Year: 2026
Type: preprint
Abstract: Background Autism spectrum disorder (ASD) involves diffuse, cell-type-specific transcriptomic dysregulation in the cerebral cortex. Single-nucleus RNA sequencing (snRNA-seq) has begun to resolve it, but analysis is dominated by per-gene differential expression, which tests one gene at a time and ignores the manifold structure that separates patient from control cells. Graph neural networks offer an alternative, but most approaches place genes at graph nodes and score them by local co-expression rather than by deviation from healthy cell states. Results We present scRED, a graph-convolutional autoencoder built on a cell–cell similarity graph rather than a gene–gene graph. The model is trained on control cortical cells; the per-gene deviation in reconstruction error between patient and control cells provides the ranking statistic. Applied to the Velmeshev ASD cortex cohort across 17 cell types, scRED recovered SFARI-curated risk genes in neuronal populations at 1.98-fold enrichment against a matched, donor-disjoint control-versus-control negative control of 1.44-fold, exceeding that control in all 9 neuronal contexts (Wilcoxon signed-rank p = 0.0039); glial and endothelial populations showed no separation. Across all 17 contexts the disease contrast exceeded its negative control (p = 0.0061). The cell-graph and gene-graph formulations performed comparably (mean difference − 0.2%, p = 0.82); both exceeded a same-shape multilayer perceptron, though the primary-depth difference did not survive Holm correction. A curated gene-symbol filter applied before highly variable gene selection improved enrichment consistently (scRED vs no-filter + 1.9%, p = 0.035). Conclusions scRED extracts cell-type-conditional disease signal concentrated in cortical neurons and not detected in the examined glial populations. The negative control shows that roughly 45% of the raw enrichment is reproduced by the matched control — reflecting scoring asymmetry together with between-donor and technical variation — leaving a residual enrichment consistent with disease, a decomposition ordinary overlap statistics do not provide. Graph-based formulations outperformed a non-graph baseline, though not after Holm correction at the primary depth, while node choice did not matter and a simple Wilcoxon rank-sum test remained competitive; we report these results without minimisation.
Source: Research Square
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7234793901
