Biology Atlas

Map / Concepts

concept · intermediate · partly written

Biomedical knowledge graphs: shapes, edges and what has worked

A biomedical knowledge graph joins genes, drugs and diseases by edges that are claims; what an edge is, and how its claim is checked, differs a great deal between graphs.

Also called: KG, knowledge graph, biomedical KG

Short answer

A biomedical knowledge graph is a network of things (genes, drugs, diseases) joined by edges that are claims. The deep difference is what an edge is: a bare relation copied from a database, a relation with its source, or a statement with its evidence and a belief score. Headline benchmark scores flatter; honest splits sit lower. No documented clinical decision was found that rested on a public graph. 123

Reported, not yet replicated

What an edge is, graph by graph more

Hetionet, PrimeKG and SPOKE store typed relations whose origin is a curated database; PrimeKG describes no per-edge evidence. RTX-KG2 keeps the source database and the supporting publications on each edge. INDRA separates a mechanistic Statement from the list of Evidence objects (sentence, article, source) behind it and attaches a belief score. Wikidata attaches references and qualifiers to statements and lets conflicting claims coexist. The counts are the builders' own descriptions of what they built. 124536

Reported, not yet replicated

Philosophies and where they disagree more

Integrate everything or curate narrowly: SPOKE draws on 41 databases and the Translator describes over 250 knowledge sources, while the Rephetio model ended with 11 metapaths chosen from hundreds and INDRA leaves gene-disease links out by design. Assertions or evidence: Hetionet and PrimeKG treat an edge as a relation; INDRA keeps evidence apart from the claim; Wikidata calls itself a database of assertions, not absolute truth. One graph or assembled on demand: PrimeKG is built once and frozen; BioThings Explorer builds its graph at query time and INDRA assembles with chosen policies. This atlas found no builder arguing the opposite camp in their own words, so this table is a framing, not a debate. 472368

Our inference

What has worked, measured more

Rephetio (Hetionet) reported AUROC 97.4% on its 755 known treatments (training positives), 85.5% on DrugCentral indications not used in training and 70.0% on clinical-trial indications, and a degree-only prior scored 97.9% on training but 54.1% on the DrugCentral set (the authors' own figures). A time-resolved test of a text-mined graph fell from random-split 0.95 AUROC to a time-split peak of 0.797. A degree-only baseline reached AUROC of at least 0.95 on 17 of 20 Hetionet-derived networks. Only two honest-split numbers were opened (about 0.70 to 0.85); that is two points, not a pattern. 2910

Reported, not yet replicated

Clinical impact: what was found more

One documented case from a proprietary graph: BenevolentAI's graph flagged baricitinib, an approved rheumatoid-arthritis drug, for 2019-nCoV; all but one author of the letter were company employees and the letter says it is not medical advice. Whether the graph was needed is untested, since AAK1 binding and JAK pharmacology were already known. The later trial figures (patient count, recovery days) could not be opened and are unverified. For public graphs no prospective test of a top prediction was found: the Rephetio authors close by saying additional studies will be needed, and the Translator consortium writes that it did not provide evidence of clinical benefit. 1127

Reported, not yet replicated

Entity linking: where it fails more

Measured failures exist mainly for the text-reading tools. Gilda scored F1 0.618 overall and 0.694 on human genes on the BioCreative Bio-ID benchmark; 3,662 synonyms are shared by at least two human genes. INDRA dropped 38% of its text Statements for ungrounded entities, and one reader's error "MAOA binds MAOB" repeated across 10 publications, so a mention count is not correctness. RTX-KG2 merges nodes and reports no merge accuracy; its own check misses merges between compatible categories. PrimeKG groups disease terms by string match, language-model similarity and manual approval, with no accuracy figure in the text read. 12351

Reported, not yet replicated

Proprietary graphs more

Two company-authored papers were read (BenevolentAI, Healx), and both are the owner's claims. QIAGEN's knowledge base was followed up separately: it keeps experimental context per finding but its quality is asserted, not publicly measured. No vendor documentation, patent or filing for the other companies was opened, so any size or accuracy figure a company gives is unverified here. 111314

Reported, not yet replicated

What an edge could mean: interventional causal edges deep

Every edge in the graphs above is an association or a literature assertion. A claim about what happens when a gene is manipulated is a different kind of statement and exists only where perturbation data exist. One theorem-level paper shows a latent causal graph can be identified up to an equivalence class from unpaired observational and interventional data, given at least one intervention per latent node; on a Perturb-seq dataset (105 genes, 217 interventions) it predicted held-out interventions with R-squared 0.99 for single and 0.98 for unseen double interventions. That is prediction, not proof that the edges are causal. The paper does not address knowledge-graph edges; the bridge is our own inference. 15

Our inference

Evolutionary information as an edge type deep

Orthology as an edge type was not found in Hetionet's 24 metaedges, PrimeKG's 30 edge types or SPOKE's examples; the full edge lists of RTX-KG2 and SPOKE were not read. Hetionet does hold one evolutionary edge: Gene-covaries-Gene, 61,690 edges built on evolutionary rate covariation, which the authors call "weak yet appreciable" in prediction. Nobody found tested orthology edges as such. 2161

Reported, not yet replicated

Plain answer

Think of a city map that also records who vouched for each road. Some graphs draw every road (PrimeKG), some keep the witnesses' notes (INDRA). Edges are claims; whether a claim is a bare relation, a relation with a source, or a statement with evidence and belief decides what the graph can honestly say. 13

Our inference

Why it matters

Any model trained on a graph is only as good as what an edge means and how its identifiers were merged. The measured gaps between headline and honest-split scores are the practical reason to ask what each edge asserts. 210

Our inference

Real numbers

PrimeKG size

129,375 nodes of 10 types and 4,050,249 relations of 30 edge types from 20 resources (the authors' counts). 1

Reported, not yet replicated

Hetionet size

47,031 nodes of 11 types and 2,250,197 edges of 24 types. 162

Reported, not yet replicated

SPOKE size

27,056,367 nodes of 21 types and 53,264,489 edges of 55 types from 41 databases, with about 92% of nodes being proteins. 4

Reported, not yet replicated

RTX-KG2 size

6.4 million nodes and 39.3 million edges in the canonical KG2.7.3 build, with 77 Biolink predicates and about a third of edges from SemMedDB text mining. 5

Reported, not yet replicated

INDRA 2023 corpus

About 900,000 unique Statements collapsed from about 2.9 million filtered ones, 63% with a single mention; about 73% of 1,800 curated Statements correct. 3

Reported, not yet replicated

Wikidata cross-reference check

Experts judged 2,007 of 2,030 user-added GARD and MeSH cross-references correct (98.9%). 6

Reported, not yet replicated

Still open here

Not yet written: Equation.

Linked from

Sources

  1. Chandak P, Huang K, Zitnik M. Building a knowledge graph to enable precision medicine. Scientific Data 10:67 (2023). paper; opened 2026-11-18 all records citing this
  2. Himmelstein DS, Lizee A, Hessler C, et al.. Systematic integration of biomedical knowledge prioritizes drugs for repurposing. eLife 6:e26726 (2017). paper; opened 2026-11-18 all records citing this
  3. INDRA paper on automated assembly of mechanistic knowledge from text and databases (PMC10167483). PubMed Central (2023). paper; opened 2026-11-17 all records citing this
  4. Baranzini SE, et al.. A biomedical open knowledge network harnesses the power of AI to understand deep human biology. AI Magazine (Europe PMC PMC9940622) (2022). paper; opened 2026-11-18 all records citing this
  5. Wood EC, et al.. RTX-KG2: a system for building a semantically standardized knowledge graph for translational biomedicine. BMC Bioinformatics (2022). paper; opened 2026-11-18 all records citing this
  6. Waagmeester A, Stupp G, et al., Su AI. Wikidata as a knowledge graph for the life sciences. eLife 9:e52614 (2020). paper; opened 2026-11-18 all records citing this
  7. Biomedical Data Translator Consortium. Progress toward a universal biomedical data translator. Clinical and Translational Science (2022). paper; opened 2026-11-18 all records citing this
  8. Callaghan J, et al., Su AI. BioThings Explorer: a query engine for a federated knowledge graph of biomedical APIs. arXiv 2304.09344 (2023). preprint, not peer reviewed; opened 2026-11-18 all records citing this
  9. Mayers M, Li TS, Queralt-Rosinach N, Su AI. Time-resolved evaluation of compound repositioning predictions on a text-mined knowledge network. BMC Bioinformatics (2019). paper; opened 2026-11-18 all records citing this
  10. Zietz M, Himmelstein DS, Greene CS, et al.. The probability of edge existence due to node degree: a baseline for network-based predictions. GigaScience (2024). paper; opened 2026-11-18 all records citing this
  11. Richardson P, et al.. Baricitinib as potential treatment for 2019-nCoV acute respiratory disease. The Lancet (2020). paper; opened 2026-11-18 all records citing this
  12. Gyori BM, Hoyt CT, Steppi A. Gilda: biomedical entity text normalization with machine-learned disambiguation as a service. Bioinformatics Advances (2022). paper; opened 2026-11-18 all records citing this
  13. QIAGEN. QIAGEN Biomedical Knowledge Base product page. qiagen.com (2026). vendor-page, not peer reviewed; opened 2026-11-18 all records citing this
  14. Krämer A, et al.. Causal analysis approaches in Ingenuity Pathway Analysis. Bioinformatics (2013). paper; opened 2026-11-18 all records citing this
  15. Zhang J, Squires C, Greenewald K, Srivastava A, Shanmugam K, Uhler C. Identifiability guarantees for causal disentanglement from soft interventions. NeurIPS 2023 (arXiv 2307.06250) (2023). preprint; opened 2026-11-18 all records citing this
  16. Himmelstein DS, et al.. Hetionet repository and metagraph. GitHub (hetio/hetionet) (2017). database; opened 2026-11-18 all records citing this
Where it sits on the map
History of this record (8)

This record has been changed 8 times, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.

Status
partly written
Last updated
2026-11-19
Record
content/concepts/biomedical-knowledge-graphs.json