This paper shows a way to turn electrical recordings from small neural cultures into a network map, and then to analyze that map with tools from graph theory. The main idea is to use a machine-learning model called Reservoir Computing (RC) to infer an Intrinsic Connectivity Map (ICM). The authors treat that ICM as an effective network map and ask how well graph measures of the map relate to observed activity such as firing rates.

The researchers start with multichannel electrophysiological recordings made on microelectrode arrays (MEA). Each electrode picks up signals from a small region of tissue, so an electrode is treated as a node that represents a local circuit rather than a single neuron. After filtering and spike detection, signals are converted into multichannel instantaneous spike rate (ISR) sequences and short network burst episodes are extracted. Those processed sequences are used to train an artificial neural network with a reservoir computing architecture.

Reservoir computing is a form of recurrent neural network that keeps a fixed, high-dimensional internal ‘reservoir’ and trains only a readout layer. In this work the reservoir provides nonlinear, leaky memory of past activity. The final readout is trained by linear regression with LASSO regularization (a method that penalizes overly complex solutions). By linearizing the trained model and removing short-term effects, the authors extract the ICM, which they interpret as an effective adjacency matrix of connections between electrodes.

The paper then applies graph-theory measures to the RC-derived ICM. The authors compute node-level centrality measures and other local and global metrics and compare them with experimentally measured observables such as firing rates and network descriptors. To validate the overall approach, they also run controlled simulations of networks (using the NEST simulator) where the true adjacency matrix is known. This allows them to benchmark how similar the inferred ICM is to the ground-truth and to study model performance across different graph structures.