Impact of the entorhinal feed-forward connection to the CA3 on hippocampal coding

Each sub-region of the hippocampus plays a critical computational role in the formation of episodic learning and memory, but studies have yet to show and interpret the individual spiking dynamics of each region and how that information is passed between each subregion. This is in part due to the difficulty in accessing individual communicating axons. Here, we created a novel microfluidic device that facilitates network growth of four separated hippocampal subregions over a micro-electrode array. This device enabled monitoring single axons over two electrodes so direction of spike propagation in interregional communication could be ascertained. In this in vitro hippocampal study, we compared spiking dynamics across two novel four-compartment device architectures: one with four sets of axon tunnels between subregions that excluded the perforant pathway from EC-CA3, and one with five sets of axon tunnels that included the EC-CA3 connection. We found 30-90% faster feed-forward firing rates (shorter interspike intervals) in axons in the five-tunnel model with 35-75% slower bursting dynamics (longer interburst intervals) compared to the four-tunnel model. Comparing the percentage of spikes in bursts between array designs, the five-tunnel architecture showed that the CA3-CA1 and CA1-EC axons had more spikes in bursts than the four-tunnel counterpart suggesting more structured information transfer. Feedback firing rates were similar between configurations. The faster feed-forward inter-regional spiking in the more natural five-tunnel configuration than with four-tunnel suggests tighter control of spiking and possibly more precise communication between subregions.

Methods
In vitro hippocampal neuronal network culture in a four-chamber device with four-way and five-way microfluidic interconnections

Microfluidic and culturing methods were previously described in more detail in Vakilna et al. (2021) and Chen et al. (2023). In short, we employed a novel four-chamber device containing microfluidic tunnels for axonal communication between chambers. Each chamber contained dissociated neurons from micro dissections of the entorhinal cortex (EC), dentate gyrus and hilus (DG), CA3, and CA1 including the subiculum from postnatal day 4 Sprague Dawley rat pups under anesthesia as approved by the UC Irvine Institutional Animal Care and Use Committee (IACUC). Brain cells were dissociated and plated at 1,000 cells/mm2 for DG, 330 for CA3, 410 for CA1 (including subiculum), and 330 for EC, in order to reflect in vivo neural densities: EC-DG 1:3, DG-CA3 3:1, CA3-CA1 1:1.25, and CA1-EC 1.25:1 (Braitenberg, 1981). Cells were in 10 μL of NbActiv4 medium (Transnetyx BrainBits, Springfield, IL; Brewer et al., 2009b) and were plated into the wells sequentially. After thirty minutes in the incubator to allow for adhesion, 0.8 mL culture medium was added.   The cultures were capped with Teflon sheets (ALA Scientific, Farmingdale NY) and incubated for 21–26 days in humidified 5% CO2 and 9% O2 (Brewer and Cotman, 1989). Half of the medium was changed every 3-4 days. Activity was recorded on days 21-26, 2-5 days after a medium change when the networks had reached maturity. MEA120 glass multielectrode arrays (MEA) equipped with 120 30-μm-diameter electrodes spaced 200 μm served as the substrates for the culture of neuronal networks (Multichannel Systems, Reutlingen, Germany; ALA Scientific, Farmingdale, NY, USA). A custom polydimethylsiloxane (PDMS) device was aligned and attached to the MEA that separates the microelectrodes into four chambers for each cell type. Axon-isolating tunnels connected each compartment. Each PDMS well was 9.7 mm2 by 1-mm high. Each microfluidic tunnel was 3 μm high × 10 μm wide × 400 μm long spaced 50 μm apart.

Two configurations of microfluidic tunnels for axonal communication were compared. In the first configuration, each compartment was connected to the adjacent one to allow feed-forward connection according to Cajal’s original understanding of the trisynaptic loop. Five of the 51 axon tunnels between each compartment were monitored with a pair of electrodes underneath. In the second configuration, two electrodes were under each of 4 of the 67 tunnels between the same subregions, but to instantiate the perforant path (Andersen et al., 1971), axonal tunnels diagonally connected the EC to the CA3 wells. Five of these 11 EC to CA3 tunnels were monitored by pairs of electrodes and two more by a single electrode. From the spike timing delay between the electrode pairs in each tunnel, we determined the direction of axonal potential propagation and categorization of action potentials as either feed-forward or feedback. In both configurations, nineteen electrodes were at the bottom of each subregion.  The substrate was treated with oxygen plasma followed by coating the wells with poly-D-lysine for cell adhesion. Finally, we placed micro dissected dissociated neurons from four-day-old rat hippocampi in the corresponding wells and cultured them in NbActive4 (Brewer 2008) for 18-22 days (Chen et al., 2023).

Multi-electrode array and recording

A Multichannel Systems MEA120 1100 (Multichannel Systems, Reutlingen, Germany) amplifier was used to record spontaneous activity on the 120-electrode microarray. Spontaneous activity was analyzed using MC_Rack software at a sampling rate of 25 kHz at 37o C for 5 min, in humidified 5% CO2, 9% O2 (custom Airgas USA, Santa Ana, CA).  Recordings were initiated several minutes after transfer from the culture incubator to the amplifier, once at least 80% of the tunnels had stable activity. Arrays with <80% active tunnels or that had poor growth in one of the compartments were rejected for recording.

Spike detection, sorting, and axonal propagation direction

Details on our axonal spike directionality algorithm were previously described in Lassers et al. (2023). Briefly, raw tunnel data sampled at 25 kHz were filtered through Wave_Clus (Chaure et al., 2018), and spike detection and clustering were computed from 5 to 50 S.D. noise and 50.1 to 500 S.D. to ensure the counting and clustering of large axonal spiking. A refractory period of 1.5 ms was specified. Any spike shapes differing by less than three standard deviations from the mean spike shape were included in a single cluster. The large, tolerated deviation was chosen to accommodate different axon-electrode coupling for a single axon on each of the two electrodes. Clustering was used to discard complex spikes from multiple axons in a microtunnel that produce overlapping spike waveforms. Similar to previous analyses of these tunnel devices that demonstrated that ∼63% of tunnels contained only one axon (Narula et al., 2017), single axons were identified by their uniform conduction velocities or spike timing delays. These timing delays were used to generate a normalized matching indexing (NMI) algorithm which was computed for every tunnel using the timing comparison made from the two electrodes spanning each tunnel. All tunnels were the same length of 400 μm except for the EC-CA3 diagonal tunnels that ranged from 400-600 μm. 

Tunnels with NMI > 0.2 (20% of spikes matched between two clusters) were considered valid. This threshold was sufficient for eliminating spurious spike pair correlations during high spike rates.

A histogram of conduction times was generated with thresholding and peak prominence values provided by the MATLAB findpeaks function. Valid delay times were between 0.2 ms and 1 ms if there was a peak at sufficiently fast conduction times. Feed-forward axons were identified by positive delay times and feed-back axons by negative conduction times, the basis for our determination of the feed-forward and feedback directionality of axonal communication. In contrast, raw data from the wells were filtered through Wave_Clus and spikes were detected at the threshold of ±3.5 S.D. noise. This lower threshold than for axons better accommodated the lower signal to noise ratio in the wells. Spike clustering was not needed since single neurons were detected in >90% of cases.

Spike dynamics and probability distributions

Vakilna et al. (2021) showed that the distribution of inter-spike intervals (ISI) and inter-burst intervals (IBI) follow log–log distributions and were visualized as normalized complementary cumulative probability distributions (CCDs) with logarithmically spaced bins (Newman, 2005). A log-transformed linear model of the slope, m, and intercept c was used to fit the CCD after log transformation. A grid search of the local maximum for was used to find the best fits with time limits varied up to 50% with a step size of 5%. A single fit was found for all ISI CCDs over the probability range from 1 to 0.1. The four-tunnel ISI CCDS intervals were different for each tunnel, for both feed-forward and feedback axons. For the five-tunnel design, intervals were from 0.01 to 0.2 s except for CA1-EC feed-forward and feedback axons which were fit for intervals between 0.01 to 0.11 s and 0.01 to 0.09 s respectively, to account for non-linearities in the distribution. Two linear fits were calculated for all inter-burst interval (IBI) CCDs piecewise to account for the “up states” and “down states,” referring to fast and slow bursting, respectively (Vakilna et al., 2021). The minimum time for the up states was used as the maximum time for the down states. We compared the distribution of ISI to a log-normal distribution instead of a log-log to show that the distributions were not log-normal. The burstiness of a neuronal unit was defined as the percentage of tagged spikes that appear in bursts, at least four spikes, each separated by less than 50 ms. Distributions of axonal and somal burstiness were plotted normalized by the number of arrays in each network design and statistical tests were performed to assess changes in burstiness between designs.

Statistics

Data were analyzed with custom MATLAB 2023a scripts. Slopes were compared for significant statistical differences at alpha=0.05 using analysis of covariance (ANCOVA) followed Tukey’s honest significant difference (HSD) test. The burstiness of the data was compared using the nonparametric Kruskal-Wallis test since the distributions of burstiness had significant deviations from the normal distribution to determine if the data came from different distributions at p<0.05. Data were combined and analyzed for nine separately plated networks for the four-tunnel model and six separately plated networks for the five-tunnel model.