Explainable multi-view graph neural networks for prioritizing autism risk genes and identifying transcriptional modules
Year: 2026
Type: preprint
Abstract: Autism spectrum disorder (ASD) has a heterogeneous genetic architecture in which risk-associated genes may participate in interconnected molecular pathways. We present an graph-explainable multi-view graph neural network (GEM-GNN) for prioritizing ASD-associated genes by integrating protein-protein interaction (PPI) and brain-specific co-expression networks with genomic features. The model uses two GraphSAGE branches to learn representations from the structural and functional network views, combines these representations through a feature bottleneck, and applies an adjusted focal-loss objective to address class imbalance. Permutation feature importance and neighborhood subgraph extraction are used to examine the genomic and network evidence underlying candidate rankings. In experiments covering 19,832 protein-coding genes, the manuscript reports an area under the precision-recall curve of 0.8008, an area under the receiver operating characteristic curve of 0.93, an F1 score of 0.74 and a precision at 100 of 76%. An ablation analysis reports an AUPRC of 0.80 for the bottleneck-fusion model, compared with 0.71 for the PPI-only model, 0.69 for the co-expression-only model and 0.75 for simple concatenation. The authors also report that 85 high-ranking candidates overlap with independent de novo mutation evidence and that candidate genes are enriched for RNA polymerase II-mediated transcriptional processes. These results suggest that combining structural and brain-specific functional networks may support ASD gene prioritization and pathway-level interpretation. The reported performance and biological findings require interpretation in light of the dataset construction, evaluation protocol and validation procedures described in the manuscript.
Source: Research Square
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225878218
