Fig. 4: Comparison with general LP baselines across tasks and runtime analysis.

a, BioPathNet benchmarked against R-GCN, HGT, RAGAT and NBFNet across multiple LP tasks, including gene function prediction, drug repurposing (indications for adrenal gland (AD), anaemia (AN), cell proliferation (CP), cardiovascular (CV) and mental health (MH)), synthetic lethality gene pair prediction (unfiltered, thr = 0; filtered, thr = 0.3) and lncRNA–target prediction. Performance metrics include MRR and Hits@1, 3, 10 for each task. Individual data points represent different seeds (N = 5) and are summarized in boxplots, with the bottom and top edges indicating the 25% and 75% percentiles, the line in the box representing the median, and whiskers extending to 1.5× the interquartile range on both sides. b, MRR vs training time (hours) for BioPathNet compared to general LP methods across all four tasks (left) and task-specific baselines (right). The circles highlight BioPathNet, and the asterisk indicates that, for the lncRNA interaction task, the KG made from the node intersection between DeepLGP and BioPathNet was used to ensure a fair comparison.