MEA-Based Graph Deviation Network for Early Autism Syndrome Signatures in Human Forebrain Organoids

Multi-electrode arrays (MEAs) are a key enabling technology in the development of cybernetic systems, as they provide a means to understand and control the activity of neural populations linking brain microtissue dynamics with electronic systems. MEAs allow high-resolution, noninvasive recordings of neuronal activity, creating a powerful interface for investigating in vitro brain development and dysfunction. In this work, we introduce a novel deep learning framework based on a graph deviation network (GDN) to analyze spiking activity from human forebrain organoids (hFOs) and predict network-level alterations associated with autism spectrum disorder (ASD) risk. Our method extends traditional spike and burst analysis by encoding amplitude-modulated spike trains as dynamic graphs, enabling the extraction of meaningful topological descriptors. These graph-based features are then processed to detect deviations in network organization induced by neurodevelopmental perturbations. As proof of concept, we examine the impact of valproic acid (VPA), a known environmental ASD risk factor. VPA disrupts synaptic signaling in hFOs, reducing efficiency, increasing path length, and decreasing input connectivity. Despite biological variability, the GDN consistently detects early dysfunction within 24 h post-exposure and captures transient millisecond-level events. This supports MEA-coupled hFOs as predictive platforms for ASD risk assessment and real-time neurotoxicity screening.