Structure-function dynamics of engineered, modular neuronal networks with controllable afferent-efferent connectivity

A soft lithography mould was made by application of a photoresist pattern on a 4-inch silicon wafer (Siltronix). The wafer was washed consecutively in acetone and isopropanol (IPA) for 1 min each to remove organic contaminants, followed by a plasma clean for 5 min in 100 sccm $\mathrm{O}_{2}$ plasma at 20 kHz generator frequency (Femto Plasma Cleaner, Diener Electronics). A 5 min dehydration bake was conducted at 120 $^{\circ}\mathrm{C}$ to remove moisture. The photoresist mr-DWL5 (micro resist Technology GmbH) was spin-coated onto the substrate in two steps of 3000 rpm for 35 s at 1000 rpm min⁻¹ and 4000 rpm for 2 s at 2000 rpm min⁻¹, respectively, to achieve a final thickness of 5.0 µm (spin150, SPS-Europe B.V.). A soft bake was subsequently conducted, starting at 50 $^{\circ}\mathrm{C}$ for 2 min before gradually increasing the temperature to 90 $^{\circ}\mathrm{C}$ (5 $^{\circ}\mathrm{C\ min^{-1}}$) where it was kept for 2 min. The wafer was left to cool down on the hotplate (5 $^{\circ}\mathrm{C\ min}^{-1}$), before being left at room temperature (RT) to rest for 10 min. A maskless aligner (MLA150, Heidelberg) was used to transfer the MEA design onto the resist with a 405 nm laser at 300 $\mathrm{mJ\,cm}^{-2}$. After exposure, the film was post exposure baked equivalently to the soft bake, before being left to rest at RT for 1 h. The film was subsequently developed in mr-Dev 600 (micro resist Technology GmbH) for 120 ± 15 s before thoroughly rinsed in IPA. The substrate was thereafter hard baked at 120 $^{\circ}\mathrm{C}$ for 30 min.

A second layer of photoresist mr-DWL40 (micro resist Technology GmbH) was spin-coated onto the substrate in two steps of 1200 rpm for 30 s at 500 rpm min⁻¹ and 3500 rpm for 5 s at 1000 rpm min⁻¹, respectively, to achieve a final thickness of 60 µm. A soft bake was consecutively conducted, starting at 50 $^{\circ}\mathrm{C}$ for 5 min before gradually increasing the temperature to 90 $^{\circ}\mathrm{C}$ (5 $^{\circ}\mathrm{C\ min}^{-1}$) where it was kept for 10 min. The wafer was left to cool down on the hotplate (5 $^{\circ}\mathrm{C\ min}^{-1}$), before being left to rest at RT for 1 h. The film was exposed with a 405 nm laser at 600 $\mathrm{mJ\,cm}^{-2}$. After exposure, the film was post exposure baked equivalently to the soft bake, before being left to rest at RT overnight. The film was subsequently developed in mr-Dev 600 (micro resist Technology GmbH) for 300 ± 15 s before thoroughly rinsed in IPA. The substrate was hard baked at 120 $^{\circ}\mathrm{C}$ for 30 min. A profilometer (Dektak 150, Veeco) was used to confirm the intended heights. The finalized mould was eventually silanized using 2 drops of 97% trichloro(1H,1H,2H,2H-perfluorooctyl) silane (Sigma-Aldrich, 448931) in vacuum for 2 h.

Silicon elastomer and curing agent (SYLGARD$\circledR$184 elastomer kit, Dow Corning) was mixed, degassed, and cast in the photoresist mould at a ratio of 10:1. The PDMS was cured in an oven (TS8056, Termaks) at 65 $^{\circ}\mathrm{C}$ for 4 h. The PDMS was peeled from the mould, and cell compartments were cut out with a 4 mm diameter puncher. PDMS debris was removed using scotch tape, and the chips subsequently washed in acetone, 96% ethanol and deionized (DI) water for 1 min each. The chips were eventually left to dry overnight.

The protocol for fabrication of microelectrode arrays was adapted and partly modified from van de Wijdeven et al [29]. 1 mm thick 4-inch borosilicate wafers (100 mm Borofloat33, Plan Optik) were used as substrates for the MEAs. The wafers were washed subsequently in acetone and IPA for 1 min each to remove organic contaminants, before being plasma cleaned for 5 min in 100 sccm $\mathrm{O}_{2}$ plasma at 20 kHz generator frequency. A 2 min dehydration bake was conducted at 100 $^{\circ}\mathrm{C}$ to remove moisture. The photoresist ma-N 440 (micro resist Technology GmbH) was spin coated onto the substrates at 3000 rpm for 42 s at 500 rpm min⁻¹ for a final thickness of 4 µm (spin150, SPS-Europe B.V.). The film was left to rest for 10 min before being soft baked at 95 $^{\circ}\mathrm{C}$ for 5 min. A maskless aligner (MLA150, Heidelberg) was used to transfer the MEA design onto the resist with a 405 nm laser at 1800 $\mathrm{mJ\,cm}^{-2}$. After exposure, the resist was developed in ma-D332/s (micro resist Technology GmbH) for 90 ± 10 s before thoroughly rinsed in DI water. Prior to e-beam evaporation, the substrates were descummed in 100 sccm $\mathrm{O}_{2}$ plasma for 1 min at 20 kHz generator frequency. An adhesive layer of 50 nm titanium was evaporated onto the substrates at 5 Å s⁻¹, followed by 100 nm platinum at 2 Å s⁻¹ (E-beam Evaporator, Pfeiffer Vacuum Classic 500). A lift-off process was subsequently conducted in acetone before the wafers were rinsed in IPA.

Prior to passivation layer deposition, another descum was conducted for 1 min. Next, a 470 nm thick insulation layer of silicon nitride ($\mathrm{Si}_{3}\mathrm{N}_{4}$) was deposited on the substrates using plasma enhanced chemical vapour deposition at 300 $^{\circ}\mathrm{C}$ for 30 min consisting of 20.0 sccm $\mathrm{SiH}_{4}$, 20.0 sccm $\mathrm{NH}_{3}$ and 980 sccm $\mathrm{N}_{2}$ gas (PlasmaLab System 100-PECVD, Oxford Instruments). A second lithographic step was conducted using ma-N 440 according to the same protocol as previously described to define an etch mask. Inductively coupled plasma with 50.0 sccm $\mathrm{CHF}_{3}$, 10.0 sccm $\mathrm{CF}_{4}$ and 7.0 sccm $\mathrm{O}_{2}$ gas for 6.5 min was subsequently used to dry etch the silicon nitride above the electrodes and contact pads (Plasmalab System 100 ICP-RIE 180, Oxford Instruments).

A thin layer of platinum black was eventually electrodeposited on the electrodes (PP-Type Wafer Plating Electroplating Laboratory System, Yamamoto). The electrolyte bath consisted of aqueous 2.5 mmol chloroplatinic acid ($\mathrm{H}_{2}\mathrm{PtCl}_{6}$, 8 wt% $\mathrm{H}_{2}\mathrm{O}$, Sigma-Aldrich, 262587), in which the wafers were partially submerged using a custom-built wafer holder so that the liquid only covered the electrodes and not the contact pads (illustrated in figure S3(B)). A Red Rod REF201 Ag/AgCl electrode (Hatch) was used as reference, and a platinized titanium plate as counter electrode. A paddle agitator was used at 60.0 rpm to reduce diffusion thickness, and the temperature was kept constant at 30 $^{\circ}\mathrm{C}$ by an external temperature controller. A Palmsens 4 potentiostat (Palmsens) was used in chronoamperometric mode with a constant voltage of −0.4 V for 3 min for all depositions. Eventually, the substrates were diced into 48 × 48 mm square MEAs with a wafer saw (DAD323, DISCO). The photoresist masks were subsequently stripped off in acetone, followed by a rinse in 96% ethanol and DI water. To remove hardened photoresist and oxidize the top few nanometres of the silicon nitride layer into silicon dioxide ($\mathrm{SiO}_{2}$), the surface was plasma cleaned for 10 min in 160 sccm $\mathrm{O}_{2}$ at 32 kHz generator frequency.

The microfluidic chips were bonded irreversibly to either MEAs or precleaned (acetone, 96% ethanol and DI water) glass coverslips (24 × 24 mm Menzel-Gläser, VWR International). Both microfluidic chips and MEAs/coverslips were plasma cleaned with $\mathrm{O}_{2}$ plasma for 1 min in 200 sccm $\mathrm{O}_{2}$ plasma at 40 kHz generator frequency. Directly after, the PDMS chips were pushed gently onto the MEAs/coverslips. To facilitate alignment of the microfluidic chips to the MEAs, two drops of 70% ethanol were placed in between the PDMS chip and the MEA, and alignment was performed under a stereomicroscope (Nikon SMZ800). For both coverslips and MEAs, bonding was finalized on a hotplate at 100 $^{\circ}\mathrm{C}$ for 1 min followed by 5 min at RT under gentle pressure. To remove traces of ethanol from the platforms, three subsequent washes in DI water were conducted with 10 min intervals. Finally, all devices were filled with DI water within 10 min after bonding to maintain hydrophilicity and sterilized in UV light overnight in a biosafety cabinet.

After sterilization, DI water was replaced by 0.1 mg ml⁻¹ Poly-L-Ornithine solution (PLO) (Sigma-Aldrich, A-004-C) and the chips incubated at 37 $^{\circ}\mathrm{C}$ and 5% $\mathrm{CO}_{2}$ for 30 min. Subsequently, the PLO was replaced by fresh PLO, and the chips incubated at 37 $^{\circ}\mathrm{C}$ and 5% $\mathrm{CO}_{2}$ for another hour. Next, the chambers were washed three times with Milli-Q (MQ) water, before being filled up with laminin solution consisting of Leibovitz-15 Medium (Sigma-Aldrich, L5520) supplemented with 45.6 mg ml⁻¹ sodium bicarbonate (Sigma-Aldrich, S6014) and 16 $\mu\mathrm{g}$ ml⁻¹ natural mouse laminin (Gibco™, 23017015) and incubated at 37 $^{\circ}\mathrm{C}$, 5% $\mathrm{CO}_{2}$ for 30 min. Subsequently, the laminin solution was replaced with fresh laminin solution, and the chips incubated for another 2 h. Laminin has been shown important for neuronal attachment and neurite guidance both in vivo [30] and in vitro [31–34], and was used to ensure neuronal attachment and network formation. During all coating steps, a hydrostatic pressure gradient was created between the chambers by filling one chamber fuller than the other to assure proper flow of coating solution through the microfluidic channels.

An astrocytic monolayer was established prior to plating of neurons to increase neuronal viability [35–37] and synaptic efficacy [38]. The laminin solution was replaced by astrocyte media consisting of DMEM, low glucose (Gibco™, 11885084) supplemented with 15% Fetal Bovine Serum (Sigma-Aldrich, F9665) and 2% Penicillin-Streptomycin (Sigma-Aldrich, P4333). Rat astrocytes (Gibco™, N7745100, astrocyte purity above 80%) were consecutively plated at a density of 100 cells mm⁻², i.e. 2000 cells per microchamber in the two-nodal devices and 4000 cells in the single-nodal devices. The astrocyte to neuron ratio was chosen based on previous optimizations [29], ensuring that the astrocytes did not cover up the entire surface and hence shielded the microelectrodes from the neuronal activity due to uncontrolled proliferation or astrogliosis [39–41]. The astrocyte to neuron density in the rat cortex in vivo has been shown to be between 1:2 and 1:3 [42]. While neurobasal media reduces astrocyte proliferation compared to the FBS supplemented astrocyte media [43, 44], the astrocytes have still been shown to continue proliferation during maturation to recapitulate the ratio seen in vivo [45].

After two days of expansion, the astrocyte media was replaced with neuronal media consisting of Neurobasal Plus Medium (Gibco™, A3582801) supplemented with 2% B27 Plus (Gibco™, A358201), 1% GlutaMax (Gibco™, 35050038) and 2% Penicillin-Streptomycin (Sigma-Aldrich, P4333). Additionally, Rock Inhibitor (Y-27 632 dihydrochloride, Y0503, Sigma-Aldrich) was added at a concentration of 0.1% during plating to increase survival rate. Pen-strep was not included in the microfluidic chips used for viral transductions until 4 DIV to avoid potential interference with transduction efficacy. Rat cortical neurons from Sprague Dawley rats (Gibco, A36511, neuronal purity: 96%) were plated at a density of 1000 cells mm⁻², equalling 20 000 cells per chamber in the two-nodal platforms and 40 000 cells in the controls. This plating density was chosen based on previous studies indicating that at least 250 cells mm⁻² are required for networks to establish synchronized activity [46], and that higher densities can lead to more complex activity dynamics [47]. Half the cell media was replaced with fresh cell media 4 h after plating, and again after 24 h. From here on, half the cell media was replaced every second day until the cultures were ended at 28 DIV.

To assess the efficacy of the microfluidic design at promoting unidirectional axonal growth, neurons in two-nodal devices on glass coverslips (n = 7 for each condition) were virally transduced with an AAV 2/1 serotype loaded with either pAAV-CMV-beta Globin intron-EGFP-WPRE-PolyA or pAAV-CMV-beta Globin intron-mCherry-WPRE-PolyA. This was done to ubiquitously express different fluorescent proteins in the two neuronal populations, aiding longitudinal monitoring of the structural connectivity in the microchannels during development. Transductions were started one day after plating. A hydrostatic pressure gradient was used during the transductions to hinder viral leakage between the two nodes. The capacity of the microfluidic design at keeping the liquid content in the two nodes separated is presented in figure S4. 3/4 of the medium in the presynaptic node was removed, and AAV 2/1 viruses with a CMV promoter encoding GFP were added at a titer of approximately $5e^2$ viruses/cell ($1e^7$ viruses per chamber). The cells were incubated at 37 $^{\circ}\mathrm{C}$ and 5% $\mathrm{CO}_{2}$ for 3 h. Next, the presynaptic chamber was filled up with fresh neuronal medium, and 3/4 of the medium in the postsynaptic node was removed. AAV 2/1 viruses with a CMV promoter encoding mCherry were furthermore added to the postsynaptic node at a titer of $5e^2$ viruses/cell. The postsynaptic node was subsequently filled all the way up again after 3 h. The devices were finally left unperturbed for 48 h. Both viral constructs yielded a consistently high transduction efficacy, with no significant difference in protein expression levels between the constructs (figure S5). While a small amount of cross-contamination of viruses was observed across the nodes, the described approach was found to yield the best trade-off between transduction efficacy and cross-contamination. Viral vectors were prepared in-house at the Viral Vector Core Facility, NTNU.

Images were acquired using an EVOS M5000 microscope (Invitrogen) using DAPI (AMEP4650), CY5 (AMEP4656), GFP (AMEP4651) and TxRed (AMEP4655) LED light cubes and Olympus UPLSAP0 4X/0.16 NA and 20x/0.75 NA lenses. Post-processing of images was conducted in ImageJ with the Fiji plugin to change the selected colour channels and add scale bars.

For immunocytochemistry, cells were plated in µ-slide 8 Wells (Ibidi, GmbH, 80826) following the same protocol as for the microfluidics. Fixation was done using glyoxal solution based on the protocol by Richter et al [48]. The solution consisted of 71% MQ water, 20% ethanol absolute (Kemetyl, 100%), 8.0% Glyoxal solution (Sigma-Aldrich, 128465) and 1% acetic acid (Sigma-Aldrich, 1.00063). Cell medium was replaced with the fixative and the cells incubated at RT for 15 min, before washing the wells with phosphate-buffered saline (PBS, Sigma-Aldrich, D8537) 3 times for 5 min each. Permeabilization was conducted using 0.5% Triton-X (Sigma-Aldrich, 1086431000) diluted in PBS for 5 min, before the wells were washed 2 consecutive times in PBS for 5 min. Next, blocking solution consisting of 5% goat serum (Abcam, ab7481) diluted in PBS was added to the wells, and the cells were left to incubate at RT on a shaking table at 30 rpm for 1 h. The blocking solution was thereafter replaced by primary antibodies mixed with 5% goat serum in PBS. A full overview of primary antibodies and their respective concentrations are listed in table 1. The µ-slides were subsequently placed on a shaker table at 30 rpm at 4 $^{\circ}\mathrm{C}$ overnight. The following day, secondary antibody solution was prepared consisting of 0.2% secondaries and 5% goat serum diluted in PBS. Prior to mixing, the secondary antibodies were centrifuged at 6000 rpm for 15 min to remove precipitates. Next, the primary antibody solution was removed from the wells, and the wells washed three times with PBS for 5 min each. Subsequently, secondary antibody solution was added to the wells, and the cells incubated at RT on a shaker table at 30 rpm for 3 h in the dark. After this, the secondary antibody solution was replaced by solution consisting of the nuclear stain Hoechst (Abcam, ab228550) at a concentration of 1/2000 diluted in PBS and the cells were incubated at a shaking table at RT for 30 min. Finally, the cells were once again washed three times in PBS for 5 min each, then once in MQ water for 5 min, before the µ-slides were eventually filled halfway up with MQ water for imaging.

Imaging was conducted using a Zeiss 800 Airyscan Confocal Laser Scanning Microscope (CLSM) with Zeiss Zen Blue software. The microscope was equipped with a LSM 800 Laser Module URGB with diode lasers 405 nm (5 mW), 488 nm (10 mW), 561 nm (10 mW) and 640 nm (5 mW). A Zeiss Axiocam 503 color camera, 2.8 Mpix, 38 fps (color) was furthermore used for image acquisition.

For the full protocol, see the supplementary materials.

The average impedance of the electrodes was measured to 34.5 ± 10.9 $\mathrm{k}\Omega$ in 0.1% PBS Solution using a MEA-IT60 System (Multichannel Systems). Recordings were conducted using a MEA2100 workstation (Multichannel Systems) at a sampling rate of 25 000 Hz. A temperature controller was used to maintain temperature at 37 $^{\circ}\mathrm{C}$ (TC01, Multichannel Systems). A 3D-printed plastic construct covered by a gas-permeable membrane was used to keep the cultures sterile during recordings. Prior to recordings, the networks were allowed to equilibrate for 5 min at the recording stage. Each recording lasted for 15 min, and recordings were always conducted on the day following media changes.

Stimulations were performed at 27 DIV using a train of 60 spikes at ±800 mV (positive phase first) of 200 µs duration with an interspike interval of 5 s. Prior to stimulation, a baseline recording of 5 min was conducted to identify the most active electrode close to the centre of the chamber measured in number of spikes/s, which was subsequently chosen for stimulation. For the two-nodal platforms, the most active electrode in the postsynaptic node was stimulated first, followed by the most active electrode in the presynaptic node. The stimulation parameters were chosen in line with previous literature [11, 49–51].

All data analysis, including filtering and spike detection, was conducted in Matlab R2021b using a combination of borrowed and custom-made scripts. Raw data was filtered using a 4th order Butterworth bandpass filter, removing high frequency noise (above 3000 Hz) and low frequency fluctuations (below 300 Hz). A notch filter was used to remove 50 Hz noise caused by the power supply mains. Both filters were run using zero-phase digital filtering with the Matlab function filtfilt to avoid changes in the relative position of the detected spikes. Spike detection was conducted using the Precise Timing Spike Detection (PTSD) algorithm developed by Maccione et al [52]. The threshold was set to 11 times the standard deviation of the noise, the maximum peak duration to 1 ms and the refractory time to 1.6 ms.

Burst detection was subsequently performed using the logISI approach developed by Pasquale et al [53], with a minimum of 4 consecutive spikes set as a threshold for a burst to be detected. The hard threshold for the interspike interval both within and outside of the burst core was set to 100 ms. This number was chosen based on the original manuscript reporting the method [53], as well as manual inspection of the burst detection performance. For analysis of spontaneous network activity, network bursts were furthermore detected using the logIBEI approach [54], with at least 20% of all active electrodes in the network required to participate for the activity to be classified as a network burst. Network synchrony was measured as the median number of active electrodes participating in each network burst during a recording.

For each network burst, the summed-up spiking activity of each node was binned into 20 ms bins, and the resulting tuning curves of the two nodes were plotted. If a network burst was detected, but only one of the nodes had more than 30 detected spikes within the time span of the network burst, the network burst was evaluated as contained within that individual node. If more than 30 spikes were detected within both nodes, the network burst was considered to have propagated between the nodes. If the network burst was found to spread between the two nodes, the temporal position of the peaks of the two tuning curves were compared to evaluate whether the network burst spread from the pre- to the postsynaptic node or vice versa. Only response delays between 20 ms and 240 ms were included in the analysis as delays shorter or longer than this made it difficult to infer directionality or made it unlikely for the activity to be part of the same network burst, respectively. This delay time is also consistent with the propagation times reported by others for in vitro neuronal networks [55]. A raster plot and plot of the binned firing rate showing the detection and analysis of the spread of networks bursts are shown in figures 2(A) and (B). Resulting tuning curves of a network burst spreading from the pre- to the postsynaptic node, and vice versa, are shown in figure 2(C).

Zoom In Zoom Out Reset image size

Stimulation data was run through the SALPA filter developed by Wagenaar and Potter [56], and each stimulated time point was followed by 15 ms blanking to avoid stimulation artifacts being detected as spikes. Peristimulus time histograms were made for each platform by binning the data in the 300 ms following a stimulation into 20 ms time bins, and the average response of the network in each bin for the 60 individual stimulations was plotted. Raster plots were made using a customized version of the SpikeRasterPlot function developed by Kraus [57]. Heatmaps showing 20 ms binned network activity following stimulations were created using a customized version of the Matlab function heatmap developed by Deoras [58].

Spike data for the spontaneously evoked activity was binned into 50 ms time bins, and functional connectivity was evaluated using pairwise mutual information through the information theory toolbox developed by Timme and Lapish [59]. Graph theoretical metrics were calculated using the Brain Connectivity Toolbox developed by Rubinov and Sporns [60]. These were furthermore used to plot custom designed graphs showing node colour as spike frequency (spikes/s), node size as PageRank centrality, edge colour as the functional connectivity (pairwise mutual information) and node edge colour as the community calculated using the Louvain algorithm [61]. Nodes were positioned according to the relative position of the electrodes on the actual microfluidic MEAs. Average graph metrics were also plotted over time.

For all plots, customized colormaps were created using the Matlab versions linspecer [62] and colorBrewer2 [63], based on the web tool colorBrewer developed by Brewer et al [64].

For statistical analysis, SPSS version 29.0.0.0 was utilized. Generalized Linear Mixed-Effect Models (GLMMs) were used to allow for comparison of the repeated measures. The network type was used as a fixed effect. In addition to treating the controls and two-nodal networks as separate network types, the two-nodal networks were also split into the pre- and postsynaptic node for additional analysis of the difference in their characteristics. DIV was used as a random effect. The various network features were used as targets. Furthermore, either a gamma distribution with a log link function or a normal distribution were chosen with the linear model based on the fit with the data distributions. The quality of the model fit was evaluated using the skewness of the data distribution, the Akaike information criterion and an evaluation of the predicted versus observed values. Sequential Bonferroni adjustment was used for multiple comparison.