In the public Hetionet dataset — 47,031 nodes and 2,250,197 edges spanning genes, compounds, diseases, and pathways — only 755 edges carry the label treats. That is the entire set of FDA-recognized compound–disease treatment relationships encoded in the graph. Compare that to 1,552 Compound nodes and 137 Disease nodes: a naive full cross-product is 212,624 possible compound-disease pairs, and the graph asserts a known therapeutic link for just 0.36% of them. That gap — the space between "known to treat" and "structurally plausible to treat" — is exactly the terrain computational drug-repurposing algorithms are built to search.
The mechanism: paths standing in for pharmacology
Hetionet, built by the Greene Lab (University of Pennsylvania) from 29 public resources, is not itself a repurposing engine — it's a substrate. Its power comes from the intermediate layer: 20,945 Gene nodes wired to compounds via 11,571 binds edges and to diseases via 265,672 regulates, 138,944 causes, and related edges. A compound-repurposing hypothesis becomes a graph-traversal question: does a path exist from Compound X, through a gene or pathway, to Disease Y, even when no direct treats edge has ever been curated? The original Hetionet paper (Himmelstein et al., eLife 2017) formalized this as "rephetio" — scoring hundreds of metapaths (typed path templates like Compound–binds–Gene–associates–Disease) with a logistic regression trained on the 755 known treats edges as positive labels.
A concrete, real path in this exact dataset: Sildenafil (DrugBank DB00203) carries 17 binds edges in the graph, including one to PDE5A, the phosphodiesterase-5 gene. That single gene-binding edge is the mechanistic thread connecting Sildenafil's original cardiovascular pharmacology to its second FDA approval — Revatio, for pulmonary arterial hypertension (approved 2005), a wholly different indication from Viagra's 1998 approval for erectile dysfunction. The graph does not "know" this repurposing story explicitly (Sildenafil has zero direct treats edges to pulmonary hypertension in this snapshot); it only encodes the binding relationship that made the discovery mechanistically plausible in the first place. That's the entire point of path-based repurposing methods: they recover signal the curated treats layer missed or hasn't caught up to.
Caveats worth stating plainly
Hetionet is a 2017 snapshot built primarily from DrugBank, GWAS Catalog, and related sources current to roughly 2015–2016; it does not reflect drugs approved or repurposed since, and its 755 treats edges are curated FDA-label indications, not off-label or investigational use. A path existing between a compound and a disease is a hypothesis generator, not evidence — the historical rephetio validation found real signal but also plenty of false positives, and any candidate still needs wet-lab and clinical confirmation. The index here is a static mirror; it carries none of the newer edges or nodes added to expanded Hetionet-style graphs since.
Why this matters: as repurposing pipelines increasingly lean on knowledge-graph traversal and embedding models trained on exactly this kind of structure, the ratio of mechanistic edges to confirmed outcomes is the real bottleneck — not compute, not embedding dimensionality, but the sparseness of ground truth to validate against. Graphs like this one are useful precisely because they make that sparsity visible and queryable.