Multi-electrode arrays (MEAs) provide a noninvasive interface with sub-millisecond temporal resolution and long-term, multi-site recordings, enabling mechanistic investigations of in vitro human brain development and disease-related dysfunction; nevertheless, conventional MEA pipelines largely rely on firing/burst statistics or channel-/waveform-level features, which can be insufficient to systematically characterize and interpret network-level organization and its subtle pathological deviations. Accordingly, representing multichannel spiking activity as time-varying graphs and leveraging graph neural architectures to extract quantitative topological descriptors offers a principled route to capture instantaneous connectivity changes and early electrophysiological phenotypes at the network scale. Within the context of autism spectrum disorder (ASD) risk, valproic acid (VPA) serves as a prototypical environmental perturbation, with prior evidence linking VPA exposure to altered expression of neurodevelopmental and synaptic-transmission pathways and overlap with ASD risk genes, alongside functional disruptions in forebrain organoids relevant to synaptic/network efficiency. “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 the author Arianna Mencattini, a researcher at University of Rome Tor Vergata.
This study used valproic acid (VPA)–exposed human forebrain organoids (hFOs) as an in vitro neurodevelopmental perturbation model to probe ASD-linked synaptic dysfunction, leveraging a 24-well, 16-electrode-per-well Axion Biosystems MEA platform for parallel recordings: seven male-donor organoids were cultured one per well, with VPA applied at 1 mM to three wells after 7 days post-plating while four wells remained untreated controls; MEA sessions were acquired at 0 h (pre-treatment), 30 min, 24 h, and 48 h post-exposure at 12.5 kHz for 5 min per session, using the first 2.5 min for GDN training and the remaining 2.5 min for testing. To ensure adequate electrode contact, non–Matrigel-embedded hFOs were generated, 42-day-old organoids were seeded into MEA wells and cultured for 7 days before VPA treatment, and the analysis focused on wells A1–A3 (U1M+VPA) and B1–B4 (U1M controls). For signal conditioning and graph construction, raw traces were band-pass filtered (200–3000 Hz, zero-phase Butterworth), spikes were detected via a 50-ms window noise-estimation scheme with threshold th_spike=4.5×ref-value, and each detected spike triggered a Kaiser-window amplitude modulation (shape factor 10, ~20 ms) applied synchronously across electrodes; a centered 18-ms segment was retained as an AM-spike and AM-spikes were concatenated chronologically across channels to form a multivariate AM-spike sequence. This sequence was then segmented with an overlapping 4.5-ms window (T=56 samples) advancing by one sample; each window was fed to the GDN to forecast the next sample and to extract a corresponding directed graph, yielding 169 graph instances per 225-point AM-spike (with the first 56 points excluded due to prediction initialization). During model development, the first half of the time series was used for training and the second half for validation, and a distinct GDN was trained for each well and acquisition session to derive graph structures and features. From each graph, 19 topological descriptors were computed to build feature matrices, descriptors were normalized to the 0 h session mean to mitigate inter-well heterogeneity, and a two-step DS-driven selection retained globally discriminative time points (>95th percentile) followed by features with DS>0.75; the resulting feature set was classified via LDA with leave-one-well-out (LOWO) cross-validation per session to quantify time-resolved connectivity deviations induced by VPA.
The results indicate that GDN forecasting exhibits no appreciable temporal shift across sessions, yet the VPA condition shows an overall decrease in prediction performance relative to CTRL, with no notable session-dependent trend, suggesting reduced predictability of network activity under VPA exposure. Discriminative score (DS) analysis further shows that certain topological descriptors (e.g., path length) remain highly informative across 30 min, 24 h, and 48 h, whereas others (e.g., diameter) contribute in a session-specific manner, supporting the central role of network-level structure in class separation. PCA visualization corroborates this interpretation by revealing higher dispersion (greater inter-organoid heterogeneity) in CTRL and reduced variability in VPA, consistent with a common treatment-induced effect. Under the LOWO protocol, a classifier trained at 30 min generalizes poorly to later sessions, with weakening separability and a bias toward VPA at 48 h, motivating session-specific models; the strongest discrimination occurs at 24 h, and well-level majority voting achieves the best accuracy (up to 100%). By contrast, waveform-based classification remains below 50% accuracy, while conventional MEA-NAP metrics (MFR/BR/SIB) still show clear separation and MFR stabilizes after ~1 minute, jointly reinforcing the conclusion that VPA induces detectable network-level functional alterations.
In conclusion, this study introduced and validates a MEA–organoid analysis framework based on a Graph Deviation Network which encodes amplitude-modulated spike trains as dynamic graphs and derives topological descriptors to detect deviations in network organization; in a VPA-based proof of concept, the approach consistently identifies early network dysfunction within 24 h and yields signatures consistent with reduced efficiency, increased path length, and decreased input connectivity. By demonstrating that abnormalities can be detected from only millisecond-scale MEA snippets with high temporal precision, the framework supports prospective use of MEA-coupled hFOs for rapid ASD-risk–related functional phenotyping and real-time neurotoxicity screening, and it lays groundwork for future “read–write” experiments involving active stimulation and dynamic response tracking. However, this study still has several limitations, including restricted biological coverage (e.g., downstream analyses focusing on the U1M line selected for higher drug sensitivity) and a preprocessing design that, while emphasizing collective activity, may partially obscure small inter-electrode timing offsets. “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 Arianna 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.
The paper, “MEA-Based Graph Deviation Network for Early Autism Syndrome Signatures in Human Forebrain Organoids” was published in the journal Cyborg and Bionic Systems on Nov. 6, 2025, at DOI: 10.34133/cbsystems.0441.
END