Graph Node Classification to Predict Autism Risk in Genes.

Area of Science:

• Genetics
• Computational Biology
• Machine Learning

Background:

• Autism spectrum disorder (ASD) has a significant genetic component, but identifying specific risk genes remains challenging.
• Understanding gene interactions and chromosomal locations is crucial for pinpointing genetic contributions to ASD.
• Graph neural networks (GNNs) offer a powerful framework for analyzing complex biological networks.

Purpose of the Study:

• To explore genetic risk associations with autism spectrum disorder (ASD) using graph neural networks (GNNs).
• To develop and evaluate GNN models for classifying the autism risk of genes.
• To assess the contribution of chromosome band location and protein interactions to ASD risk classification.

Main Methods:

• Constructed a gene network using the Sfari dataset and protein interaction network (PIN) data, with genes as nodes and interactions as edges.
• Employed GNN architectures including graph convolutional networks, Graph Sage, and graph transformer, alongside a Multi-Layer Perceptron baseline.
• Performed three classification tasks: binary risk association, multi-class risk association, and syndromic gene association.

Main Results:

• The Graph Sage model consistently outperformed other architectures across all classification tasks.
• Ablation studies confirmed the importance of chromosome band location and protein interactions in model performance.
• Achieved high accuracies: 85.80% for binary risk, 81.68% for multi-class risk, and 90.22% for syndromic classification.

Conclusions:

• GNNs, particularly Graph Sage, are effective tools for classifying ASD-related genes.
• Integrating gene interaction and chromosomal location data enhances the accuracy of ASD genetic risk prediction.
• This approach holds promise for advancing the understanding and identification of genetic factors in ASD.