Multi-electrode arrays (MEAs) are widely used to study human brain development and disease by enabling detailed, noninvasive recordings of neural activity. However, traditional MEA analysis methods often focus on simple firing or burst statistics, which may not capture the complexity of network-level changes associated with neurological disorders.
A recent study led by Arianna Mencattini from the University of Rome Tor Vergata introduced a new approach that models multichannel spiking activity as time-varying graphs. The research team developed a deep learning framework based on a Graph Deviation Network (GDN), which encodes amplitude-modulated spike trains into dynamic graphs. This method aims to detect deviations in network organization and predict alterations linked to autism spectrum disorder (ASD) risk, specifically those induced by valproic acid (VPA).
"Therefore, we introduces a deep learning framework based on a Graph Deviation Network (GDN) that encodes amplitude-modulated spike trains into dynamic graphs to model deviations in network organization and to predict ASD-risk–associated, VPA-induced network-level alterations from MEA-coupled human forebrain organoids," said Arianna Mencattini.
The study used human forebrain organoids exposed to VPA as an in vitro model for ASD-related synaptic dysfunction. Researchers recorded neural activity using an Axion Biosystems MEA platform with 16 electrodes per well across multiple sessions before and after VPA exposure. Data processing involved filtering signals, detecting spikes, and constructing dynamic graphs from the resulting spike sequences. The GDN was trained separately for each well and session, allowing extraction of topological features that were then analyzed for differences between treated and control samples.
Results showed that while GDN forecasting did not shift over time within sessions, prediction performance decreased under VPA exposure compared to controls. Certain graph-based descriptors remained informative across different time points post-exposure, highlighting their role in distinguishing between conditions. Principal component analysis indicated reduced variability among VPA-treated samples compared to controls.
Classification accuracy was highest at 24 hours post-treatment when using graph-based features; traditional waveform-based methods performed poorly by comparison. Conventional MEA metrics also confirmed clear differences between groups following VPA treatment.
In summary, the study demonstrates that modeling neural activity as dynamic graphs can reveal early connectivity changes linked to ASD risk factors like VPA exposure—sometimes within just 24 hours—using only brief segments of data with high temporal precision. "Future studies should strengthen generalizability via independent datasets (across iPSC lines and laboratories) and extend the paradigm toward broader donors/perturbations and stimulation-enabled protocols to improve robustness and mechanistic interpretability," said Mencattini.
Authors of the paper include Arianna Mencattini, Giorgia Curci, Alessia Riccardi, Paola Casti, Michele D'Orazio, Joanna Filippi, Gianni Antonelli, Erica Debbi, Elena Daprati, Wendiao Zhang, Qingtuan Meng, and Eugenio Martinelli.
