Microfluidic platforms for probing spontaneous functional recovery in hierarchically modular neuronal networks Authors: Keita Watanabe, Hideaki Yamamoto, Takuma Sumi, Hakuba Murota, Hironobu Osaki, Kenshiro Kawamoto, Teppei Matsui, Yoshito Masamizu, Shigeo Sato, Ayumi Hirano-Iwata Published: 2026 ABSTRACT The inherent capacity to flexibly reorganize after injury is a hallmark of brain networks, with recent studies suggesting that the functional consequences of damage are strongly influenced by the network’s nonrandom connectivity. However, experimental platforms that enable bottom-up investigations of the structure–function relationships underlying damage and recovery processes remain limited. Here, we used polydimethylsiloxane microfluidic devices to construct hierarchically modular neuronal networks that mimic the topological features of the mammalian cortex. Laser microdissection was employed to selectively sever intermodular connections, enabling controlled damage to either hub or peripheral connections. Damage to hub connections leads to delayed recovery, requiring more than three days for correlations to re-emerge. Contrarily, peripheral damage resulted in faster recovery. Repeated injury to neuronal networks further revealed that recovery occurs through both new pathway formation and restoration of originals. These findings provide mechanistic insights into the intrinsic self-repair capacity of living neuronal networks. Introduction Stroke and traumatic brain injury disrupt motor and cognitive functions, with severe damage often leading to irreversible loss of regional brain functionality1–4. In contrast, depending on the extent and location of the injury, function can sometimes partially recover, highlighting the brain’s capacity for reorganization5–7. For example, in macaque monkeys, ibotenic acid-induced lesions in the primary motor cortex hand area caused motor impairments that recovered with rehabilitation, involving early compensation by the ventral premotor cortex and later compensation by the perilesional primary motor cortex, indicating time-dependent network reorganization8. In addition, rodent studies have shown that lesions in the primary motor cortex impair sensory processing in remote regions, which again can recover through the loss and gradual compensation of inhibitory connections9. The importance of the damage location in the recovery process is underscored by observations that damage to the corticospinal tract, a hub pathway connecting the motor cortex to the spinal cord, results in prolonged motor impairment compared with a stroke directly in the cortex10,11. At a broader scale, computational simulations using a neuronal mass model in structured brain networks derived from diffusion magnetic resonance imaging have shown that the impact of focal lesions depends on network topology, with damage to central hubs disrupting widespread functional connectivity across the brain, whereas lesions in peripheral regions induce more localized effects12. Collectively, these studies indicate that recovery after brain injury involves complex and dynamic reorganization across local and large-scale networks. Recently, in vitro neuronal cultures are increasingly recognized as versatile platforms for elucidating the cellular and network mechanisms underlying spontaneous recovery after injury owing to their precise controllability of the extracellular environment and ease of observation and manipulation13,14. In particular, the integration of cultured neurons with bioengineering tools has gained attention for guiding neuronal growth and reproducibly constructing well-defined networks with living cells15–20. Among these approaches, microfluidic devices have become widely used because of their ability to define structures precisely and their ease of use, making them an excellent method for patterning neurons. This capability enables real-time monitoring of axonal degeneration and regeneration following injury, providing a robust platform for investigating the mechanisms of damage tolerance and self-repair in living neuronal networks21. To date, several in vitro injury models utilizing microfluidic devices have been established. These include physical injuries induced by vacuum aspiration or compression22,23, chemical injuries caused by the application of neurotoxins24, and optical injuries achieved through laser microdissection25–27. Such approaches have contributed to understanding the effects of localized damage to individual axons and randomly connected neuronal networks. However, despite these advances, an experimental platform that enables systematic evaluation of interactions between network architecture and injury remains to be established, leaving behind a comprehensive understanding of structure–function relationships at the network level. In this work, we used microfluidic devices to constrain the growth of primary cortical neurons in dissociated cultures and fabricated living neuronal networks bearing a hierarchically modular structure—a network structure that is evolutionarily conserved in animal nervous systems28,29. Using laser microdissection, we applied focal damage to microchannels coupling the modules and investigated the changes in spontaneous and evoked neuronal activity by calcium imaging. The experiments show that damage induced in hub and peripheral links has profoundly different effects on activity and its spontaneous recovery. Intriguingly, the recovery process is highly dependent on the topological role of the damaged connections even in this simplified model of biological neuronal networks: while hub damage leads to a slower recovery than that of peripheral damage, it ultimately results in a greater rebound in network synchronization. This restoration is not solely attributable to the re-establishment of the original connections but is also supported by the emergence of alternative functional pathways, indicating a compensatory reorganization of network activity. Results Laser microdissection of engineered neuronal networks To induce localized damage to cultured neuronal networks, we constructed a fully optical system that integrates a 355 nm, nanosecond-pulse laser into an upright microscope (Fig. 1a). The laser beam is directed through mirrors and a dichroic mirror and is focused onto nerve fiber bundles within the microchannels using a 20× objective lens. The structure of the polydimethylsiloxane (PDMS) microfluidic device used to fabricate engineered neuronal networks is illustrated in Fig. 1b. The network features a hierarchical modular architecture, consisting of 2 × 2 groups, each of which is further subdivided into a 2 × 2 array of smaller subgroups30,31. Each subgroup is implemented as a 100 × 100 μm2 through-hole, which we refer to as modules. The through-holes are interconnected by microchannels measuring approximately 6.0 ± 0.9 μm (n = 12 microchannels) in height and 4.5 ± 0.4 μm (n = 12 microchannels) in width. These microchannels restrict soma from entering while permitting the extension of multiple neurites from the two ends. Primary rat cortical neurons were cultured in the microfluidic device and used for recordings at designated days in vitro (DIV) (Supplementary Figs. S1, S2).Fig. 1Overview of the experimental setup.(a) Sketch of the laser microdissection system. (b) Phase-contrast micrograph of an artificial neuronal network at days in vitro (DIV) 10. A microfluidic device was attached to a poly-D-lysine coated coverslip and primary rat cortical neurons were seeded onto the device. (c) Schematic illustrations of the injury experiment (upper panels) and representative fluorescence traces of GCaMP6s before and after injury (lower panels). The red traces represent neurons within the module corresponding to the damaged region. The traces are arranged from bottom to top according to the order of the region of interest (ROI) numbers shown in Fig. 3b. The schematics are not based on experimental data and are included for conceptual clarity. Overview of the experimental setup. (a) Sketch of the laser microdissection system. (b) Phase-contrast micrograph of an artificial neuronal network at days in vitro (DIV) 10. A microfluidic device was attached to a poly-D-lysine coated coverslip and primary rat cortical neurons were seeded onto the device. (c) Schematic illustrations of the injury experiment (upper panels) and representative fluorescence traces of GCaMP6s before and after injury (lower panels). The red traces represent neurons within the module corresponding to the damaged region. The traces are arranged from bottom to top according to the order of the region of interest (ROI) numbers shown in Fig. 3b. The schematics are not based on experimental data and are included for conceptual clarity. Localized damage was induced in the network by irradiating a single laser pulse at the center of the microchannel region. Although the amount of physical disruption depended both on the innate thickness of the neurite bundle (Supplementary Figs. S3, S4) and the number of applied pulses (Supplementary Fig. S5), this procedure selectively truncated the neuronal processes within the microchannels through energy absorption and subsequent ablation32,33. To assess neuronal viability following the damage, we compared the spontaneous activity of the neurons before and after laser irradiation. Prior to damage, the entire network, including the neurons at the target site, exhibited spatially and temporally complex activity that exhibited globally synchronized activity together with asynchronous activity (Fig. 1c, left)30,31,34. Following damage, the globally synchronized activity diminished, and the firing rate of neurons at the damage site decreased. Nevertheless, these neurons continued to exhibit spontaneous activity, indicating that they remained viable (Fig. 1c, right; Supplementary Fig. S6). As a pilot experiment of laser microdissection, laser pulses were delivered to the two microchannels connecting the four modules on the bottom-right (Fig. 2a, designated as Group B) to the rest of the network. Experiments revealed that this intervention functionally isolated the group from the rest of the network, decreasing the degree of neuronal correlation between those in Group B and those in the rest of the network (Fig. 2b). Quantitatively, the correlation between neuron pairs involving Group B (i.e., A–B, B–C, and B–D) decreased after the injury in 94.1% of the pairs (n = 2168 pairs), whereas the value for those not involving Group B (i.e., A–D, C–D, and A–C) was 22.7% (n = 1883 pairs), indicating a selective loss of functional connectivity (Fig. 2c, Supplementary Fig. S4). Not only the fraction of neuron pairs but also the difference in pairwise correlation after the damage was significantly greater for the neuron pairs involving Group B than for those that did not involve Group B (Fig. 2d). These findings indicate that laser-induced damage selectively impairs the functional connectivity of the targeted population, effectively isolating the fraction from the network.Fig. 2Effects of injury on artificial neuronal networks.(a) Left: Schematic illustration of the micropattern structure. The damaged microchannels are colored in red. Right: Fluorescence image of the artificial neuronal network labeled with tdTomato under the Synapsin-1 promoter. The arrows indicate the location of the damage. Neurite bundles transected by focused laser irradiation are marked with red arrowheads. (b) Correlation coefficients between neurons calculated from the estimated firing rates before (bef) and after (aft) the damage. (c) Changes in the correlation coefficients before and after the damage, grouped according to the spatial location of the neurons (n = 2304 pairs). Gray indicates pairs where both neurons belong to nonisolated modules; orange indicates pairs where one neuron belongs to an isolated module. (d) Changes in the correlation coefficients for individual neuron pairs (n = 2304 pairs). The black square dots represent the medians. The color coding corresponds to Panel (c). *p < 0.05 (Mann–Whitney U test). (e) Comparison of neuronal activity between damaged (n = 12 networks from 9 coverslips) and undamaged (control, n = 31 networks from 15 coverslips) conditions. The difference in the number of networks is caused by the undamaged group including control samples taken in a different set of experiments. Left: Mean correlation coefficient across the whole network. Right: Mean firing rate across the whole network. The bars represent the means. *p < 0.05, **p < 0.01, and ***p < 0.001 (Wilcoxon signed-rank test for within-group comparisons; Mann–Whitney U test for between-group comparisons). Effects of injury on artificial neuronal networks. (a) Left: Schematic illustration of the micropattern structure. The damaged microchannels are colored in red. Right: Fluorescence image of the artificial neuronal network labeled with tdTomato under the Synapsin-1 promoter. The arrows indicate the location of the damage. Neurite bundles transected by focused laser irradiation are marked with red arrowheads. (b) Correlation coefficients between neurons calculated from the estimated firing rates before (bef) and after (aft) the damage. (c) Changes in the correlation coefficients before and after the damage, grouped according to the spatial location of the neurons (n = 2304 pairs). Gray indicates pairs where both neurons belong to nonisolated modules; orange indicates pairs where one neuron belongs to an isolated module. (d) Changes in the correlation coefficients for individual neuron pairs (n = 2304 pairs). The black square dots represent the medians. The color coding corresponds to Panel (c). *p < 0.05 (Mann–Whitney U test). (e) Comparison of neuronal activity between damaged (n = 12 networks from 9 coverslips) and undamaged (control, n = 31 networks from 15 coverslips) conditions. The difference in the number of networks is caused by the undamaged group including control samples taken in a different set of experiments. Left: Mean correlation coefficient across the whole network. Right: Mean firing rate across the whole network. The bars represent the means. *p < 0.05, **p < 0.01, and ***p < 0.001 (Wilcoxon signed-rank test for within-group comparisons; Mann–Whitney U test for between-group comparisons). Interestingly, localized damage also induced widespread effects on the entire network. For example, the mean correlation in the damaged network decreased by 59.3 ± 20.5% (n = 12 networks) immediately after the damage (Fig. 2e), independent of the extent of damage (Supplementary Fig. S7). A reduction of 55.6 ± 31.1% was also observed in the mean firing rate. In contrast, when the damage was not applied to the network, no significant changes in either the mean correlation or the mean firing rate were observed between the two recording sessions. The reduction of activity observed in modules distant from the injury site is attributable to a decrease in effective input connectivity to individual neurons, as well as the transection of axons that extend beyond adjacent modules. Additionally, we performed a control experiment in which the laser was irradiated on regions outside the network structure and confirmed that the network-wide effects of laser microdissection are caused by the disruption of intermodular connections rather than nonspecific effects of laser exposure (Supplementary Fig. S8). In summary, laser microdissection, coupled with cell patterning via microfluidic devices, allows targeted disruption of neuronal connectivity and widespread changes in network activity. Effect of the damage sites To investigate how the damage location influences the dynamics in a nonrandom network possessing hierarchically modular connectivity, we applied laser microdissection either to hub connections, which serve as central links, or to peripheral connections that connect terminal nodes. To assess both the impact of damage and the recovery process, spontaneous neuronal activity was recorded for 15 min at five time points: before the injury, immediately after the injury, and on days 1, 3, and 7 after the injury (Fig. 3a).Fig. 3Hub and peripheral damage elicit distinct functional responses.(a) Experimental protocol. Injury was induced by laser microdissection on days in vitro (DIV) 10–11. Spontaneous activity was recorded for 15 min at predefined time points: immediately before and after the damage, and 1, 3, and 7 days after the damage. (b) Representative raster plots of neuronal activity (top) and the corresponding population firing rates (bottom) recorded immediately before the injury, immediately after the injury, and 7 days after the damage. The left panel illustrates the spatial layout of the region of interest (ROI), with neurons located near the damage site highlighted in red. The red arrowhead indicates the damaged hub connection. Neurons are ordered from bottom to top on the basis of ascending ROI numbers. (c) Representative maps of correlation-based connectivity (top rows) and effective connectivity (bottom rows) for networks with damage to the hub or peripheral connections. A control network with no damage is also presented. The correlation map shows only structurally neighboring neuron pairs. Effective connections were estimated using transfer entropy, with values at each time point converted to z-scores. Only connections with z > 1 are visualized. The blue curves indicate connections that existed before the injury, whereas the red curves indicate newly formed connections. Line thickness is scaled according to the z-scored value. (d) Changes in the mean correlation across the entire network. Sample sizes: hub-damaged (n = 32 networks), peripheral-damaged (n = 11 networks), and control (n = 31 networks). *p < 0.05; **p < 0.01; ***p < 0.001; and N.S., not significant (Wilcoxon signed-rank test with Bonferroni correction). (e) Evolution of effective connectivity upon damage. Focusing on neuron pairs that exhibited a z-scored transfer entropy (TE) greater than 1 before the injury, we estimated changes in their mean strength over time on the basis of relative TE (RTE, see Methods). The plots and error bars represent the means and the standard error of the mean (SEM) across networks, respectively. Hub and peripheral damage elicit distinct functional responses. (a) Experimental protocol. Injury was induced by laser microdissection on days in vitro (DIV) 10–11. Spontaneous activity was recorded for 15 min at predefined time points: immediately before and after the damage, and 1, 3, and 7 days after the damage. (b) Representative raster plots of neuronal activity (top) and the corresponding population firing rates (bottom) recorded immediately before the injury, immediately after the injury, and 7 days after the damage. The left panel illustrates the spatial layout of the region of interest (ROI), with neurons located near the damage site highlighted in red. The red arrowhead indicates the damaged hub connection. Neurons are ordered from bottom to top on the basis of ascending ROI numbers. (c) Representative maps of correlation-based connectivity (top rows) and effective connectivity (bottom rows) for networks with damage to the hub or peripheral connections. A control network with no damage is also presented. The correlation map shows only structurally neighboring neuron pairs. Effective connections were estimated using transfer entropy, with values at each time point converted to z-scores. Only connections with z > 1 are visualized. The blue curves indicate connections that existed before the injury, whereas the red curves indicate newly formed connections. Line thickness is scaled according to the z-scored value. (d) Changes in the mean correlation across the entire network. Sample sizes: hub-damaged (n = 32 networks), peripheral-damaged (n = 11 networks), and control (n = 31 networks). *p < 0.05; **p < 0.01; ***p < 0.001; and N.S., not significant (Wilcoxon signed-rank test with Bonferroni correction). (e) Evolution of effective connectivity upon damage. Focusing on neuron pairs that exhibited a z-scored transfer entropy (TE) greater than 1 before the injury, we estimated changes in their mean strength over time on the basis of relative TE (RTE, see Methods). The plots and error bars represent the means and the standard error of the mean (SEM) across networks, respectively. A representative progression of neuronal activity following damage to the hub connection is summarized in Fig. 3b. Immediately after the injury, a reduction in the firing rate and global synchrony was observed. By day 7, however, the synchrony had partially recovered. In particular, neurons near the damage location (shown in red) resumed to coactivate with other neurons. Irregular firing patterns and persistent decreases in activity were observed in some neurons, suggesting that recovery was heterogeneous across the network. Figure 3c compares the changes in network activity following damage to either hub or peripheral connections. While a reduction in correlation coefficients was consistently observed immediately after damage, the timescale of recovery differed depending on the damage site (Fig. 3c, upper panels). To assess this trend in detail, we compared changes in the mean correlation of the entire network under each condition (Fig. 3d). In the case of hub connection damage, the mean correlation decreased by 35.7 ± 19.0% after injury and required approximately three days to recover. Peripheral damage also caused a transient reduction of 33.3 ± 25.4% but returned to the baseline level within one day. Notably, by day 7 after the injury, only the hub-damaged group presented a mean correlation exceeding the baseline level by 48.6 ± 71.4%. These changes were not observed in the control group (without damage), suggesting that damage location plays a critical role in determining both the speed and extent of network recovery. The evaluation of effective connectivity using transfer entropy further revealed that many of the original connections (blue) disappeared immediately after the damage, regardless of the damage site, and that new connections (red) emerged with development (Fig. 3c, lower panels). In the absence of damage, an immediate reduction in neural correlation was not observed; however, a transition to new effective connectivity patterns was consistently observed. Quantitatively, focusing on connections with a z-score > 1, the transfer entropy values immediately decreased after the injury 0.40 ± 0.29-fold (n = 32 networks, hub) and 0.41 ± 0.23-fold (n = 11 networks, peripheral) relative to baseline levels, both of which were significantly lower than those of the control group that were not subjected to any damage (n = 31 networks; Mann–Whitney U test, p < 0.001; Fig. 3e). Interestingly, no significant differences were observed between the hub and peripheral injury groups in the overall trajectory of these connectivity changes (Mann–Whitney U test, p > 0.05), implying a common mechanism of network reorganization across damage locations. Furthermore, despite the large-scale reorganization, certain original connections still remained on day 7 after the injury. For example, stable edges remained within the four modules located in the upper right of the hub-damaged network and the upper left of the peripheral-damaged network in Fig. 3c. This suggests that functionally important core structures may persist even after injury, potentially serving as scaffolds for re-establishing global network connectivity. We note that whereas the immediate postinjury reduction in the effective connectivity is clearly caused by the damage, the changes observed in the subsequent days likely reflect a combination of both damage-induced and developmental factors. Moreover, as effective connectivity evaluates functional information flow rather than direct anatomical connections35, the detection of interactions between spatially distant modules is not unexpected. Mechanisms of spontaneous recovery Finally, to elucidate the underlying mechanisms of this recovery, we used optogenetic tools to stimulate neurons in each module and analyzed changes in the propagation delay of the induced activity. The experiments were performed immediately after spontaneous activity was recorded on each measurement day (Fig. 4a). To perform optogenetic stimulation, neurons were transfected with the red-shifted channelrhodopsin ChrimsonR at DIV 4 using viral vectors. Optogenetic stimulation was applied using patterned red light controlled by a digital mirror device, with the illumination area defined as a rectangle surrounding the neurons in each module. The stimulation sequence was set on the basis of our previous study31, with a duty cycle of 5%, a repetition frequency of 10 Hz, and an exposure duration of 1 s (Fig. 4b). The stimulation across all 16 modules was repeated five times per experiment.Fig. 4The propagation delay transiently increases following damage but recovers over time.(a) Experimental protocol. After each recording session, the network was optogenetically stimulated to assess evoked responses. The other experimental conditions were identical to those described in Fig. 3. (b) Schematic description of the stimulation pattern. Each stimulus consisted of 5 ms light pulses delivered at 10 Hz for 1 s, with a 10 s interval between stimuli. (c) Representative responses to a stimulation in module 2 upon damage to the hub connection are shown at three time points: before damage (bef), immediately after damage (aft), and 7 days after the damage (7d). The blue and gray traces indicate neurons within module 2 and those in other modules, respectively. The red vertical line marks the time point at which any neuron within the stimulated module first exceeded a firing rate of 1 Hz. (d) Propagation delay map of a representative network. The red arrowhead indicates the damaged hub connection, whereas the red square indicates the stimulated module. Propagation delays were spatially mapped onto the somatic locations. The colors represent the average delay, and the transparency represents the response probability. A darker (more opaque) color indicates a higher response probability. (e) Evolution of the average propagation delay. Hub connection damage was applied between modules 2 and 10 (n = 7 networks), whereas peripheral connection damage was applied between modules 3 and 6 (n = 6 networks). Stimulation was applied five times to one of the two modules neighboring the damaged connection, and the time required for the evoked activity to propagate to the other module was averaged for each sample. Plots and error bars represent the mean and the standard error of the mean (SEM), respectively. The propagation delay transiently increases following damage but recovers over time. (a) Experimental protocol. After each recording session, the network was optogenetically stimulated to assess evoked responses. The other experimental conditions were identical to those described in Fig. 3. (b) Schematic description of the stimulation pattern. Each stimulus consisted of 5 ms light pulses delivered at 10 Hz for 1 s, with a 10 s interval between stimuli. (c) Representative responses to a stimulation in module 2 upon damage to the hub connection are shown at three time points: before damage (bef), immediately after damage (aft), and 7 days after the damage (7d). The blue and gray traces indicate neurons within module 2 and those in other modules, respectively. The red vertical line marks the time point at which any neuron within the stimulated module first exceeded a firing rate of 1 Hz. (d) Propagation delay map of a representative network. The red arrowhead indicates the damaged hub connection, whereas the red square indicates the stimulated module. Propagation delays were spatially mapped onto the somatic locations. The colors represent the average delay, and the transparency represents the response probability. A darker (more opaque) color indicates a higher response probability. (e) Evolution of the average propagation delay. Hub connection damage was applied between modules 2 and 10 (n = 7 networks), whereas peripheral connection damage was applied between modules 3 and 6 (n = 6 networks). Stimulation was applied five times to one of the two modules neighboring the damaged connection, and the time required for the evoked activity to propagate to the other module was averaged for each sample. Plots and error bars represent the mean and the standard error of the mean (SEM), respectively. Following hub connection damage, neuron-to-neuron variability in the rise and peak times of the evoked activity increased, but these timings realigned after seven days (Fig. 4c). Moreover, analysis of propagation delays revealed that, prior to damage, responses could spread rapidly across multiple modules, whereas such propagation was suppressed immediately after the damage (Fig. 4d). In particular, the propagation of activity through damaged connections was inhibited. By day 7 after the damage, both the extent and delay of propagation approached the baseline levels, suggesting a recovery of the network’s signal transmission capacity. Next, we investigated the temporal changes in propagation delay around the damaged areas. Specifically, in the hub-damaged condition, modules 2 and 10 were stimulated and the connections between them were damaged, whereas in the peripheral damage condition, modules 3 and 6 were stimulated, and the connections between them were damaged. The mean propagation delays were then evaluated for each condition (Fig. 4e). In the hub-damaged condition, the propagation delay increased from 132 ± 139 ms to 258 ± 138 ms and gradually returned to 121 ± 135 ms by day 7. In the peripheral damage condition, the delay similarly increased from 65.6 ± 74.5 ms to 234 ± 146 ms and returned to 39.2 ± 50.1 ms by day 7. These results show that signal propagation across microchannels is spontaneously restored irrespective of the damage site. Although the propagation delays recover over time, it remains unclear whether this reflects the restoration of the original damaged pathway or the recruitment of alternative paths through other modules. To address this, we conducted a re-lesion experiment in which the same hub connection was damaged again seven days after the initial damage (Fig. 5, n = 14 networks). As a control, a separate group received the first hub damage on day 7 (n = 10 networks).Fig. 5Prior injury increases the tolerance of neuronal networks to subsequent perturbations.Sample sizes: reinjury (n = 14 networks) and control (n = 10 networks). (a) Mean correlation coefficients across the network. **p < 0.01; N.S., not significant (Wilcoxon signed-rank test). (b) Mean correlation coefficients and firing rates around the damaged region. Top: schematic of the damage location (the highlighted area indicates the modules analyzed); bottom: mean correlation coefficient. *p < 0.05; **p < 0.01; and N.S., not significant (Wilcoxon signed-rank test). (c) Change in the correlation coefficients around the damaged region. The center line indicates the median; the box limits are the upper and lower quartiles; and the whiskers indicate 1.5 × the interquartile range. **p < 0.01 (Mann–Whitney U test). Prior injury increases the tolerance of neuronal networks to subsequent perturbations. Sample sizes: reinjury (n = 14 networks) and control (n = 10 networks). (a) Mean correlation coefficients across the network. **p < 0.01; N.S., not significant (Wilcoxon signed-rank test). (b) Mean correlation coefficients and firing rates around the damaged region. Top: schematic of the damage location (the highlighted area indicates the modules analyzed); bottom: mean correlation coefficient. *p < 0.05; **p < 0.01; and N.S., not significant (Wilcoxon signed-rank test). (c) Change in the correlation coefficients around the damaged region. The center line indicates the median; the box limits are the upper and lower quartiles; and the whiskers indicate 1.5 × the interquartile range. **p < 0.01 (Mann–Whitney U test). As presented in Fig. 3d, the damaged networks exhibited an immediate reduction in correlation and the mean firing rate following the initial damage, with both metrics recovering by day 7. Interestingly, following the second lesion on day 7, the networks did not show any significant change in the mean correlation. In contrast, networks subjected to the first damage on the same DIV showed a decrease in the mean correlation by 18.1 ± 29.2%, confirming that the difference was not due to the timing of the initial damage (Fig. 5a). This enhanced damage tolerance in the reinjury group was more pronounced when the metrics were evaluated for the neurons directly neighboring the targeted connections. Specifically, following the injury on day 7, the mean correlation decreased 16.4 ± 25.0% in networks that received a second injury, whereas the value was substantially greater (53.7 ± 25.9%) for those that received their first injury (Fig. 5b). Statistical comparison further revealed that the change in correlation coefficients following the damage on day 7 was significantly greater in the control group (Fig. 5c). The smaller impact of the second lesion therefore suggests that the network was no longer solely dependent on the original pathway but was maintaining synchronization via alternative routes established during recovery. This is further supported by the gradual regrowth of neurites over the 7-day period (Supplementary Fig. S9), which paralleled the recovery of correlation-based connectivity (Fig. 3c), effective connectivity (Fig. 3e), and propagation delays (Fig. 4e). In summary, the results above point to two key mechanisms underlying spontaneous functional recovery in neuronal networks: (1) reconnection of original pathways, supported by a modest reduction in the mean correlation following the second injury, and (2) emergence of alternative routes that compensate for the lost connections, supported by the increased tolerance in the reinjury groups. The increased tolerance in the reinjury groups suggests that network reconfiguration, rather than exact restoration, plays a dominant role in recovery. Discussion We fabricated cultured neuronal networks with a hierarchically modular structure and selectively ablated specific intermodular connections using focused laser pulses. By combining calcium imaging with optogenetic stimulation, we quantitatively tracked changes in spontaneous and evoked activity. Our findings revealed that (i) global network synchronization in the cultured neuronal networks, initially disrupted by the damage, gradually recovered over time, (ii) the rate and extent of functional recovery varied depending on whether the disrupted connections were hub or peripheral, and (iii) recovery was mediated not only by reconnection in the original pathways but also by the activation of alternative pathways. The third point suggests that recovery in neuronal networks involves functional compensation through reconstruction rather than a simple restoration of structure and function, providing complementary insights to the findings reported in animal and modeling studies36–39. The approach presented herein enables bottom-up investigations of how physical damage impacts function in biological neuronal networks, providing insight into the mechanisms that underlie robustness and functional recovery in complex networks. Microfluidic devices enable precise control of network connectivity in cultured neurons, uniquely positioning them as powerful tools to constructively investigate how structural features of neuronal networks influence their responses to perturbations. Various injury modalities—such as traction, compression, suction, laser ablation, and chemical treatments—have been shown to be compatible with microfluidic platforms to study axonal regeneration, axonal degeneration, neuroprotection, neuron–glia interactions, and other neural responses to injury22–24,26,27,40. Our work extends these approaches to investigate injury responses in networks bearing a biologically plausible modular organization, clarifying the structural role of damaged regions and their influence on global network dynamics. Previous in vitro studies have reported that localized injury to neuronal networks leads to a temporary reduction in the firing rate and synchrony across the entire network, followed by partial recovery over the course of several hours to days25,41. In addition, a study combining experimental data with mathematical modeling suggested that modular organization may contribute to redundancy in signal transmission and increase damage tolerance36. In this study, we found that recovery of network activity differed depending on the location of the damage in the hierarchically modular network: disruption of hub connections led to delayed recovery and prolonged desynchronization of the entire network. This indicates that the properties of damaged regions critically influence the recovery process in structured networks of biological neurons. More broadly, such effects have also been observed in other complex networks, such as power and urban transportation networks42,43. Our results are consistent with previous in vivo studies44–46, which have shown that damage to central structures in brain networks disrupts both structural and functional integration of the entire network. In particular, the widespread changes in network activity observed after damage to hub connections in our experiments closely match the functional impairments associated with injury to the internal capsule, a clinically important, highly centralized white matter structure. The internal capsule is a key structure in the central nervous system where the main conductive pathways connecting the cerebral cortex to the brainstem and spinal cord converge, and even small, localized lesions are known to affect diverse functional domains, including motor, sensory, and cognitive functions47–49. Furthermore, in the context of postinjury recovery, the formation of alternative pathways using residual structures is known to contribute substantially to functional recovery, alongside the regeneration of the originally damaged tracts37–39, further highlighting the strong parallels to our in vitro observations. In the present work, optogenetic stimulation was employed as a diagnostic tool to assess functional connectivity and signal propagation, as previous studies have suggested that controlling network-level plasticity in neuronal cultures through external stimulation is challenging31,50. Nevertheless, in the mouse visual cortex, repetitive optogenetic stimulation could be used to imprint cell assemblies51. Furthermore, studies in both human and animal models have shown that activity-dependent interventions can facilitate recovery through network reorganization5 and recruitment of alternative pathways39. Future investigations of stimulation protocols during the recovery process may provide insights for investigating similar effects in vitro. Damage to neuronal networks is common in many neurological disorders, and understanding its consequences is essential for both basic and applied research. Therefore, it would be interesting, e.g., to perform in-situ calcium imaging during laser microdissection, as mechanical stimulation has been reported to induce calcium waves in astrocytic networks52,53, which are important for triggering acute neuroprotective responses and coordinating synaptic activity. Our platform may not only contribute to the understanding of damage-induced functional impairment and the recovery process but also lead to advances in the treatment of neurological disorders, the development of regenerative medicine, and the design of damage-resistant engineering systems54. Methods Cell culture All animal experiments and gene introduction procedures were performed with the approval of the Tohoku University Center for Laboratory Animal Research (2020AmA-001) and the Tohoku University Center for Gene Research (2019AmLMO-001). Primary rat cortical neurons were prepared from embryonic day 18 rat embryos following previously published protocols15,55,56. Before cell seeding, a microfluidic device was attached to a poly-D-lysine coated coverslip and filled with plating medium [minimum essential medium (11095-080, Gibco) supplemented with 5% fetal bovine serum and 0.6% D-glucose]. To remove air bubbles from the microchannels and modules, the device was placed in a vacuum chamber under reduced pressure. The medium was then replaced with fresh plating medium, and the coverslip was incubated in a humidified CO2 incubator (37 ∘C, 5% CO2) for at least 3 h. Rat cortical neurons were seeded onto coverslips at a density of 4.2 × 104 cells cm−2, incubated for 3 h, and subsequently transferred to a culture dish containing 5 mL of Neurobasal Plus medium [Neurobasal Plus (A3582901, Gibco) supplemented with 2% B-27 Plus Supplement (A3582801, Gibco) and 0.25% GlutaMAX-I (35050-061, Gibco)]. At DIV 4, half of the medium was removed, and adeno-associated virus (AAV) vectors encoding the genetically encoded calcium indicator GCaMP6s (#100843-AAV1, Addgene) and the red-shifted channelrhodopsin ChrimsonR (#59171-AAV1, Addgene) were applied separately. The transduction efficiency of the AAV was estimated to be 83.3% (evaluated with GCaMP6s; n = 150 neurons from n = 3 coverslips). On DIV 5, 2.5 mL of fresh Neurobasal Plus medium was added. On DIV 8, half of the culture medium was replaced with fresh Neurobasal Plus medium, and 50 μL of penicillin–streptomycin (P4333, Sigma) was added to prevent contamination. Laser microdissection A system for inducing localized nerve damage was constructed by integrating a passive Q-switched solid-state laser (FTSS-355-50, CryLas; wavelength: 355 nm, pulse width: 1.1 ns; repetition frequency: 5 Hz) into an upright microscope (BX-51WI, Olympus). The laser beam was introduced into the optical path via a UV-compatible beam expander (BE03-355-3X, Thorlabs) and a 355-nm compatible dichroic mirror, and then focused onto a bundle of nerve fibers inside a microfluidic chamber using a 20 × water immersion objective (UMPLFLN 20 × W, Olympus; numerical aperture NA = 0.5). The laser spot size estimated from ablation marks on a glass substrate was 1.2 ± 0.1 μm (Supplementary Fig. S10), approximately 1.4 times the theoretical spot size determined by the diffraction limit. A neutral density filter (NDHN-U100, Sigma Koki) was used to adjust the laser power, with the final power at the sample surface set at 0.03 μW. A mechanical shutter (SSH-25RA, Sigma Koki) was used to isolate a single laser pulse by setting the open time to 0.2 s. Device fabrication The microfluidic device was fabricated according to a previously reported protocol31,55. The master mold was fabricated by patterning two layers of SU-8 photoresist using photolithography. First, to form the first layer containing the module and microchannel structures, SU-8 3005 (Kayaku Advanced Materials) was spin-coated on a 2-inch silicon wafer at 3000 rpm. This mold was soft baked at 65 ∘C for 1 min, followed by 95 ∘C for 5 min, and then UV light was irradiated through a photomask using a mask aligner (MJB4, Suss). This was followed by postbaking at 65 ∘C for 1 min and 95 ∘C for 3 min, development with an SU-8 developer, and washing with 2-propanol. For the second layer, SU-8 3050 was spin-coated at 1500 rpm, and baked at 65 ∘C for 1 min and 95 ∘C for 30 min. A photomask was used to form only the module pattern on the second layer, followed by photolithography. The sample was postbaked at 65 ∘C for 1 min and 95 ∘C for 5 min. The development process was the same as that for the first layer, and finally, the mold was completely cured by baking at 60 ∘C for 3 h. The master mold was used to fabricate a microfluidic device by replica molding. Sylgard 184 (Dow Corning) was mixed with a 7.5:1 weight ratio of base and curing agent, degassed in a vacuum chamber, poured onto the master mold, and cured by heating at 70 ∘C for 2 h. After curing, the PDMS was peeled from the mold, ultrasonically cleaned in 100% ethanol and deionized water, and then sterilized by exposure to ultraviolet light for more than 30 min. Calcium imaging Calcium imaging recording was first performed on DIV 10 or 11, immediately before injury treatment. Subsequent recordings were performed under identical conditions immediately after the injury treatment and on days 1, 3, and 7 after the injury. For recording, the coverslip with micropatterned neurons was transferred to a glass-bottom dish (3960-03535, Iwaki) containing 1 mL of culture medium and 1 mL of fresh Neurobasal Plus medium. All subsequent cultures and imaging were performed in the same dish. Culturing was continued until the final recording on day 7 after the injury. Following the recording on day 3 after the injury, half of the medium was replaced with fresh medium. Fluorescence intensities were acquired using an inverted microscope (IX83, Olympus) equipped with a 10 × objective lens (UPLXAPO, Olympus; NA, 0.40), an LED light source (Lambda HPX, Sutter Instrument), a scientific CMOS camera (Zyla 4.2P, Andor), and an incubation chamber (Tokai Hit) with integrated temperature and CO2 control. A field of view measuring 2000 × 2069 μm was imaged at a resolution of 1450 × 1500 pixels, providing an effective pixel size of 1.38 μm per pixel. During imaging, the CO2 concentration in the stage area was maintained at 5% using a CO2 controller, and the temperature was maintained at 37 ∘C. All recordings were performed using Solis software (Andor). Spontaneous neuronal activity was recorded at 20 frames per second, whereas responses to light stimulation were recorded at 100 frames per second. Optogenetic stimulation Stimulation of the neuronal network was performed by illuminating patterned light to neurons expressing the red-shifted channelrhodopsin ChrimsonR. The red light was emitted from a red LED light source (Solis 623-C, Thorlabs) and projected onto specific regions of the neuronal network using a digital micromirror device (DMD; Polygon400-G, Mightex). Stimulation pattern design and control were carried out using PolyScan2 software (Mightex). Data analysis To extract neuronal activity from the calcium imaging data, we manually selected 32 neurons from a single network (two neurons per module). The neurons were selected based on the following criteria: (1) clearly defined somatic boundaries without overlap with neighboring cells, and (2) detectable calcium responses throughout the 15 min recording period. Regions of interest (ROIs) were manually defined around each neuronal soma using Fiji software (NIH). As reliable tracking of individual neurons across recording sessions was technically challenging, the neurons selected as ROIs were not necessarily identical across sessions conducted over multiple days. The fluorescence intensity F(t) at each time point t was calculated as the mean pixel value within the ROI. For neurons with poorly defined somatic boundaries, we set a circular ROI with a diameter of approximately 15 μm, corresponding to the typical size of the soma. The relative fluorescence intensity (RFU, ΔF/F) for each cell was computed using the equation ΔF/F = (F − F0)/F0, where F represents the fluorescence intensity at time t and F0 denotes the baseline intensity. The resulting RFU time series was converted into instantaneous firing rates r(t) using CASCADE, a deep learning-based algorithm for spike inference from calcium dynamics57. The CASCADE model was trained on a dataset consisting of simultaneous patch-clamp and calcium imaging recordings reported in the literature. To ensure compatibility with the trained model, the calcium signals obtained in this study were adjusted to match the same frame rate, noise level, and RFU amplitude (RFU = 1) as those used during model training. Instantaneous firing rates inferred by CASCADE were used for subsequent analyses. This approach enables the removal of signal artifacts caused by the rise and decay time constants inherent to the calcium indicator. The population firing rate at each time frame was obtained by averaging r(t) over all cells. Neuronal correlation The functional correlation between neurons i and j, CCij, was quantified using the Pearson correlation coefficient: 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{CC}}}_{ij}=\frac{{\sum }_{t}[{r}_{i}(t)-{\overline{r}}_{i}][{r}_{j}(t)-{\overline{r}}_{j}]}{\sqrt{{\sum }_{t}{[{r}_{i}(t)-{\overline{r}}_{i}]}^{2}}\sqrt{{\sum }_{t}{[{r}_{j}(t)-{\overline{r}}_{j}]}^{2}}},$$\end{document}CCij=∑t[ri(t)−r¯i][rj(t)−r¯j]∑t[ri(t)−r¯i]2∑t[rj(t)−r¯j]2,where ri(t) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{r}}_{i}$$\end{document}r¯i represent the instantaneous and time-averaged firing rates of neuron i, respectively. The mean correlation for a network was calculated by taking the mean of CCij across all the neuron pairs, excluding the diagonal elements. Propagation delay To evaluate the propagation delay of neuronal activity between modules, we analyzed each cell’s response to optical stimulation using the following procedure. First, for each cell in the stimulated module, a firing event was detected when the instantaneous firing rate exceeded 1 Hz. The time point at which the firing rate first exceeded this threshold was defined as the reference time. The analysis time window was set to 1 s, spanning from 0.5 s before the onset of stimulation to 0.5 s before its offset. For cells in other modules, the response firing time was defined as the first point after the reference time at which the firing rate exceeded 1 Hz. The propagation delay time was then calculated as the time difference between the response firing time and the reference time. Effective connectivity Causal interactions between neurons were quantified using generalized transfer entropy (GTE)58,59. Given spike trains of two neurons, X = {xt} and Y = {yt}, the amount of information transferred from neuron Y to neuron X is evaluated as: 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{TE}}}_{Y\to X}=\sum P({x}_{t+1},{{{{\bf{x}}}}}_{t}^{(k)},{{{{\bf{y}}}}}_{t+1}^{(k)}){\log }_{2}(\frac{P({x}_{t+1}| {{{{\bf{x}}}}}_{t}^{(k)},{{{{\bf{y}}}}}_{t+1}^{(k)})}{P({x}_{t+1}| {{{{\bf{x}}}}}_{t}^{(k)})}),$$\end{document}TEY→X= ∑P(xt+1,xt(k),yt+1(k))log2(P(xt+1∣xt(k),yt+1(k))P(xt+1∣xt(k))),where xt+1 represents the state of neuron X at time t + 1, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\bf{x}}}}}_{t}^{(k)}=\{{x}_{t},{x}_{t-1},\ldots ,{x}_{t-k+1}\}$$\end{document}xt(k)={xt,xt−1,…,xt−k+1}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\bf{y}}}}}_{t}^{(k)}=\{{y}_{t},\,{y}_{t-1},\,\ldots ,\,{y}_{t-k+1}\}$$\end{document}yt(k)={yt,yt−1,…,yt−k+1} denote the past k states of neurons X and Y, respectively. In this study, the Markov order was set to k = 2. Binary spike trains were obtained from 15 min spontaneous activity recordings with a time step of 50 ms. A spike was detected by applying a Schmitt trigger algorithm (upper threshold of 1 Hz and lower threshold of 0.5 Hz) to the instantaneous firing rate r(t)35. Once the upper threshold was crossed, the neuron was considered active until r(t) fell below the lower threshold. To evaluate changes in effective connectivity within a network, we focused on functionally relevant connections, defined as those with a z-scored transfer entropy greater than 1 before the injury. The raw transfer entropy for each connection was then normalized to its preinjury value, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{TE}}}_{Y\to X}^{(0)}$$\end{document}TEY→X(0), and finally averaged across all connections to obtain the mean relative transfer entropy (RTE) of the network: 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{{\rm{RTE}}}=\frac{1}{N}{\sum}_{(X,Y)\in C}\frac{{{\rm{TE}}}_{Y\to X}}{{{\rm{TE}}}_{Y\to X}^{(0)}},$$\end{document}RTE¯=1N∑(X,Y)∈CTEY→XTEY→X(0),where C is the set of functionally relevant connections, and N is the number of elements in C. Statistical statement All values in the analyzed data are presented as the mean ± SD or as the means, unless otherwise noted. The sample size (n) is indicated in each section of the main text and in the figure captions. Statistical comparisons were performed using the two-sided Wilcoxon signed-rank test for paired (within-group) data and the two-sided Mann–Whitney U test for unpaired (between-group) data. For multiple group comparisons, p-values were adjusted using the Bonferroni correction. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Supplementary Information Reporting Summary