Autism Risk Classification Using Graph Neural Networks Applied to Gene Interaction Data

We use a gene interaction network to predict which genes are associated with Autism Spectrum Disorder (ASD), thus allowing for earlier detection of ASD. ASD is a disorder that affects the development of the brain. No medical tests can diagnose ASD, so the diagnosis is based on the individual's behavior and development. We used gene location in the chromosome band and their possible interaction with each other to help classify the gene association with ASD. Graph sage and graph convolutional network (GCN) models were studied to organize these associations. The model's performance was analyzed using various parameters to obtain optimal performance. The graph sage model obtained an area under the ROC curve (AUC) of 0.85 percent on testing and 0.85 on validation. The GCN model achieved 0.78 AUC on testing and 0.78 AUC on validation. GCN achieves a specificity of 0.96 and a sensitivity of 0.63. Graph Sage achieved a specificity between 0.96 and a sensitivity of 0.94. Our classification performance demonstrates the capability of graph sage to classify genes carrying association risk for ASD. This analysis also demonstrates the inability of GCN to recognize association, as determined by the low sensitivity. Our results show that it is possible to apply graph neural networks to understand the link between diseases and genetics.