfunction

Neuroanatomy; Cognitive neuroscience

Introduction

Learning and memory are fundamental cognitive processes that rely on the coordinated interaction of distributed brain networks. Substantial evidence indicates that sleep plays a critical role in supporting these functions, contributing to memory consolidation and synaptic homeostasis.^1,2,3,4,5,6 Conversely, sleep deprivation (SD) severely disrupts cognitive performance, leading to significant impairments in learning and memory.^7,8,9 Understanding the neural mechanisms underlying these SD-induced deficits remains a central goal in neuroscience.

The hippocampal-prefrontal circuit constitutes a fundamental pathway for cognitive control and memory consolidation.^10,11,12 Functionally, the ventral hippocampus (vHPC) routes spatial and contextual information to the medial prefrontal cortex (mPFC), which integrates these inputs to support rule-based decision-making and working memory.^13,14,15 This bidirectional communication is highly state-dependent. During active task engagement, coherent theta oscillations (4–12 Hz) synchronize the two regions, with the vHPC typically leading the mPFC, thereby providing a “feedforward” drive essential for encoding task-relevant information.^16,17,18 Conversely, during sleep, this circuit undergoes a critical shift: Hippocampal sharp-wave ripples coordinate with cortical spindles to facilitate the transfer of labile memory traces from the hippocampus to neocortical sites for long-term storage, a process known as systems consolidation.^19,20 Consequently, SD may disrupt these precise temporal dynamics, impairing the circuit’s ability to stabilize and retrieve memory representations. Importantly, SD has been shown to alter neuronal activity patterns in both the HPC and PFC, suggesting that functional disruption within this circuit may be a key contributor to cognitive impairment.^1,21,22

While local field potentials (LFPs) provide a high-resolution measure of regional brain activity, elucidating the directed information flow between regions requires analytical approaches that can infer causality. Effective connectivity offers a framework to characterize these causal interactions, serving as a fundamental mechanism for inter-regional communication.²³ Among available methods, dynamic causal modeling (DCM) is a powerful, model-driven approach for quantifying effective connectivity from neurophysiological signals.²⁴ It has been successfully applied to data ranging from non-invasive human recordings^{12,25,26,27,28,29,30} to intracranial recordings in animals,^23,31,32,33 making it ideally suited to dissect circuit mechanisms underlying behavior.

Despite the established roles of the vHPC and mPFC in learning and memory, the specific causal influence of SD on the dynamic interplay within this network remains poorly understood. Current evidence largely describes correlational changes, leaving a gap in our knowledge of how SD alters the directed, causal flow of information that is necessary for optimal cognitive function.

While anatomical projections and synchronized oscillations between the ventral hippocampus (vHPC) and medial prefrontal cortex (mPFC) are well-documented, these static or correlational measures do not explain how information flows directionally during the specific phase of memory acquisition. Crucially, it remains unknown whether SD impairs performance by globally suppressing the circuit or by selectively disrupting specific directional pathways (e.g., bottom-up vs. top-down). To address this gap, we combined continuous LFP recordings from the vHPC and mPFC with DCM in rats performing a T-maze task under control or sleep-deprived conditions. By moving beyond correlation to quantify causal dynamics, we aimed to test the hypothesis that SD impairs learning by selectively disrupting the effective connectivity within this network. Our results reveal that SD preferentially compromises the forward (from vHPC to mPFC) pathway during learning acquisition, characterizing the altered circuit dynamics by which sleep loss leads to cognitive deficits.

Results

Histological verification of electrode placement in vHPC and mPFC

Extracellular neurophysiological recordings were obtained from vHPC and mPFC (Figure S1B). Representative histological examinations of lesion sites are depicted in Figure S1C.

Sleep deprivation delays the acquisition of the T-maze task

To investigate the effects of sleep deprivation on learning-related hippocampal–prefrontal communication, rats were assigned to either a control group or a sleep deprivation (SD) group and trained in a T-maze task according to the experimental timeline shown in Figure 1. Rats were trained in a delayed alternation T-maze task consisting of a free-choice phase, a delay period, and an alternate-choice phase (Figure 2). We assessed the accuracy of the control and SD groups during task acquisition (Figure 3A). This indicates that the behavioral correctness of rats increases as the learning stage progresses. The control group reached the acquisition criterion faster than the SD group (control: 6.4 ± 0.48 days; SD: 9.3 ± 0.42 days; independent-samples t test, t (12) = 6.32, p < 0.001; Cohen’s d = 3.38). (Figure 3B).

Figure 1.

Experimental timeline and daily protocol

(A) Overall experimental timeline. Rats underwent electrode implantation on day 0, followed by 5–7 days of recovery and 3 days of handling acclimation (30 min/day). They were then trained in the T-maze task for 12 consecutive days with simultaneous LFP recordings. During this training period, rats in the sleep deprivation (SD) group received 8 h of gentle-handling sleep deprivation per day during the light phase (ZT0–ZT8; 08:00–16:00).

(B) Daily schedule during the 12 training days. Control rats remained undisturbed in their home cages from ZT0 to ZT8, whereas SD rats underwent gentle-handling sleep deprivation during the same period. Immediately after this interval (∼ZT8), both groups performed the daily T-maze session at the same time of day with simultaneous LFP recordings from the vHPC and mPFC, after which they were returned to their home cages.

Figure 2.

Schematic diagram of the T-maze task

(A) Free choice.

(B) Delay.

(C) Alternate choice; n = 14.

Figure 3.

Results of behavioral statistics

(A) Daily accuracy of T-maze task learning for both groups of rats.

(B) Days to reach the acquisition criterion; control rats learned faster than SD rats (independent-samples t test, t (12) = 6.32, p < 0.001; Cohen’s d = 3.38); ∗∗∗p < 0.001; n = 14; Error bars refer to SEM.

Figure 4.

Typical microcircuit model (cmc) for dcms (spm_fx_cmm_nmda)

Learning modulates bidirectional effective connectivity between vHPC and mPFC

To estimate directed interactions between the vHPC and mPFC, we applied dynamic causal modeling using a canonical microcircuit model incorporating NMDA receptor-mediated synaptic dynamics (Figure 4). Figure 5 illustrates the connection model and Bayesian model selection outcomes for the ventral hippocampus (vHPC) and medial prefrontal cortex (mPFC). Based on existing literature, six potential cases of bidirectional connectivity between the ipsilateral mPFC and vHPC are recognized (Figure 5A). In our study, Bayesian inversion was initially conducted to optimize the selection of these models, evaluating the relative logarithmic evidence based on the LFP data from two groups totaling 12 rats performing the T-maze task pre- and post-acquisition. Bayesian model selection results indicated that Model 5 had the highest relative log model evidence value (248), suggesting it is the optimal connection model between vHPC and mPFC (Figure 5B). This model involves a forward connection from vHPC to mPFC and a backward connection from mPFC to vHPC. Furthermore, posterior probability analysis confirmed the credibility of Model 5, with a posterior probability exceeding 0.99. Therefore, potential scenarios regarding the influence of pre- and post-connectivity on rat learning were further analyzed based on Model 5.

Figure 5.

Depiction of the several types of between-source connections and the Bayesian model comparison result

(A) Diagrams of 6 possible vHPC-mPFC connection models: Model 1: Bidirectional transverse connections. Model 2: Backward connections from mPFC to vHPC. Model 3: Forward connections from mPFC to vHPC and backward connections from vHPC to mPFC. Model 4: Forward connections from vHPC to mPFC. Model 5: Forward connections from vHPC to mPFC and backward connections from mPFC to vHPC. Model 6: Bidirectional forward and backward connections between the two brain regions.

(B) Bayesian model comparison results (model 1 as the standard comparison model). Model 5, the winning model, has the largest relative log evidence value (248).

(C) Diagrams of 3 possible modulation models based on Model 5. The green line indicates the inter-brain area connections modulated by learning. FCM, only forward connection from vHPC to mPFC is modulated by learning; BCM, only backward connection from mPFC to vHPC is modulated by learning; BiCM, bidirectional forward connection and backward connection between mPFC and vHPC are both modulated.

(D) Bayesian model comparison results (FCM as the standard comparison model). BiCM has the largest relative log evidence value (2.5), which indicates both forward and backward connections between mPFC and vHPC are modulated by learning; n = 12.

Based on the chosen optimal connectivity model, there are three possible scenarios for the effect of learning this factor on the connectivity between the vHPC and the mPFC (Figure 5C). Similarly, Bayesian inversion and relative logarithmic evidence of the models were applied for the optimal selection of the aforementioned models. Connection Modulation Model Results: The horizontal axis of the graph represents the results of modulation model analysis for forward connection affected, backward connection affected, and bidirectional connection affected, respectively, while the vertical axis denotes the relative logarithmic model evidence value. It can be observed from the figure that the model where both bidirectional connections are affected has the largest relative log model evidence value (2.5) and is deemed as the optimal modulation model (Figure 5D). Further analysis of its confidence level revealed that the model had a high confidence level (98%) with a posterior probability of 0.98. These findings suggest that learning a memory task can simultaneously impact the bidirectional connectivity between vHPC and mPFC.

Successful learning is characterized by strengthened forward and weakened backward connectivity

Given that Model 3 was selected based on frequency range responses, we proceeded to test for the co-moderating effects of learning. In other words, we looked for parameters that positively or negatively modulated the effect. The two moderating parameters examined included: deep prefrontal pyramidal cells driving the extrinsic connectivity of vHPC excitatory receptors, and deep hippocampal pyramidal cells driving the extrinsic connectivity of mPFC excitatory receptors. To assess the significance of each modulatory effect, we utilized Bayesian parameter averaging for the posterior densities of both bands. Hereafter, we report effects with a confidence level of 95% or higher. We initially analyzed changes in connection strength between pre- and post-acquisition in control rats. These data encompassed trials where the task was completed correctly and incorrectly while maintaining consistent proportions and behavioral correctness. The results indicated that forward connections strengthened by 180% and backward connections weakened by 23% with a posterior probability exceeding 0.99 post-acquisition compared to pre-acquisition (Figure 6A). The results in Figure 6A encompass tasks completed correctly and incorrectly by the rats. To further investigate the effects of correctly and incorrectly performed tasks on the results in Figure 6A. We analyzed the changes in connection strength when rats performed the memory task correctly and incorrectly, respectively. The results show that when rats completed the task correctly, the strength of the forward connection after learning increased to 230% of the pre-acquisition level, while the strength of the backward connection decreased to 52% of the pre-acquisition level, with a posterior probability exceeding 0.99 (Figure 6B); when rats completed the task incorrectly, the strength of the forward connection and backward connection after learning decreased to 82% and 59% of the pre-acquisition level, respectively, with a posterior probability exceeding 0.99 (Figure 6C).

Figure 6.

Comparisons of connection strength between vHPC and mPFC in control and SD rats

(A) Comparison of learning-affected connection strength between vHPC and mPFC in control rats.

(B) Comparison of learning-affected connection strengths between vHPC and mPFC when the task was performed correctly in control rats.

(C) Comparison of learning-affected connection strengths between vHPC and mPFC when the task was performed incorrectly in control rats.

(D) Comparison of learning-affected connection strength between vHPC and mPFC in SD rats.

(E) Comparison of learning-affected connection strengths between vHPC and mPFC when the task was performed correctly in SD rats.

(F) Comparison of learning-affected connection strengths between vHPC and mPFC when the task was performed incorrectly in SD rats; n = 12.

Subsequently, we analyzed changes in effective connection strength between brain regions in the sleep-deprived group of rats. As depicted in Figure 6A, the data in Figure 6D consisted of trials where the task was completed correctly and incorrectly, with consistent proportions and behavioral correctness. Changes in connection strength pre- and post-acquisition in rats in the sleep-deprived group differed from those in the control rats. The results showed that forward connections were enhanced by 39% and backward connections were weakened by 38% with a posterior probability of 98% post-acquisition compared to pre-acquisition (Figure 6D). When the rats completed the task correctly, the strength of forward connections post-acquisition was enhanced to 102% of that pre-acquisition, and the strength of backward connections was weakened to 18% of that pre-acquisition, with a posterior probability exceeding 0.99 (Figure 6E); when the rats completed the task incorrectly, the strength of forward connections post-acquisition was weakened to 23%, and the strength of backward connections was enhanced to 386% of that pre-acquisition, with a posterior probability of 97% (Figure 6F).

Sleep deprivation disrupts the learning-induced enhancement of forward connectivity

To quantify the parametric effects of SD on learning and memory, we directly compared the effects of SD on the strength of connections between vHPC and mPFC. The results showed that in the pre-acquisition phase of rats, the forward connection strength of the SD group was enhanced to 166% of the control group, and the backward connection strength was weakened to 17% of the control group, with a posterior probability exceeding 0.99 (Figure 7A). In the post-acquisition phase of the rats, the forward connection strength of the SD group decreased to 83% of the control group, and the backward connection strength decreased to 17% of the control group, with a posterior probability exceeding 0.99 (Figure 7B).

Figure 7.

Comparisons of connection strength between vHPC and mPFC in two groups

(A) Comparison of learning-affected connection strengths between vHPC and mPFC when rats perform a task pre-acquisition.

(B) Comparison of learning-affected connection strengths between vHPC and mPFC when rats perform a task post-acquisition; n = 12.

Discussion

The brain’s capacity for functional integration depends on sophisticated interactions among specialized regions that communicate through directed pathways.³⁴ Our study advances current understanding by determining the directional nature of the circuit dysfunction induced by sleep loss. While previous studies have established the necessity of vHPC-mPFC coherence for learning, we provide novel evidence that SD does not uniformly inhibit the network; rather, it selectively abolishes the learning-dependent strengthening of the forward (vHPC to mPFC) drive. This finding offers a precise circuit-level mechanism: SD prevents the prefrontal cortex from effectively integrating spatial/contextual inputs from the hippocampus, which is critical for decision-making.

Our model comparisons identified a bidirectional vHPC-mPFC architecture as the most plausible network structure, with a forward connection from vHPC to mPFC and a backward connection in the reverse direction. This finding aligns with earlier anatomical and electrophysiological studies demonstrating reciprocal projections between these regions.^15,23,35,36 Furthermore, we found that learning modulates both of these connections simultaneously, underscoring the integrated nature of this circuit in memory processes.^37,38

The learning-related enhancement of forward connectivity observed in control rats is consistent with the putative role of this pathway in relaying spatial and contextual information to the PFC for executive control and decision-making.^13,20,39,40 This enhanced effective connectivity may correspond to functional correlates of activity-dependent synaptic plasticity, potentially involving NMDA receptor-mediated mechanisms.^41,42 Conversely, the concomitant weakening of the backward pathway may be functionally related to network selection or active forgetting, whereby top-down control from the PFC suppresses irrelevant hippocampal activity to prioritize salient information.^15,43 This is supported by evidence that prefrontal projections to the hippocampus can inhibit neuronal activity and gate memory retrieval.

Critically, SD disrupted this finely tuned balance. SD rats exhibited a profound impairment in the strengthening of the forward connection during learning. This deficit provides a compelling circuit-level explanation for their delayed behavioral acquisition. The persistence of weaker forward connectivity suggests that SD may prevent the functional coupling necessary for efficient communication from the hippocampus to the prefrontal cortex. The observed disruption in forward effective connectivity aligns with established SD-induced alterations in glutamatergic transmission. Previous studies have shown that SD downregulates the expression of AMPA (GluA1) and NMDA receptor subunits in the hippocampus,^44,45 which likely serves as the molecular basis for the functional disconnection we modeled. Furthermore, sleep is known to promote synaptic downscaling and renormalization^5,46,47; SD may therefore disrupt this process, potentially leaving circuits in a saturated state that may be less capable of efficiently encoding new information by strengthening relevant pathways.^5,6

The differential effect of SD on forward versus backward connections is particularly intriguing. It suggests that SD does not uniformly suppress all neural activity but rather disrupts specific, behaviorally relevant information streams. This specificity aligns with emerging views that SD impairs cognitive function not through global brain inhibition but through a breakdown in the precise temporal coordination and information transfer between key brain regions.^4,7,48

From a translational perspective, impaired hippocampal-prefrontal communication may represent a common mechanism linking sleep disturbances to psychiatric and neurodegenerative disorders, including depression, Alzheimer’s disease, and schizophrenia. By pinpointing forward connectivity as a critical substrate, our results raise the possibility that interventions enhancing hippocampal drive to the prefrontal cortex could mitigate sleep-related cognitive dysfunction.

We acknowledge that SD can introduce confounding factors such as physical fatigue and stress. However, several aspects of our study design mitigate these concerns. First, we employed the “gentle handling” method, which has been shown to induce significantly lower HPA-axis activation compared to forced deprivation techniques.⁴⁹ Second, regarding motor fatigue, it is important to note that all SD rats successfully completed the T-maze trials and made navigational choices, albeit with lower accuracy. If physical exhaustion were the primary driver, we would expect a failure to traverse the maze or prolonged immobility, rather than the specific decision-making errors observed. Therefore, while we cannot fully rule out minor contributions from fatigue, the deficit manifests primarily as a failure of spatial working memory (a vHPC-mPFC function) rather than a motor incapacity.

In conclusion, our study demonstrates that SD induces a selective deficit in hippocampus-prefrontal forward connectivity during learning. By applying computational modeling to the neurophysiological data, we move beyond correlation to characterize the directed effective connectivity associated with SD-induced cognitive impairment. These findings not only elucidate a novel mechanism by which SD disrupts memory but also highlight the primacy of directed information flow from the vHPC to the mPFC in supporting learning and adaptive behavior.

Limitations of the study

Several limitations should be considered. First, while our sample size is consistent with similar electrophysiological studies, future work would benefit from a larger cohort to enhance generalizability. Second, sleep/wake states during the SD window were not quantified by polysomnography; thus, brief microsleeps cannot be ruled out. Third, while DCM estimates parameters of synaptic coupling based on biophysical neural mass models, it represents effective connectivity—the causal influence that one neuronal system exerts over another—rather than a direct measurement of physical synaptic weights. Consequently, the changes in forward and backward connectivity observed here reflect functional modulation inferred from LFP dynamics. Although these findings provide a compelling circuit-level mechanism, future studies employing intracellular recordings or optogenetic manipulations are needed to directly verify the specific synaptic and cellular alterations underlying these inferred connectivity changes. Finally, our study focused on a specific frequency range (4–100 Hz); future studies could explore the contributions of specific frequency bands (e.g., theta and gamma) to these causal interactions.^35,37

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Duyan Geng (dygeng@hebut.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •All data reported in this article are available upon request from the lead contact.
• •This article does not report original code.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (grant no. 52277230), the Science and Technology Cooperation Special Project of Shijiazhuang (grant no. SJZZXB24006), the Hebei Province Provincial Science and Technology Programme Project International Science and Technology Co-operation Special Project (grant no. 25292001D), and the National Key R&D Program of China (grant no. 2022YFC2402203).

Author contributions

Conceptualization, W.Z. and D.G.; methodology, W.Z. and D.G.; investigation, W.Z.; formal analysis, W.Z.; writing – original draft, W.Z.; writing – review and editing, W.Z. and D.G.; supervision, D.G.

Declaration of interests

The authors declare no conflict of interest.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, peptides, and recombinant proteins
Isoflurane	RWD Life Science	N/A
Experimental models: Organisms/strains
Adult male Wistar rats	Institute of Zoology, Chinese Academy of Sciences	N/A
Software and algorithms
SPM12	Functional Imaging Laboratory, UCL Queen Square Institute of Neurology, London, UK	https://www.fil.ion.ucl.ac.uk/spm/software/spm12/
MATLAB	MathWorks	R2021b
CinePlex Editor V3	Plexon Inc	Version 3
Other
OmniPlex/128 neurophysiological signal acquisition system	Plexon Inc	OmniPlex/128
Nichrome electrodes (35 μm)	Custom-made	N/A
Stereotactic frame	RWD Life Science	N/A
Overhead video tracking system	Plexon Inc	CinePlex system
Vibrating microtome	Leica Microsystems	VT1200S
Optical microscope	Olympus	BX51WI

Experimental model and study participant details

Animals

Adult male Wistar rats, 10-12 weeks old and weighing 280-320 g (n = 18), were acquired from the Institute of Zoology, Chinese Academy of Sciences. These rats were kept on a 12-hour light/dark cycle (lights on at 08:00, defined as Zeitgeber time 0 [ZT0]) and had free access to food and water until food restriction for behavioral experiments. The rats were randomly assigned to different experimental groups. To ensure impartial results, the experimenters were blind to the group assignments when the treatment conditions were the same. All research protocols followed the National Institutes of Health (NIH) guidelines and were approved by the Biomedical Ethics Committee of Hebei University of Technology, with the license number HEBUTaCUC2022062.

Method details

Surgical procedures

Rats were anesthetized with 5% isoflurane in a flow-through chamber until sedation was reached. Subsequently, they were placed in a stereotactic frame, where the head fur, scalp, and subcutaneous connective tissue were sequentially removed to expose the skull. Using classic bregma and lambda landmarks, the skull was then flattened. Nichrome electrodes (diameter: 35 μm) were implanted into specific sites (Figure S1A): the mPFC (2.5–4.5 mm anterior to bregma, 0.2–1.0 mm lateral to midline, 2.6 mm depth) and the vHPC (4.6–6.0 mm posterior to bregma, 4.4–5.4 mm lateral to midline, 7.7 mm depth). Additionally, a grounding screw and a reference screw were inserted into the skull. Dental cement was then applied to secure the electrodes in position. Throughout the procedure, isoflurane concentration was maintained at 1.5%, and heating pads were used to keep the rats warm. Post-surgery, the rats were allowed to recover for 5-7 days until they regained their pre-surgery weight.

Sleep deprivation model preparation

To acclimate the rats to the experimenter, each of the eighteen rats underwent a 30-minute handling session for the first three days. Nine rats were randomly designated as the sleep deprivation (SD) group. Subsequently, the SD protocol was aligned with the T-maze training period, lasting for 12 consecutive days. SD was performed for a fixed duration of 8 h per day during the light phase (08:00–16:00; ZT0–ZT8, with lights on at 08:00), which corresponds to the rats' primary resting period. This duration was selected to maximize homeostatic sleep pressure immediately prior to the behavioral task.⁵⁰ During SD, rats were gently stroked or tapped on their cages to prevent sleep onset upon signs of drowsiness, a “gentle handling” method chosen to minimize physical stress compared to forced locomotion.⁴⁹ We note that gentle handling cannot guarantee 100% wakefulness without electrophysiological monitoring; therefore, SD here refers to behaviorally enforced wakefulness, and brief microsleep episodes cannot be completely excluded. Immediately after the daily SD session (ZT8), rats performed the T-maze task with simultaneous LFP recording. Control rats remained undisturbed in their home cages during ZT0–ZT8 and were tested at the same time of day. Following testing, all rats were returned to their housing with ad libitum access to water and restricted food to maintain body weight at ∼85% of baseline. This repeated daily deprivation model mimics chronic sleep restriction, allowing us to examine the cumulative impact of sleep loss on the gradual acquisition of the spatial memory task. Our primary goal is to model the common real-world scenario where learning or critical decision-making occurs under conditions of significant sleep debt. The experimental flowchart is depicted in Figure 1.

T-maze task

Behavioral training and testing took place in a modified T-maze, as illustrated in Figure 2, with the center arm measuring 70 cm in length and each target arm 50 cm. Our T-maze task involved a delayed non-matching to sample task, incorporating a “shaping” session to train rats on spontaneously alternating between target arms before formal testing. The conditions, including color, odor, brightness, and reward quantity, were identical for both target arms to ensure uniformity. The T-maze task was structured into three phases: free choice, delay, and alternate choice. At each trial’s outset, rats were held at the start point behind a barrier. Once the barrier was removed, the rat traversed the maze’s center arm, proceeded to an open target arm, and secured a reward in the target zone. After consuming the reward, the rat returned across the center arm to the starting area, where it was again confined behind the barrier for 30 seconds. Subsequently, starting anew from the initial point, if the rat chose the target arm that was different from the free-choice phase at the choice point, it was able to receive the food reward again and was considered to have performed the task correctly. The opposite was considered as incorrect execution of the task. Rats achieving a behavioral accuracy rate above 80% for two consecutive days were deemed to have met the “acquisition” criterion.

LFPS signal acquisition

A neurophysiological signal acquisition system (OmniPlex/128, Plexon, USA) was applied to synchronously record the LFPs signals (sampling frequency 1000 Hz) acquired from the vHPC and mPFC brain regions of the rats while they were performing the T-maze memory task. At the same time, the position of the rats was obtained by an overhead video tracking system. The data analyzed in this study were the data when the rats were located at the selection point during the alternating selection phase, and the data length was 0.8 s. Because the electrodes were dislodged from some of the rats during the task, the data from six rats per group (control and sleep deprivation) were included in this research. Before applying DCM to analyze the LFPs signals, they were first filtered from 4 to 100 Hz, as well as trap-filtered to eliminate 50 Hz IF interference. To explore the changes in the connectivity characteristics between vHPC and mPFC during the execution of the T-maze memory task in rats pre- and post-acquisition, the LFPs data from rats during the initial stage of task learning (1∼3 days) and the end of the learning stage (rats' correctness on the task was not less than 80% for 2 consecutive days) were selected for the subsequent studies in this paper corresponding to the state of pre- and post-acquisition. DCM was performed on a total of 300 trials from 12 rats (an average of 25 trials per rat). For each trial, LFP epochs were extracted around the choice/selection point during the alternation phase (0.8 s per epoch, Figure S2). Epochs were screened for gross artifacts (e.g., amplifier saturation or movement-related transients) prior to subsequent modeling. After preprocessing (band-pass 4–100 Hz and 50 Hz notch filtering), cross-spectral density features across 4–100 Hz were computed and used as the summary statistics for DCM inversion (DCM for cross-spectral densities, as implemented in SPM12). The same preprocessing and feature extraction pipeline was applied to both regions and both experimental groups.

Dynamic causal modeling

In Dynamic Causal Modeling (DCM), neural mass models are built upon established cortical anatomy and physiology. We utilized conductance-based models representing brain regions as interconnected neural blocks with specific inter- and intra-regional connectivity. These models have been widely used for analyzing electrophysiological data in both animal and human studies.^{51,52,53,54,55} In our study, we employed a typical microcircuit model (CMC) of DCM, incorporating NMDA receptors (spm_fx_cmm_nmda) to address laminar differences in the origin of anterograde and retrograde connectivity within the brain, especially between superficial and deep pyramidal cells.

The CMC utilized in our study included excitatory and inhibitory outward connectivity parameters from four distinct cell layers: superficial pyramidal cells, spiny stellate cells, deep pyramidal cells, and inhibitory interneurons. When considering a single signaling source, excitatory spiny stellate cells transmit received excitatory signals to superficial pyramidal cells and inhibitory interneurons (see Figure 4). In scenarios involving two signaling sources, superficial pyramidal cells in the parietal source convey signals in a forward manner to spiny stellate cells and deep pyramidal cells in the upper layers of the cerebral cortex. Meanwhile, deep pyramidal cells transmit top-down signals in a backward manner from the frontal cortex to superficial pyramidal cells and inhibitory interneurons in the parietal source.⁵¹

DCM of the prefrontal-hippocampal circuit

The extrinsic connectivity types in our DCM are based on connections between different cortical regions²³; there is no clear hierarchical relationship between vHPC and mPFC. Therefore, our initial step was to ascertain the connection type that most effectively elucidates the measured LFP data. Drawing from relevant literature,^28,56 we assumed the existence of bidirectional connections between ipsilateral mPFC and vHPC in this study (totaling 6 distinct scenarios). Bayesian inversion and relative logarithmic evidence of the model were then employed to select the optimal connectivity model, identifying the model most likely to influence interbrain connectivity, corresponding to the factor under study (learning in this case), termed the optimal modulation model. The model with the highest relative logarithmic evidence value was deemed the optimal connectivity or modulation model. Simultaneously, the confidence level of the selected optimal connectivity and modulation models is assessed based on the posterior probability of the model, with higher confidence indicated by posterior probabilities closer to 1. Using the described methodology, this paper comparatively analyzed changes in connection strength between the mPFC and vHPC brain regions in control and sleep-deprived rats pre- and post-acquisition of the T-maze task, as well as during correct and incorrect task performance, to investigate the impact of learning on neural information transfer in these brain areas. The DCM analyses in this manuscript were conducted using SPM12 (https://www.fil.ion.ucl.ac.uk/spm).

Bayesian model inversion and selection

DCM employs a biophysical model of neuronal responses based on a neural mass model to predict electrophysiological data.^11,23,56 Synaptic responses between prefrontal and hippocampal regions (afferent pyramidal cells) are modeled. Model inversion entails fitting each competing model to real data and optimizing its parameters (including alterations in synaptic connectivity parameters) to most accurately predict a given dataset.^53,54 Model inversion yields a probabilistic estimate of the optimized neural mass model parameters along with the model itself. This approximation (the posterior probability of a given model) is utilized to directly assess the model connection structure that provides the most accurate explanation of the data.⁵⁵ Based on the winning model, we assessed the posterior densities of learning-dependent or sleep-dependent modulations to ascertain their consistency in delineating changes in actual data. To establish prior parameter values, DCM inversions were conducted on three randomly chosen rat controls, and the average a posteriori parameter values were employed as the a priori values for DCM inversions across all subjects in each group. This step was taken to ensure the identification of the (local) maximum, i.e., the optimal parameter settings for each group, and to prevent bias in model parameters among groups.²⁹

Histology

Following the experiment’s conclusion, rat brain tissue was sectioned to verify the placement of the microelectrode array within specific rat brain regions. The entire brain of the rat was extracted intact and immersed in a paraformaldehyde solution for 24 hours for fixation. It was then transferred to a 30% sucrose solution for dehydration. Following a 24-hour settling period, blocks of the prefrontal cortex and hippocampal brain regions were carefully preserved. Subsequently, a vibrating microtome was utilized to create coronal brain slices along the direction of electrode implantation, with a slice thickness of 150 μm. These sliced brain sections were sequentially mounted on glass slides and sealed with neutral resin. Finally, the actual electrode implantation sites were photographed using an optical microscope and compared with the rat brain atlas to verify their placement within the hippocampal brain region and prefrontal cortex.

Quantification and statistical analysis

Statistics

This study entails an analysis of 300 trials to examine variations in effective connectivity and connection strength across different brain regions. For behavioral outcomes, group differences were assessed using independent-samples t-tests. Effect sizes are reported as Cohen’s d for t-tests and partial eta-squared (ηp²) for ANOVA. Errors were reported as mean ± SEM. For effective connectivity, Bayesian inference was utilized within the DCM framework. Inferences were based on the posterior probability of the models/parameters given the data. Crucially, the posterior probability serves as the inferential statistic in this framework (analogous to 1 - p value). A posterior probability exceeding 0.95 (providing strong evidence) was considered statistically significant.²³ To address potential instability in parameter estimates due to varying trial counts across conditions (e.g., correct vs. incorrect trials), group-level inference was performed using Bayesian Parameter Averaging (BPA). A fundamental feature of BPA is that it weights individual parameter estimates by their posterior precision (the inverse of variance). Consequently, estimates derived from conditions with fewer trials—which typically yield higher uncertainty and lower precision—contribute less to the group mean. This precision-weighting mechanism ensures that the reported connectivity dynamics are driven by robust, high-precision estimates, thereby mitigating the impact of limited trial numbers in specific sub-conditions.