Modular architecture and robustness to damage in spiking neural networks modeling in vitro neurons

• Modular architecture in spiking neural networks modeling in vitro neurons confers robustness to damage and facilitates recovery (Frontiers).¹
• The findings inform resilient neuroprosthetic and BCI design.¹
• Modularity and recovery are relevant to long-term implant design.^{1 1}

Gardner updates

• Modular architecture in spiking neural networks modeling in vitro neurons confers robustness to damage and facilitates recovery, with relevance to resilient neuroprosthetic and BCI design (Frontiers, tier-2). ¹

Weekly enrichment (2026-07-20)

• The underlying study (Frontiers in Neuroscience, Volume 19, 2025; DOI 10.3389/fnins.2025.1570783, also posted on bioRxiv) paired in vitro experiments with in silico spiking-network modeling to ask how modular topology and synaptic plasticity jointly shape a network’s response to and recovery from focal damage.^{2 3}
• Modular cultures were built by growing primary rat cortical neurons on PDMS topographical substrates patterned as ~300 μm-wide parallel “tracks,” so neurons connected strongly along tracks (forming modules) and weakly across them; a ~4 mm scalpel incision on day in vitro 12–13 delivered the focal lesion.²
• Wide-field calcium imaging (50 frames/s over a 7.1×7.1 mm field, ~1,400 regions of interest) tracked spontaneous network bursts before injury and at 0 min, 15 min, 2 h, 6 h, and 24 h post-injury; activity dropped sharply immediately after the cut but recovered to pre-damage levels within about 24 h.²
• The computational model used ~2,800 Izhikevich neurons (80% excitatory, 20% inhibitory) with axon-growth-based connectivity, and reproduced the experimentally observed post-lesion decline and recovery when spike-timing-dependent plasticity (STDP) was included.²
• Simulated damage across six conditions showed intra-modular cuts (perpendicular to tracks) were more disruptive than inter-modular cuts; immediately after injury, burst frequency fell to about 0.27 of control for an intra-modular half-cut and 0.06 for a full intra-modular cut, versus 0.60 (half) and 0.41 (full) for inter-modular cuts.²
• Recovery depended on lesion size, direction, and the presence of modular structure: small along-track damage recovered by ~2 h, but full intra-modular (across-track) damage stayed depressed (about 0.70 of control) even at 24 h, and non-modular control networks failed to restore activity regardless of cut size.²
• Using a reservoir-computing readout, the network classified spoken digits (“zero,” “one,” “two”) with pre-damage accuracy of about 66.2% for the modular network and 69.8% for the unpatterned control, both well above the 33.3% chance level.²
• After retraining the linear output weights, classification performance was maintained under three of four modular damage conditions (intra-modular half, inter-modular half, inter-modular full) but dropped for full intra-modular and non-modular damage (e.g., ~59.8% for full intra-modular), indicating modularity protects sensory information representation against small lesions.²
• The authors note STDP was not directly demonstrated in the in vitro cultures (a stated limitation), positioning the combined experimental-numerical platform as a way to predict recovery in damaged networks and to inform resilient neuroprosthetic and long-term BCI implant design.²

Footnotes

1. https://news.google.com/rss/articles/CBMilAFBVV95cUxOSzBrM3JDQkd1QXRmejF6V0pHM01HMjZMX05aS213a2ZjTlEzYkVESjMzZFFIUVhNaUdqSzdPa2pZNEhuNi03VTFta0o1d0NBczduVExKNERVSktvLUVEeEpoT1RDRHNQeFVPa0IzVXZPZ2p1ay03bWM1VUF2MWRnaDFRcW1mQUc5U2VFOWlzTkhCUEFa?oc=5 ↩ ↩² ↩³ ↩⁴ ↩⁵

2. https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2025.1570783/full ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

3. https://doi.org/10.1101/2025.02.04.636401 ↩