Hippocampal network axons respond to patterned theta burst stimulation with lower activity of initially higher spike train similarity from EC to DG and later similarity of axons from CA1 to EC

Objective. Decoding memory functions for each hippocampal subregion involves extensive understanding of how each hippocampal subnetwork processes input stimuli. Theta burst stimulation (TBS) recapitulates natural brain stimuli which potentiates synapses in hippocampal circuits. TBS is typically applied to a bundle of axons to measure the immediate response in a downstream subregion like the cornu ammonis 1 (CA1). Yet little is known about network processing in response to stimulation, especially because individual axonal transmission between subregions is not accessible. Approach. To address these limitations, we reverse engineered the hippocampal network on a micro-electrode array partitioned by a MEMS four-chambered device with interconnecting microfluidic tunnels. The micro tunnels allowed monitoring single axon transmission which is inaccessible in slices or in vivo. The four chambers were plated separately with entorhinal cortex (EC), dentate gyrus (DG), CA1, and CA3 neurons. The patterned TBS was delivered to the EC hippocampal gateway. Evoked spike pattern similarity in each subregions was quantified with Jaccard distance metrics of spike timing. Main results. We found that the network subregion produced unique axonal responses to different stimulation patterns. Single site and multisite stimulations caused distinct information routing of axonal spikes in the network. The most spatially similar output at axons from CA3 to CA1 reflected the auto association within CA3 recurrent networks. Moreover, the spike pattern similarities shifted from high levels for axons to and from DG at 0.2 s repeat stimuli to greater similarity in axons to and from CA1 for repetitions at 10 s intervals. This time-dependent response suggested that CA3 encoded temporal information and axons transmitted the information to CA1. Significance. Our design and interrogation approach provide first insights into differences in information transmission between the four subregions of the structured hippocampal network and the dynamic pattern variations in response to stimulation at the subregional level to achieve probabilistic pattern separation and novelty detection.


Introduction
The human hippocampus plays a central role in the formation of episodic memory (Rolls 2010, Roy 2017 and contributes to profound cognitive functions (Lisman et al 2017). In particular, the four main subregions involved in memory formation and consolidation are the entorhinal cortex (EC), dentate gyrus (DG), cornu ammonis 1 (CA1) and CA3. Anatomical and behavioral studies together with electrophysiological investigation through in-vivo recording consistently show functional distinctions among the subregional layers (Langston 2010, Masser 2014, Roy et al 2017, such as pattern identification through primary inputs from EC to sparse representations in DG's larger number of neurons (Leutgeb 2007a, Chavlis et al 2017 and pattern completion in CA3 recurrent collaterals Leutgeb 2007b, Poli et al 2018). The divergent nature of the perforant path projections suggests that information originating focally in EC is distributed widely along the septotemporal axis (Cappaert et al 2015).
To understand these functioning mechanisms as computational properties (Rolls 2010), one must access the hippocampal neuronal activities at the network level. However, recording and decoding the signal with single-axon spatial and single spike temporal resolution are not attainable in vivo, particularly in the axon bundles that relay information from one region to the next. Reconstituting the living connections to study electrophysiology in vitro in a controlled manner is vital for gaining insight into the coding and decoding pathways in memory formation.
Extensive work has demonstrated the high accessibility and practicability of cocultured neuronal networks (Kim et al 2017, Poli et al 2017. In this approach, qPCR tests demonstrated that the subregional cells maintained their identity in the reconstructed culture (Brewer et al 2013). Poli et al (2018) have demonstrated the ability of pattern separation for two-compartment co-cultures. However, most of these investigations were limited to a twocompartment system, e.g. CA3-CA1 co-culture, missing relevant interactions with other subregions of the neuronal circuit, such as time encoding from the EC and pattern separation in the DG (Rolls 2010). The more integrated hippocampal network reconstituted in the four-compartment in-vitro electrophysiological system reported here could provide substantially richer information and insights into the subnetwork communication and processing activities, especially since it includes feed-forward axons from the EC to CA3.
Most of the questions about decoding hippocampal functions arise from how the network responds to outside behavioral stimuli. While this approach is certainly relevant to the real world, it has the disadvantage of not knowing the actual stimuli delivered to the hippocampus. In work with hippocampal slices, stimuli are usually delivered to an entire axon bundle to elicit responses in as many target neurons as possible. Although unphysiological, this approach has contributed to our understanding of how long-term potentiation (LTP) and depression (LTD) might depend on stimulus frequency. Empirical tests indicated that stimulation at theta frequencies (four-10 Hz) optimally evoked increased output LTP of the population spike as the summation of responses from a large number of neurons. Later studies revealed the presence of natural theta rhythms in the hippocampal network and posited their importance for interregional communication (Larson andLynch 1988, Lisman 2017). Theta burst stimulation (TBS) recapitulates natural brain rhythms with short bursts of high frequency gamma at 100 Hz in groups with theta frequency spacing and has been particularly effective in stimulating Schaffer collateral axons in CA1 within the hippocampal slice resulting in LTP (Shimono et al 2002). Valuable circuit architecture and dynamics could be explored further in this reverse engineered hippocampal network with multisite recording of axonal activities in response to TBS. Given the fact that memories are distributed across multiple synapses (Wixted et al 2014, Guzman et al 2016, it is difficult to establish statistically significant causal links between plasticity and outside stimulus, particularly when the actual stimulated neurons cannot be identified. Using the stimulus input distributed to a limited set of neurons or axons would more precisely prescribe one brain region receiving information from others. Therefore, to monitor the information routing in the hippocampal circuit and the localized plasticity that results from defined stimulus patterns, we reverse engineered the hippocampal network with four subregions in a four-compartment platform integrated onto a multielectrode array (MEA). We micro dissected four subregions as the EC, the DG formation, including the hilus (DG), the CA3 and the CA1 including the subiculum. Axonal connections between subregions were enabled by the narrow microfluidic tunnels between the compartments, from which the electrical activity of the axons passing through them were sampled by each of the two embedded electrodes. From the two electrodes, analysis of spike timing allowed the determination of directionality as either feed forward or feedback. We designed the patterned TBS delivered to three sites in EC to facilitate studying the evoked responses and corresponding axonal transmission among the four subregions. Several key questions we sought to answer were: (1) Is the network capable of transmitting the patterned stimulus to the other subregions? (2) How do the communications between subregions differ to affect their encoding of that stimulus and how long do these differences persist post-stimulus? (3) For stimuli repeated at different timescales, does the network make adjustments in plasticity of axonal transmission between each subregion?
To these ends, a digital quantification method was employed that directedly compared the temporal spiking patterns at single-synaptic delay resolution of 3 ms (Poli et al 2018, Chung andEdwards 2019), down sampled from spikes recorded at 40 µs resolution. We first demonstrated that this in vitro hippocampal network uniquely distinguished single site and multisite stimuli. The most similar betweenaxon similarity in CA3-CA1 output showed striking agreement with anatomical CA3 recurrent networks. Furthermore, through the specific firing timing and propagation directionality, the constructed routing maps both confirmed the distinct processing pathways associated with different stimulation patterns and exhibited stochastic responses. Finally, the direct comparison of similarity in responses arising from both shorter (0.2 s) and longer (10 s) timescales elucidated a functional transfer from early subregions to later ones for recognizing and remembering the repeated stimulations. We were able to map variations in the timing of axonal transmission from one subregion to another, in both feed-forward and feedback directions.

Device fabrication and assembly
The microfluidic platform for neuron culture with electrophysiological monitoring consisted of two main components: (1) a custom designed and fabricated four-compartment culture chamber integrated on top of the MEA120 (figures 1(A) and (B)) and (2) a commercially available MEA120 (figure 1(C); MultiChannel Systems, Reutlingen, Germany) with 120 low impedance 30 µm diameter electrodes with 200 µm spacing and electrically insulated (with silicon nitride) interconnects serving as the bottom substrate ( figure 1(B)). The culture chamber (figure 1(B)) was made from Polydimethylsiloxane (PDMS) (SYLGARD™ 184, Dow, Inc) with micromolding techniques. The master mold was fabricated with photolithography on two layers of thick photoresist (PR). The supporting substrate was a 4inch silicon wafer cleaned with a brief dip in hydrofluoric acid followed by a thorough rinse in deionized (DI) water. In order to align the two PR layers, a set of chrome (Cr) alignment marks was first photolithographically patterned with lift-off technique on the wafer. Shipley 1827 positive PR was spin coated on the wafer at 4500 rpm for 30 s to a thickness of 2 µm, followed by a soft bake at 120 • C for 1 min and exposed for 20 s through the alignment-mark mask in a mask aligner (Karl Suss MA6) under 10 KJ cm −2 . The exposed PR was then developed in MF-319 developer (MICROCHEM) for 45 s, rinsed with DI water and blow dried using a nitrogen gun. A thin layer of Cr (1000 Å) was deposited over the entire wafer in an Ebeam evaporator (Airco Temescal CHA-600S/CV-8) and then the PR lifted off in acetone. After the alignment marks were finished, two sequential PR layers were deposited and patterned. The first layer was 3 µm thick SU-8 2000 negative PR. It was patterned and developed to form 8 µm wide lines that were 3 µm high and oversized 600 µm long, which would subsequently form the mold for the PDMS microtunnels. In the finished platform, these microtunnels allow isolation of axons that grow from one chamber to the connecting chamber through the microtunnels. The second layer was 500 µm thick SU-8 3500. The photomask for this second layer was used to pattern the parts of the mold for the four main culture chambers separated by 400 µm and were 500 µm tall. These two photomasks were aligned to the Cr alignment marks to ensure correct feature definition. After developing the second PR layer, the wafer was hard baked at 150 • C for 30 mins to further solidify and stabilize the SU-8 mold. Before the molding process, the mold was first vapor coated with release agent triethoxysilane (Gelest, Inc) for 2 h. SYLGARD™ 184 Silicone Elastomer Kit (Dow, Inc) was used at the recommended 10:1 mix ratio of the polymeric base and curing agent. After thorough mixing and degassing under partial vacuum, the mixture was poured over the mold wafer to slightly more than 500 µm thickness. A final degassing step was performed until all gas bubbles were visibly dissipated. To ensure a flat and even PDMS top surface, a PET plastic film was applied on top of the PDMS with a 355 gm weight added to make sure that the 0.2 mm thick PET film was in contact with the top surface of the SU-8 chamber mold. The PET film would also serve to ease the final demolding step. The entire assembly was then cured at 60 • C for 12 h in an oven on a shelf that had been carefully leveled. Finally, the demolding process involved gently peeling off the PET film, which would carry the PDMS piece with it. Each PDMS device was punched out with an 11 mm biopsy punch (Acuderm Inc.) and aligned onto the MEA120 substrate under a microscope with the aid of 10 µl 70% ethanol to promote sliding mobility.

Device preparation for cell culture
To promote device sterilization and enable filling the microfluidic tunnels with media by gravitational force, we performed the following surface preparation steps. The assembled MEA and PDMS device were cleaned with 70% ethanol and rinsed with DI water twice. The liquid residue was vacuumed out and device was left to dry for 1 h. Then the device was treated for 2 mins with 35 mA air/oxygen plasma (EMS Sciences Emitech model BMS100x-025, Ashford, Kent, England) to create a hydrophilic surface. Poly-D-lysine (PDL) solution (Sigma-Aldrich P6407, 0.1 mg ml −1 in water) was loaded into each of the four chambers sequentially to fill the tunnels at room temperature overnight. On the second day, the PDL solution was vacuumed out and rinsed with DI water once. The device was left in a biosafety hood for 1 h to dry. Before cell plating, the device was filled with culture medium and placed in the incubator to equilibrate. Each mold created on a 4-inch silicon wafer contained 28 PDMS devices. The 2-layer process included a thin (3 µm) SU-8 layer that defined the 10 µm wide microfluidic tunnels that connected neighboring chambers and a thick (500 µm) SU-8 pattern that precisely defined the chamber height and lateral geometries. (B) The four-compartment culture chamber was aligned to the electrodes on the (C) multi-electrode array (MEA) under a microscope (not shown). The alignment ensured that some of the tunnels would have two electrodes underneath each of them and the rest of the electrodes would be evenly distributed among the four compartments.

Cell culture
The primary rat hippocampal cultures procedure was based on an established lab protocol (Brewer et al 2013). All procedures were approved by the UC Irvine Institutional Animal Care and Use Committee. For this purpose specifically, the whole hippocampus and overlying EC was removed from postnatal day 4 (P4) Sprague-Dawley rat (Charles River), then the four subregions (EC, DG, CA3, CA1) were separately dissected en bloc (Vakilna et al 2021) (figure 2). The subregional tissues were collected in Hibernate A and processed into neuron suspensions in culture medium with targeted densities: EC: 800; DG: 2500; CA3: 800; CA1: 1000 cells µl −1 . After aspiration of equilibrated media with care to avoid emptying the tunnels, subregional neurons were plated sequentially into four chambers on MEAs in Neurobasal A/B27 media (BrainBits, Springfield, IL) medium (Brewer and Torricelli 2007) or BrainPhys plus SM1 (Stem Cell Technologies) both with 0.5 mM Glutamax (Fisher Scientific) and growth factors FGF2 and PDGFbb (Peprotech) at 10 ng ml −1 each. Plating densities were 330/1000/330/410 cells mm −2 ,at ratios intended to mimic the in vivo anatomical density ratio for the subregions in the rat hippocampus (Braitenberg 1981). Cultures were incubated in a humidified incubator with 5% CO 2 , 9% O 2 at 37 • C (Forma Model 3520). Half of the culture medium was removed and replaced with the same volume of fresh medium every 3-4 d. Growth factors were added at 2x concentration to replenish the original concentration during the first week only. Neuron growth was monitored in each chamber through phase contrast microscopy every 7 d.

Fluorescent imaging
To verify the presence of axons in the microfluidic tunnels, Calcein AM (Thermo Fisher), a cell-permanent dye that stains live neuronal cells was used. The dye was first dissolved in DMSO at 1 mg ml −1 . Approximately 2 µM final concentration was added to the cell culture and then incubated for 30 mins at 37 • C in the 5% CO 2 , 9% O 2 incubator. A fresh wash with medium was done for 15 mins to remove excess ester and reduce the background intensity. NucBlue Live cell stain (Thermo Fisher) that emits blue fluorescence from UV excitation on binding to nuclear DNA was used to identify cell bodies and facilitate counting. Two drops of NucBlue were added to the 1 ml media well, and then the cell culture was incubated for 15 mins. The dyes were excited with GFP and DAPI filter Hippocampal dissection from P4 rat brain. (A) The whole brain was removed. (B) The hippocampus was located, and each subregional tissue was directly separated under the microscope. The illustration shows a cross section, but the dentate gyrus (DG) was dissected from the whole hippocampus as a string of tissue followed in sequence by the whole CA3, CA1 and entorhinal cortex (EC). cubes, respectively in an inverted Olympus IX83 fluorescence microscope. Images were acquired with a Hamamatsu camera (ORCA-flash 4.0, Hamamastu Photonics) driven by MetaMorph Basic software (Molecular Devices).

Electrophysiology and stimulation protocol
The cultures were kept in the incubator for three weeks to differentiate dendrites, axons, and form synaptic connections. The mature network ensures optimally active recording channels. Prior work established maintenance of subregional neuron types by PCR analysis of gene expression enrichment specific for DG, CA3 and CA1 (Brewer et al 2013). To examine axonal responses between each subregion of the hippocampal network, we delivered patterned TBS to one, two and then three electrodes in the EC chamber (figure 3(A)) because the EC gates hippocampal inputs in the brain for processing to downstream subregions. Specifically, the 100 Hz high frequency pulse trains had five biphasic pulses of 10 ms inter-pulse interval, 400 µs pulse duration and 10 µA amplitude with negative pulse first. This low amplitude of stimulation was chosen to reduce responses to about 100 ms of spiking activity so as not to carry over into the next 200 ms time window. The trains were delivered at 200 ms intervals (5 Hz). Three TBS with varying train numbers were delivered to three electrodes (A, B, C figure 3(B)) in a staggered order. Therefore, the single site (Stim site A), two-site (Stim site A, B) and three-site (Stim site A, B, C) stimulations were delivered to the network consecutively.
Note that a pre-stimulation recording served as the control for the single-site stimulation (S1), which was the control for the two-site stimulation (S2), which served as the controls for the three-site pattern (S3). This order was designed to stimulate a limited number of neurons in the network and allowed analysis of the responses based on the timing delays. Note also that S3 was repeated four times (S3-1, S3-2, S3-3, S3-4), followed by a spontaneous recording control. Three distinct EC electrode channel combinations were chosen as stimulating sites, based on robust spontaneous activity. Each set of multisite combinational stimulations had ten repetitions with 10 s intervals and 2 min breaks in-between the sets. To investigate the prospect for pattern completion, a 'partial-cue' signal was designed as the stimulus train only delivering to 1st and 2nd site (AB). These sets of stimulations were performed in the same temporal order 9 min after initial training. The different repeats allowed us to evaluate the ability of network to 'learn' and detect synaptic plasticity in changes in axonal transmission at different timescales of 200 ms, and 10 s.

Data processing
The raw data was collected by MC_Rack software (MultiChannel Systems) at 40 kHz through 1100x amplifiers with hardware cutoffs for bandpass between 1 and 50 kHz. The files were converted into the hierarchical data format, H5 and then to single channel files (.mat format) using Multi-Channel Datamanager and MATLAB (version 2020a MathWorks Inc.). See table 1 for files for six arrays with cultures recorded from six animals pooled into one preparation and ten repetitions of each stimulus per array. Since the spike detection algorithm was based on the threshold detected as a multiple of standard deviation (S.D.) of the signal in 200 ms contiguous windows, the artifacts from stimulation would contaminate measures of the noise. Therefore, the first 50 ms segment following each stimulus delivery were blanked out through all stimulated recordings. Spikes in each single channel were detected after the raw signal was filtered through a 300-3000 Hz bandpass filter, followed by application of an optimal amplitude threshold calculated from Wave_clus MATLAB toolbox (Quiroga et al 2004).
Because the axonal coupling to electrodes varied from tunnel to tunnel, the signal amplitudes between tunnels exhibited large deviations, ranging from tens of microvolts to a few thousand microvolts. To optimize the spike detection performance from Wave_clus, differential thresholds were chosen as −4.5 or −11 times the noise of the signal that was calculated for 200 ms contiguous windows. The spikes were detected with a refractory period of 2 ms, to preclude mistiming spikes from multiple axons even though most tunnels contained only one or two axons (Narula et al 2017). For each 150 ms window following a stimulus train, the probability distribution of interspike intervals (ISI) followed a log-log distribution.
The distributions were normalized as complementary cumulative probability distributions (CCDs) with logarithmically spaced bins (Newman 2005, Vakilna et al 2021. A linear model was used to fit the CCD after log transformation (1), where P is the cumulative probability, α the slope, t the interspike interval in ms and c the intercept. The best fits (highest R 2 ) were found through performing a grid search to identify the local maximum with time limits shortened up to 50% with a step size of 5%. The relative order of two spike trains was determined from the lag of the cross correlation (MATLAB xcorr).

Jaccard distance (JD)and similarity
The processing was based on custom-written MATLAB scripts to realize certain functions. The spike train was used to characterize pattern separation in the DG on the scale of milliseconds (Madar et al 2019). Spike train similarity and synchrony were also important measurements for in-vitro neuronal networks (Ciba et al 2020). For quantifying the similarity and dissimilarity of the spike patterns, we used JD; (Jaccard 1912) calculation of the binary vectors derived from temporal response patterns of spikes in 3 ms bins (Poli et al 2018). We considered the neuron spike width and related refractory period of 1-2 ms to decode the binary information at 3 ms bins. The activities being evaluated were based on each of the 150 ms post stimulus windows. Also, two segments of the same length before and after the stimulation were used as spontaneous controls ( figure 3(B)).
(Labeled as 'Pre' and 'Post'). The binary vector converted from each spike train had 50 × 3 ms bins (150 ms) indicating whether spikes were present in the bins at certain times. The calculation of JD between two vectors was defined in equation (2).
where j 01 is the count of event bins with no spikes in the first train and bins with spikes in the second train. J 10 is the opposite and j 11 is the number of time bins with spikes in both trains. A detailed example was in figure 4(B). The value of JD quantified the dissimilarity between two spike trains while similarity was represented as 1-JD.

Statistics
Data analysis was performed using custom-written scripts in MATLAB (2020a), including functions from the statistics toolbox on ten repetitions withing each array and averages from six arrays, as stated in each figure. We performed one and two-tailed t-tests, ANOVA or ANCOVA whenever the data were normally distributed and represent variation from the mean by 95% confidence intervals. Where multiple comparisons were made, Tukey's Honestly Significant Difference test was employed. For non-Gaussian distributions we used the Wilcoxon Ranked sum test. Monte-Carlo chance was calculated in MATLAB by shuffling the axon identities one hundred times. All values were reported as mean ± S.E.M. unless otherwise noted. Criterion for significance was pvalue < 0.05.

Results
We present measures of spike dynamics of the axons that transmitted information between the four compartments of a reverse engineered hippocampus in response to TBS delivered to one, two or three sites in the EC subregion. Table 1 shows the 10 repetitions of each stimulation after averaging from 6 arrays of networks. In section 3.1, we explain the Jaccard distance measure of dissimilarity and its converse similarity between evoked spike trains. In section 3.2, we analyzed differences in axon dynamics for each of the three sites of stimulation. In section 3.3, we further compared each axon with another from the same subregion to document low levels of similarity with repetition for the three-site stimulation. In section 3.4, we examined the network response to different threesite locations. In section 3.5, we compared repetitions at 0.2 s to 10 s repetitions to discern shortterm plasticity. In section 3.6, we used cross correlation measures for individual axons traversing two electrodes to determine directions of feed-forward and feedback propagation and graphically mapped the progression of communication within the network.

Use of Jaccard distance to measure similarities and differences in axonal spike patterns evoked by patterned stimulation
In order to determine specificity for variations in stimulation sites, the reliability of responses and their subregional specificity, we stimulated the network with a short theta burst at either one, two or three sites (figure 3) and measured responses in the next 150 ms after each stimulation in four of the 67 tunnels between each subregion (11 diagonal tunnels connecting EC-CA3). In figure 4(A), the 50 ms stimulus (marked with blue bars) evoked responses in the following 150 ms post stimulus windows (red boxes). The selected examples in figure 4(A) show four parallel tunnels in CA3-CA1, in which single site (S1), two-site (S2) and three-site (S3) stimuli were delivered to evoke repeated responses (3 of 10 repetitions shown). A zoom-in to 150 ms analog signals (figure 4(B)) shows the activity patterns in the individual tunnels. The four tunnels exhibited roughly synchronized activities with varying amplitudes from tunnel to tunnel ranging from −50 µV to −2000 µV ( figure 4(B)). This variation in amplitudes was likely the result of differences in the axon-electrode coupling. Different amplitudes on an electrode suggest separate axons generating the signals. For example, in tunnel 2, there was clear evidence of two axons in contact with the same electrode, one with amplitudes around −200 µV and the other around −800 µV at their peaks. Differences in how axons responded to different stimuli were quantified by converting the analog signals to a binary 'digital' vector ( figure 4(B)). The Jaccard distance between two tunnels represented the dynamic diversity of response patterns better in 3 ms bins than at 10 ms or larger which caused lower resolution of differences in temporal pattern variations. This example shows that the single site stimulation evoked patterns with high Jaccard similarity of 0.59 between tunnels 1 and 2, while three-site stimulation S3-1 for tunnels 1 and 2 (figure 4(C)) evoked a lower similarity (greater distance) of 0.4, apparently blocking the larger spikes. This suggests that stimulation at three sites evoked greater diversity of axonal information transmission. Next, we examined these differences between tunnels by subregion to provide evidence for unique information processing in each subregion.

3.2.
In response to three-site stimulation compared to one and two-site stimulation, spike rates changed with unique patterns of axonal transmission in tunnels between subregions, consistent with pattern recognition Since each three-site stimulation included stimulation at the same one-and two-sites, we wanted to determine whether a pattern of stimulation at three sites evoked additive or different results from single and two-site stimulation. We examined the axonal spiking patterns of each 150 ms post-stimulus Spike patterns are shown in four tunnels for the S1 stimulation showed high similarity between axons from CA3 to CA1. (C) With the first three-site stimulation (S31), the large spikes from S1 were blocked resulting in a less similar pattern of activity in each of the four tunnels from the same axons.
window for the three-site stimulation compared to a single-site stimulation and two sites. The three-site stimulation was repeated four times with 0.2 s delays to evaluate the reliability of responses, all compared to the pre-stimulation and post-stimulation activity patterns as baseline references. Of the 67 tunnels between each pair of four subregions, four tunnels were monitored by electrodes to sample activity. For example, in the tunnels between CA3 and CA1 (CA3-CA1), responses evoked by single site sulation (S1) could be compared to three-site stimulation (S3) for the same axons to detect distinct patterns between the two kinds of patterned stimulation ( figure 5(A)). Vakilna et al (2021) previously reported that neurons code information over a log-log distribution of ISI, not a Gaussian distribution of firing rates. Here, we accumulated the distribution of spike intervals on a log-log scale (Newman 2005) to compare the spontaneous activity to S1, S2 and S3 responses in linear fits ( figure 5(B)). Up to two-fold differences in initial slopes were observed. Stimulation at a single site reliably evoked spikes with about a 20% increase in slope for CA3-CA1 axons (shorter ISI, faster spiking) with 30% more spikes (figure 5(C)). Strangely, two-site stimulation evoked slower spiking and 60% as many spikes as the pre-stimulation control. Threesite stimulation restored spike rates to slightly higher levels than pre-stimulation. One hundred-fifty ms after these stimuli, including three more repeats of the three-site stimulation, lowered the CA3-CA1 axonal ISI slopes by about 25% below those at the start. Responses of axons in other subregions (figures 5(D)-(G)) were less dramatic with each type of stimulation but consistently showed declines in slopes and as much as 50% less evoked spikes, compared to the preand post-stimulation controls.
To evaluate the precision of pattern differences between one, two and three-site stimulation, we compared the binary presence or absence of a spike in 3 ms bins (temporal spiking) by computing JD for dissimilarity and (1-JD) for similarity as in figure 4. The JD calculations were performed on the same axons for the responses to repeated stimulation every 0.2 s. Results for one array showed that four tunnels exhibited S1 vs. S3 dissimilarity around 0.8 and dissimilarity between S3 and S2 close to 1 ( figure 5(H)). Combining the data from six arrays, distinct responses were evoked for the different patterned stimulations across all subregions ( figure 5(I)). This suggests that these non-identical networks did not additively respond to increasing stimulation sites but robustly distinguished threesite stimulation from single and two-site stimulation, both in terms of ISI and JD measures of dissimilarity. These data indicate that the four-compartment hippocampal networks readily and robustly recognized different stimulus patterns, consistent with pattern separation and inconsistent with additive summation of input signals.

Between axon comparisons within subregions showed low similarity of spiking that indicated different information transmitted for each stimulation pattern. Spiking patterns of output axons from CA3 to CA1 were more similar than between other subregions, consistent with recurrent connections within the CA3 network
The above analyses compared responses within axons to stimuli. Next, we examined the similarity of responses (or synchrony) between different axons at the subregion-specific level. Brain function is partly realized through synchronized activities (Nikolić et al 2012, Gollo et al 2014. We examined our hippocampal networks to determine similarity of evoked responses (synchrony) between axons within their respective subregions when presented with different stimuli. In our design, there were four axon tunnels with electrodes from among the 67 tunnels that structurally connected axons between adjacent subregions. The similarity of activities in these tunnels indicated how synchronized the outputs were with each other when assessed by pairwise axon similarity using the (1-JD). The averages from all axon combinations for each subregion were used to evaluate subregional similarity. To begin with transmission from CA3 to CA1, the responses to S1 were more similar ((1-JD) = 0.59) than responses to S3-1 (0.16), as color coded in figure 6(A) in most of the repeats. A large gray square (such as figure 6(A), rep 3) means no response was recorded in at least three channels in that stimulation repetition so that similarity could not be calculated. A few reasons reflect the plastic dynamics of brain circuits: (1) alternate routing of axonal communication through tunnels not monitored. (2) In the momentary context, a certain stimulation could occur at a network down-state (Vakilna et al 2021) or at a time of strong inhibitory signaling. (3) As seen in subsequent repetitions, axonal signaling adapts to certain stimuli over time, becoming less (or more) responsive to subsequent stimulation. Averages for each of the five subregions showed distinct variations in synchrony among axons for this array (figure 6(B)), suggesting that each subregion transmitted information in a unique fashion. In CA3-CA1, each kind of patterned stimulation (single, two-site or three-site) evoked the most distinct between-axon similarities, further indicating that the output responded differently to different inputs, which would be consistent with novelty detection. In EC-DG, and EC-CA3, the stimulus evoked between-tunnel similarity did not change significantly from the 'pre' control, indicating a relatively stable synchrony. Interestingly, for the output of DG to CA3, the synchrony levels observed at different sites started to show distinguishable differences, where S1 increased the synchrony substantially (0.4), S2 declined, then S3-1 reached 0 synchrony, implying unique outputs, indicative of pattern separation for this unique new stimulus. Note subsequent Figure 5. Three-site pattern stimulation evokes unique axon spike patterns in all subregions compared to one and two-site in separate tunnels at a 0.2 s timescale. (A) 150 ms raw spike patterns showed differences in spike timing evoked by S1 vs. S3-1 in four parallel tunnels (F9, F10, G11, G12 of array 33152). (B) Cumulative log-log distributions of interspike interval (ISI) from before stimulation (pre-control), S1, S2, S3-1. and post-stimulation control revealed large differences in slopes of the distributions (n = 6 arrays, 10 repetitions per stimulus, all axon spike intervals combined). Best fit slopes are shown for the initial linear slope. (C) The CA3-CA1 slopes of fitted lines on cumulative ISI distributions showed differences with significant p values by ANOVA for each kind of stimulation (n within bars is the total spikes (ISI's +1) in 150 ms, 10 repeats, 6 arrays). P-values calculated using ANCOVA followed by Tukey-HSD. The error bars represent the 95% confidence intervals of all the ISI's obtained by ANCOVA. (D) Slopes for EC-DG axon ISI's. (E) Slopes for EC-CA3 axon ISI's. (F) Slopes for CA1-EC axon ISI's. (G) Slopes for DG-CA3 axon ISI's. (H) Jaccard dissimilarities in the same tunnel for different patterns of stimuli S3-1 vs. S1 and S3-1 vs. S2 for four tunnels (array 33152, CA3-CA1, n = 10 reps). The subregional mean was calculated from 10 repetitions with 3 ms bin size. One-tail t test showed that the four tunnels all had S1/S3-1, S2/S3-1 dissimilarity greater than 0 (p < 0.001). Two-tailed t-test showed that only tunnel 3 was significantly different between S1/S3-1 and S2/S3-1 (zero bins were redacted). (I) Multiple array average showed distinct responses to patterned stimulation across all subregions (dissimilarity approaching 1). Figure 6. At 0.2 s repeat comparisons, between axon similarity and subregional differences to assess synchrony (A) Between-axon similarity for the activities evoked by S1 and S3-1 are calculated as the pairwise combinations among each of the four axon tunnels for each repetition, respectively. S3-1 has less similar output at CA3-CA1 than S1 (n = 1 array 33152). Non-diagonal NaN gray squares are not-a-number because one of the inputs had no spikes. (B) Stimulation sites S1/S2/S3-1 evoked distinct similarities). Different similarity trends are seen in different subregions (n = 10 reps x 6 pairwise combinations = 60 for array 33152). (C) For average data from six arrays, CA3-CA1 had the highest similarity during the three-site stimulation and demonstrates high discrimination (synchrony) between stimulation patterns (p-values from t tests).
repetitions of S3 in the context of prior same-site stimulation again increased the similarity of axon responses. The EC to CA3 axons behaved similarly. This could be the evidence that EC directly innervated CA3 with less synchronized activities instead of the output from CA3 feeding back to EC, which should have higher synchrony for specific sites as in CA3-CA1. For CA1-EC, there was a drop in similarity on S2, which might be caused by sparse data from S2. However, the other sites evoked activities at a similar level of low axon synchrony as the 'pre' and 'post' control, implying a stable finish of the neural circuit. In figure 6(C), we compiled the results of multiple arrays (n = 6). This confirmed that three-site stimuli evoked the highest between-tunnel similarity in CA3-CA1 output axons with values around 0.3. CA3-CA1 also showed the most distinctions from post-stimulation synchrony over the pre-stimulation control activities. Subregions EC-DG, and CA3-EC showed unchanged levels of synchrony from spontaneous spiking suggesting no new mechanisms of information processing were evoked by stimuli, but a significant off response post-stimulus. On the other hand, DG-CA3 showed moderate increases in synchronized transmission S2 to S3-4, and CA1-EC at S3-2. These results demonstrated that the network processed patterned stimulations embedded with spatial diversity and subregional differentiations. Finally, the putative auto-associative network in CA3 contributed to the highest similarity for three-site stimulation in responses sent to the CA1 (CA3-CA1).

Evoked patterns of axonal activity changed with different stimulation sites: the pattern novelty test
To determine which stage of the hippocampal network best recognized a change to a novel stimulation, following the first set of ten repeats of three-site stimulations, two more sets of repeated stimuli were delivered on different site combinations in EC. In an example from CA3-CA1 channel G11, the different three-site stimulation resulted in burst patterns entirely different from the original three stimulation sites for ten repeats ( figure 7(A)). The 150 ms analog signal further highlighted the differences of spiking patterns of S1 from different stimulation sites ( figure 7(B)). Upon calculating pairwise dissimilarity, a large portion of the data was within the high dissimilarity range of 0.9-1. Specifically, 60% S1, 77% S2 and 51% S3-1 activity metrics were above 0.9 dissimilarity, meaning that the patterns of activity from different stimulation site combinations were unique (figure 7(C)). Statistical analyses confirmed the differences in similarity between single site and multisite stimulations (S1, S2, S3-1). Lower S3-1 dissimilarity than S2 or S1 indicated that despite including two of the prior sites of stimulation, the network recognized a third site as novel and transmitted mostly unique axonal information. For the 5-way axonal connections between subregions, S3-1 responses from different stimulation site combinations showed distinctly different distributions. Except for CA3-CA1, all other subregions showed that 80% of the pairwise dissimilarities were 0.9 or above ( figure 7(D)). These results demonstrated unique axon spike patterns in response to novel stimulation patterns, i.e. novelty detection.
3.5. Ten second repeat three-site stimulation evoked less similarity than 0.2 s repeat in early subregions, but higher similarity in CA1-EC Network recognition of novelty and even shortterm memory formation are affected by different patterns of responses and information routing in the hippocampus, putatively specific in the CA1. The nervous system uses different timescales to achieve efficient and essential memory tasks (Tetzlaff 2012). Information about how the network recognizes repeated stimulation and generates corresponding responses would help in understanding the memory formation process. A simplified interrogation of the effects of memory formation as levels of axonal transmission were evaluated by examining the response retention with two stimulus repetition timescales of 0.2 s and 10 s for three-site stimulation ( figure 3(C)). In the upstream subregion of EC-DG, spike patterns in the shorter 0.2 s repetition timescale repeat with high similarity half of the ten repetitions (figure 8(A)) but turned into less similar patterns for longer 10 s repeats ( figure 8(A)). The opposite phenomenon was seen in the downstream subregion CA1-EC, where longer timescales showed more similarity in spike timing than shorter timescales ( figure 8(B)). Average statistics for these response similarities by subregion from six arrays and the two repeat times were calculated pairwise on the same axons (figure 8(C)). In contrast to the 0.2 s timescale, the longer interval of 10 s showed a different distribution of similarity levels with axons to and from the CA1 evoking higher similarities. As a control, we calculated similarities from the Monte Carlo randomized evoked spikes from shuffled axons 100 times (red lines in figure 8(C)). To better compare the similarities on different time scales, similarities were normalized to chance values as 'fold above chance' in figure 8(D). At 0.2 s repetition, axons to and from the DG were most similar (earlier subregions), in contrast to less similarity for 10 s stimulation repeats. At the longer 10 s repetition, these earlier DG-connected subregions in the circuit loop changed from modest or low similarity, to increased similarity in the later subregion axons connected to CA1, exceeding the 0.2 s for CA1-EC ( figure 8(D)). The results indicated a transfer of response in hippocampal subregions when recognizing repeated stimulation between two repeat Figure 7. Different three-site stimulation sites evoke substantially different CA3-CA1 responses (A) Overview of channel CA3-CA1 tunnel G11 burst activities from 3 different stimulation sites in EC number 1 B3A5C4, 2 D1E1C2, 3 E5D4E4 (array 33152). (B) 150 ms window showed dissimilar spiking patterns for single site (S1) stimulation compared to the prior three-sites combination (Rep 9, 9, 6 respectively). (C) Frequent dissimilarity of 0.9 or above for histogram of new activities compared to prior combinations of stimulating sites for S1, S2 and S3-1 in CA3-CA1 from array 33152, calculated for ten repetitions, all permutations. With redaction of zero-response events. (D) For S3-1 in one array, all subregions produced highly dissimilar spike patterns compared to prior three-site stimulation. Figure 8. Hippocampal EC-DG axons report higher similarity spiking for 0.2 s repeats of three-site stimuli than 10 s repetitions; CA1 axon similarities are more equal in similarity for both times. (A) Spikes in EC-DG channel F2 showed more similar patterns at 0.2 s than 10 s repeats. (B) More similarity of spike patterns in CA1-EC after 10 s repeats than 0.2 s repeats. (C) Averages of six arrays showed early subregional higher similarity for 0.2 s than 10 s repeats, significantly above Monte Carlos chance levels for both timescales. (D) Subregional response similarities normalized to Monte Carlo chance (red line) showed higher similarity at 0.2 s repeats than 10s s repeats at earlier subregions EC-DG, DG-CA3, and EC-CA3 before transfer to downstream subregion CA1-EC with higher similarity at 10 s than 0.2 s repeats and higher than 10 s similarities than earlier 10 s repeats. timescales, a stronger coordination of responses to 0.2 s repetitions in the DG and to 10 s repetitions in CA1-EC transmission, suggestive of short-term novel pattern recognition for three-site stimulation compared to one and two-site stimulation.

Order of routing through the network changed with the type of patterned stimulus
Lastly, to understand the adaptability and specificity of the routing of responses to stimulation, we compared the routing of responses to the different stimuli by determining the timing of the first spikes and their feed-forward or feedback direction. We used cross correlation on the high pass filtered data of an axon over two electrodes in the same microtunnel to determine their prominent directionality ( figure 9(A)). The lags represented the delays between spike times recorded from the two electrodes. For example, the CA3-CA1 subregional tunnel electrodes F12-G12 showed a positive time delay, which means that the spike propagation was from G12 to F12, and thus was forward from CA3 to CA1. Based on the Figure 9. Subregional routing order of network processing of patterned stimuli. (A) The use of cross correlation of the signals in the four tunnels connecting CA3 and CA1, F9-G9, F10-G10, F11-G11, and F12-G12 to determine the feed forward or feedback directionality (arrows) of signal propagation of spikes for 3-site stimulation (array 33152). (B) Time delays were compared after each stimulus to determine routing (S3-1 stimulation, array 33152). (C) Routing maps for S2 response within each of ten 15 ms bins of the 150 ms recording (array 33152), then grouped in three-time segments of 1-4, 5-7 and 8-10. Map shows direction (color coded blue for feedback and red for feed forward) and timing (shading and width). (D) Routing maps of average response from six arrays and ten repeats (n = 60) for S1, S2 and S3-1 demonstrated both routing and directionality dominant differences associated with three stimulating site patterns (n = 6 arrays). (E) Time delays in five axonal subregions for S1, S2 and S3-1 (n = 6 arrays). For the last pair of CA1-EC bars, the blue bar represents EC to CA1 (feed forward), and the red bar CA1 to EC (feed forward) (and feedback cannot be discriminated). averaged time delay from ten repeats, each tunnel's axonal transmission direction was determined and labeled as either 'FF' for feed forward or 'FB' for feedback. After each stimulus, the first response in each subregion and its directionality were extracted and recorded ( figure 9(B)). The time at the start of each subregional response (first spike after end of stimulus) was used to generate a color-coded routing map from this stimulation of a single array for the ten possible orders (2 directions × 5 subregional connections) (figure 9(C)). Results from six arrays and ten repeats showed the distinct response routings from the patterned simulations (figure 9(D)). After single site stimulation S1, the response routing in the network appeared to exhibit quicker FB than FF axonal signaling, possibly due to sparse coding of the feed forward and our limited sampling from only four of the 51 communicating axon tunnels, some of which would likely be feed forward. Another interpretation is that single site stimulation activated prior spontaneous episodes, i.e. memory reactivation. The average post stimulus delays between FF and FB directions at most subregions did not show significant differences (figure 9(E)). However, the post-stimulus responses from S2 and S3-1 for the fast FF at EC to DG and EC to CA3 as well as FB from EC to CA1 demonstrated different routes for the flow of information in the network in response to the stimuli, despite initial stimulation at EC. Note that because EC in vivo normally feeds forward into CA1 (Rockland andVan Hoesen 1999, Kitamura et al 2014) as well as the longer routes through DG and CA3, we could not distinguish the EC feed forward into any subregion from the EC feedback to any region, but especially the CA1. Most likely, both directions were represented in CA1-EC axon tunnels that would need to be determined by axon tracing. The directionality timing maps associated with patterned stimuli revealed different information routing in distinct pathways and thus demonstrated that this in-vitro hippocampal network incorporated robust network processing instead of a unidirectional monolithic feed-forward processing.

Overview
In this four-compartment reconstruction of the hippocampal network, we gained electrical access to the communicating axons between subregions. This allowed us to begin to ascertain the flow of information in response to patterned input stimuli from the EC. Individual axonal transmission is hard to monitor in vitro or in vivo, and impossible to realize at the network level with current technology. Although input-output models predict the spiking patterns of specific subregions in the hippocampus (Geng et al 2019, She et al 2022 and have even helped memory in patients with brain injury (Roeder et al 2022), the axonal transmissions provide more direct insight into the computational properties of each subregion. Single subregion dynamics of somal calcium spiking indicates cell assemblies involved in CA1 place cells in freely behaving mice (Ziv et al 2013), but fail to track the activation of upstream hippocampal subregions. The digital encoding of pattern separation and completion (Poli et al 2018) using Jaccard Distance measurements led to the ability of decoding the network processing properties as they respond to stimuli. At least six major observations emerge from this hippocampal model. (1) Axonal spike times between each subregion are log-log distributed and shifts in response to stimulation with no evidence of a peak spike rate. (2) Spontaneous wiring of neurons in the network produces widely disparate axonal spike trains between axons in each subregion in response to similar input stimulation. (3) Only CA3-CA1 axons become more similar in response to stimulation, but only for the first novel exposure. (4) For the same axons, stimulation at different sites produces different responses, indicative of novelty detection; the same axons can transmit different information, based on spike timing. (5) The same stimuli at different delay times between repetitions are preferentially encoded in different inter-regional axons. (6) The traditional sequential processing of information in the hippocampus is controlled by considerable feedback activity that is stronger for single-site than three-site stimulation. These short-term responses to stimulation are unique in axonal access and subregional specificity compared to long-term stimulation for weeks that alters the spiking dynamics of the neurons in cultures of unseparated hippocampus (Ide et al 2010) or cortical neurons in a structured PDMS device (Ihle 2022).

Log-log distributions of spike times shift in response to stimulation
Unlike sensory systems where spike rates center around a Gaussian mean that shifts with stimulation (e.g. sensory-visual system (Reich et al 2000)), hippocampal spike times are distributed in a log-log (akin to power law) distribution. This permits a larger dynamic range, characteristic of multiple non-linear systems, such as synaptic release probability, synaptic receptor IPSC to trigger an action potential. Multiple connected neurons evoke avalanches of spiking then go quiet for a while (Beggs andPlenz 2003, Plenz et al 2021). In our system, each subregion displayed these log-log dynamics.

Spontaneous wiring of neurons in the network produces widely disparate axonal spike trains between axons in each subregion in response to similar input stimulation
Several possibilities exist for how axons from the same region could respond to stimulus. (1) They could all respond together because they received the same input, so they should relay some aspect or representation of that input.
(2) At the other extreme, each axon emanates from a single neuron which was most likely not the neuron that was stimulated, so each transmission neuron would receive the input stimulus through a different processing pathway. We observed most of the spikes on an axon within the first 50 ms after the stimulus. Therefore, if we had chosen to integrate responses over this 50 ms, and a simple rate code would find that most axons had similar responses. However, it is questionable whether this is truly the code of transmission. Instead, we chose an integration window of just 3 ms such that most bins had 0 or 1 spike and we could detect precise spike timing coding. At this resolution, we observed low spike train similarities of 0.1-0.2 and not different from before stimulation for axons from the EC into DG, CA3 or CA1. Only axons into and from the CA3 (DG-CA3 and CA3-CA1) showed similarity measures that rise above chance at 0.1-0.2 or 0.3. These levels are still low, consistent with a coding function of pattern separation from DG to CA3. We previously demonstrated evidence for pattern separation in pairs of hippocampal subregions in vitro stimulated by pair-pulse, strongest in the EC to DG (Poli et al 2018). Here we extended these findings of pattern separation by DG in response to TBS at multiple sites. Yassa and Stark (2011) reviewed the convergent evidence that the DG is the primary hippocampal field in humans to accomplish pattern separation. In older adults with amyloid deposition, high resolution fMRI showed hyperfunctional connectivity between EC and DG that associated with impaired mnemonic object discrimination (Adams 2022), suggesting overactivation leading to confusion in pattern separation.

4.4.
Only CA3-CA1 axons become more similar in response to stimulation, but only for the first novel exposure Slice physiologists commonly stimulate many Schaffer collateral axons entering the CA1 from the CA3 as a group with the same extracellular theta burst stimulus to evoke LTP in CA1 (Larson and Lynch 1988). Here we looked for the signs of this coordinated stimulus in the CA3-CA1 axons from upstream excitation in the EC and CA3. A single site of stimulation in EC or three sites evoked a three-fold increase in the coordinated (similarity) spiking of CA3-CA1 axons into CA1 over ten repetitions, suggestive that our theta-burst-like stimulus to the EC reached the CA3 neurons to produce an output from their axons that would be likely to evoke LTP in the target CA1. Surprisingly, stimulation at two sites resulted in a loss of activity and uncoordinated axonal spiking, a drop that was repeated after the three-site stimulation. These declines may be examples of LTD (Malenka and Bear 2004) that persists until the fourth repetition.

For the same axons, stimulation at different sites produced different responses indicative of novelty detection; the same axons can transmit different information, based on spike timing
An important function of the hippocampus is to distinguish between unique inputs using the same circuitry. We applied the theta-burst stimulation at different sites in the EC and saw large differences in spike train similarities in all communicating axons, with especially strong novelty in the CA3-CA1 axons. These findings are consistent with the ability of the hippocampus to recognize novel objects and places (Larkin 2014). From a network perspective, it is interesting that the same axons can carry different information streams, characteristic of a small world network Zwi 2004, Downes et al 2012). That is, different inputs take unique routes through the same subnetworks, yet they transmit their signals to the next subnetwork over the same limited number of axons in a manner characteristic of a small-world network (Poli 2015, Vakilna et al 2021. 4.6. The same stimuli at different delay times between repetitions are preferentially encoded in CA3 or CA1 EC (Kraus et al 2015), CA1 (Kraus 2013), and CA3 (Salz 2016) neurons encode time-related aspects of experience. For rats running for variable times or speeds on a treadmill without access to spatial cues, 'time cells' have been located and recorded in these hippocampal subregions, with the same cells even reporting different time scales (Mau et al 2018). Buzsaki and Tingley (2018) proposed a mechanism of timing involving varying frequencies within the 4-10 Hz theta cycle for phase-amplitude coupling or even counting theta cycles. Sanders et al (2018) proposed that non-spatial aspects of an episodic memory are encoded by a rate remapping of new cell assemblies. Here we found a distinction in preferential coding of 0.2 and 10 s repeated stimuli. The 0.2 s repetitions evoked stronger similarity in the axons early in the circuit from EC to DG and DG to CA3, than the 10 s repetitions. By subregion, 10 s repetitions were less similar in DG axons than the higher levels of similarity of spike patterns in the CA3-CA1 and CA1-EC axons, like the CA1 preference for even longer times (Mankin et al 2012). The similar spike patterns in different axons were likely to activate more strongly a different set of target neurons in downstream subregions and suggested that different cells assemblies would code for differentially timed events. It will be interesting to learn from ongoing studies whether theta oscillations are increased for the 0.2 sec repeats and whether the 10 s repeats are associated with a slower 0.1 Hz oscillation (Mankin et al 2012).

The traditional sequential processing of information in the hippocampus is controlled by considerable feedback activity that is stronger for single-site than three-site stimulation
The sequential feed-forward activation of the hippocampal tri-synaptic loop needs mechanisms to control run away excitation. One clear mechanism is to employ local interneurons in each subregion, but each subregion appears to have only 10%-15% local inhibitory interneurons, although these cells have extensive axonal branching for local control (Pelkey et al 2017). This local GABAergic control of pyramidal neuron activation occurs by either feed-forward (Lawrence and McBain 2003) or feedback directions (Pelkey et al 2017) to limit burst durations. However, much less attention has emerged from inter-regional feedback. CA3 GABAergic feedback to DG hilar neurons disynaptically limits the 'detonation' of DG neurons (Lawrence andMcBain 2003, Scharfman 2007). CA1 GABAergic backpropagation to the CA3 was also documented (Sik et al 1995). But these reports are rare and fail to address the system-wide feedback control of the entire hippocampus. Our reductionist model of the hippocampal subregions provides further evidence for feedback transmission between each of the subregions that varied with stimulus pattern and was largely contemporaneous with feedforward activity (Vakilna et al 2021;Lassers, submitted). Pharmacologic studies are needed to determine the GABAergic inhibitory nature of the timed feedback axons (McKay et al 2013).

Limitations
Many of the limitations of this model hippocampal system are readily evident such as lack of behavioral and other brain modulatory inputs. Our 2D reductionist hippocampus does not model the lamellar organization of the native hippocampus including the layered inputs of the EC into the 3D hippocampus (Marks et al 2020;Hanssen 2023). We have not yet taken advantage of the easy pharmacological access to further define the inhibitory components of the networks. Further, definition of the slow wave oscillations associated with these networks is in progress.

Conclusions
Averaging categorical responses from six independent network cultures enabled the first characterization of subnetwork axonal responses to patterned stimulation in each of the four subregions from a new reverse-engineered hippocampus. Most notable was access to activity in the communicating axons between subregions, which responded to EC stimulation with log-log distributed spike times that differed between subregions. The distinctly dissimilar responses in communicating axons suggest a functional pattern separation in response to stimulation at 1, 2 or 3 sites in the EC. Repeated stimulation evoked low levels of similar spiking patterns only in the axons from CA3 into CA1 revealing the dynamic plasticity of the hippocampal network. Changing the site of the stimulus evoked new patterns of spiking in the same communicating axons for facile detection of novel stimuli. Changing the time between repeated stimulation shifted spike pattern similarity from the DG for short times to the CA1 for longer times. Finally, we detected robust feedback signaling between each subregion to keep the network responses within a limited range. Further investigation of this control centered around a critical state (Beggs andPlenz 2003, Plenz et al 2021) is ongoing. Together, each group of axons communicating between hippocampal subregions display a rich repertoire of dynamics, reflecting robust plasticity in response to prior activity for largely novel responses.

Data availability statement
The data that support the findings of this study are available upon reasonable request from the authors.