MitoKG: A Pretrained Mitochondrial Heterogeneous Knowledge Graph

MitoKG is a mitochondrial-focused heterogeneous knowledge graph. It extracts only the mitochondria-relevant subset of five public biomedical databases (MitoCarta 3.0, STRING v12.0, Reactome, DrugBank, and CORUM 5.2). Starting from the MitoCarta 3.0 human core mitochondrial gene list, 984 unique gene symbols are retained; these map to 1,132 canonical UniProt entries, which form the gene/protein nodes of the graph. Only edges and auxiliary nodes connected to this core set are kept. Graph composition:- 4 node types: gene/protein (1,132 UniProt entries, corresponding to 984 MitoCarta core gene symbols), pathway (815), drug (305), protein complex (297). Total nodes: 2,549.- 5 forward edge types: protein-protein interaction from STRING at high confidence (14,763 edges), gene-pathway membership from Reactome (5,990 edges), drug-target from DrugBank (656 edges), gene-complex membership from CORUM (806 edges), and pathway hierarchy from Reactome (804 edges). Total forward edges: 23,019.- Coverage over the 984 core mitochondrial genes: 970 with at least one PPI, 860 with at least one pathway, 345 with at least one complex membership, and 292 with at least one drug targeting. A Heterogeneous Graph Transformer (HGT) was pretrained on this graph, producing 64-dimensional node embeddings that can be used directly as mitochondria-aware gene features in downstream models. Released artifacts: the HeteroData graph object, HGT model weights, 64-dim embeddings for the 984 unique gene symbols (gene_embeddings.npy), 64-dim embeddings for all 1,132 UniProt entries in the graph (all_gene_embeddings.npy), a UniProt ID whitelist of the 1,132 entries, and node-index mappings. Source databases retain their original licences; MitoKG does not redistribute raw tables from MitoCarta, STRING, Reactome, DrugBank, or CORUM.