AI-Powered Pathogenicity Prediction for Mitochondrial DNA Variants
Mutations in mitochondrial DNA cause rare neurodegenerative and metabolic diseases, but roughly 40% of known variants remain classified as “Variants of Uncertain Significance” (VUS), meaning clinicians cannot tell patients whether their mutation is harmful. MitoGraph uses a Graph Attention Network trained on a knowledge graph of genes, diseases, and conservation data to predict which VUS are likely pathogenic. Explore the knowledge graph, inspect per-edge attention weights learned by the model, browse ranked predictions, and examine how variants cluster in the neural network's latent space.
Mitochondrial Complex Graph
Force-directed layout: Complexes → Genes → Variants → Phenotypes • Edge thickness = GATv2Conv attention weight (α) • Click any node to inspect features • Drag nodes to explore
Variant–Phenotype Predictions
Top-ranked phenotype association per VUS • Hard negative mining + attention scoring • △ = probability > 85%
Latent Space Explorer
UMAP projection of GATv2Conv variant embeddings • ★ = VUS in pathogenic cluster • Click a point to filter the table above