Fig. 2: Overview of the four LP tasks across different KGs and impact of BioPathNet’s design choices on performance.

a, Schematic of the gene function prediction KG, linking genes, chemicals and cellular pathways. Training triplets (G1: gene–pathway links extracted from KEGG) are in black; BRG edges (extracted from Pathway Commons for message passing) are in red. b, Simplified schematic of the PrimeKG graph for disease–drug indication prediction, connecting, for example, genes, diseases, drugs and phenotypes. G1 edges (disease–drug links) are in black; BRG edges (all other relations) are in red. c, Schematic of synthetic lethality KG for gene pair prediction. G1 edges (SL links) are in black; BRG edges are in red and include additional SL interactions alongside associated nodes (cellular components, molecular functions, biological processes and pathways). d, LncRNA-mediated regulation KG from LncTarD 2.0, with BRG from Pathway Commons. The graph features six node types (lncRNAs, microRNAs, mRNAs, pseudogenes, transcription factors and proteins), with regulatory relationships in black and BRG protein interactions in red. e,f, Ablation study results on the NTA and SANS (e) and on the BRG (f) showing performance changes vs original BioPathNet, that is, increases in green and decreases expressed as percentages in red across five seeds for each experimental task.