CNS*2026 Halifax

Introduction
Electrophysiological studies in neuroscience probe interactions among neuronal populations across multiple scales, from single-cell activity to large network dynamics [1,2]. Here, we present a computational framework to decode signals from microelectrode array (MEA) recordings [3]. The model is based on Reservoir Computing (RC) and learns spike-rate sequences to reproduce network responses to external stimuli. A key outcome is a macroscopic connectivity map capturing effective connectivity with higher accuracy than standard statistical methods such as cross-correlation and transfer entropy. We describe the model, discuss its implications and limitations, and present applications to cultured neuronal networks under different interventions.


Methods
The approach relies on electrophysiological recordings from mouse cortical cultures acquired via microelectrode arrays (MEA). After preprocessing (filtering and spike detection), signals are converted into multichannel instantaneous spike-rate (ISR) sequences, from which bursting episodes are extracted. These are used to train an artificial neural network with a reservoir computing (RC) architecture to learn the synaptic transmission function underlying rate-coded activity. The network is represented macroscopically, with nodes corresponding to MEA electrodes. The RC reservoir performs nonlinear transformations with leaky memory, and outputs are obtained via LASSO-regularized linear regression [4].


Results
Model validation followed two complementary approaches. First, the inferred connectivity map was benchmarked against a ground-truth network generated in silico, with simulations designed to replicate MEA measurements. Second, both in silico and in vitro (real neuronal cultures) data were used in a predictive framework: the model was trained and validated on spontaneous activity, while testing was performed using responses to controlled local stimuli, including optogenetic perturbations. Model predictions under identical stimuli were then compared with the recorded responses.
In the presentation, we will report the model’s performance and highlight selected applications along with their results.

Discussion
In this study, we developed a computational model that decodes spatio-temporal data from electrophysiological measurements of neuronal cultures. The model reconstructs the network structure on a macroscopic domain and predicts the response to a localized stimulus. Our primary goal was to create an advanced experimental data analysis tool for processing complex time-series. The results obtained indicate that the model not only serves as a data analyzer but can also function as a network simulator.


References
[1] Llinás, R. R. (1988). The intrinsic electrophysiological properties of mammalian neurons: insights into central nervous system function. Science, 242(4886), 1654-1664.
[2] Contreras, D. (2004). Electrophysiological classes of neocortical neurons. Neural Networks, 17(5-6), 633-646.
[3] Auslender, I., Letti, G., Heydari, Y., Zaccaria, C., & Pavesi, L. (2025). Decoding neuronal networks: A Reservoir Computing approach for predicting connectivity and functionality. Neural Networks, 184, 107058.
[4] Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology, 58(1), 267-288.



Acknowledgement
This work was financed by the European Union - NextGenerationEU - National Recovery and Resilience Plan (NRRP) - Mission 4 Component 2 Investment 1.2 - "Funding projects presented by young researchers" MSCA PNRR Young Researchers, "CIRCUS project" - MSCA20240000106 - CUP E63C25000820007.