Knowledge Graphs in Biomedicine: Connecting Genes, Diseases, Pathways, and Drugs

Biomedical data is, by its nature, fragmented. Gene function is documented in one database, protein interactions in another, disease associations in a third, drug-target relationships in a fourth, and the clinical evidence linking all of these to patient outcomes sits in yet another set of repositories, each with its own schema, nomenclature, and update cycle. The scientific knowledge needed to understand any single disease mechanism or evaluate any single drug target is spread across sources that were not designed to talk to each other.

Knowledge graphs are a structural response to this fragmentation. By representing biological entities as nodes and the relationships between them as edges, they make heterogeneous biomedical knowledge queryable, traversable, and amenable to computational reasoning in ways that isolated databases and unstructured literature cannot support. Over the past decade, biomedical knowledge graphs have moved from a data management concept to an active infrastructure for drug repurposing, target identification, disease mechanism research, and precision medicine, with a growing body of validated applications demonstrating where they add genuine value and where their limitations remain unresolved.

What a Biomedical Knowledge Graph Is

Mathematically, a knowledge graph is a labeled directed graph in which nodes represent entities, directed edges represent relations between them, and each edge carries a type that specifies the nature of the relationship. In a biomedical context, the entities are things like genes, proteins, diseases, drugs, pathways, phenotypes, and biological processes. The edges encode relationships such as "gene encodes protein," "drug inhibits target," "variant associates with disease," "pathway includes gene," and "drug causes adverse event."

The graph structure is what distinguishes a knowledge graph from a flat database table or an unstructured literature repository. A flat database can answer the question "which drugs target BRCA2?" A knowledge graph can answer the question "which existing drugs target proteins that interact with BRCA2-associated pathway members, where that pathway is implicated in the disease of interest, and what evidence supports each step of that chain?" That kind of multi-hop reasoning across heterogeneous data types is the core capability that knowledge graphs add to biomedical research.

Biomedical knowledge graphs integrate entities such as genes, proteins, diseases, drugs, and phenotypes through structured relationships drawn from multi-source data, followed by integration and quality control. The underlying data sources span curated biological databases (UniProt, Reactome, ChEMBL, DrugBank), disease-gene association resources (DisGeNET, OMIM, ClinVar), ontological vocabularies (Gene Ontology, Human Phenotype Ontology, MONDO Disease Ontology), and, increasingly, information extracted from the biomedical literature using natural language processing.

The Major Biomedical Knowledge Graphs

Several large-scale biomedical knowledge graphs are publicly available and actively used in research.

Hetionet, developed through Project Rephetio, is an integrative biomedical network encoding knowledge from millions of biomedical studies. Hetionet v1.0 consists of 47,031 nodes of 11 types and 2,250,197 relationships of 24 types. It was originally constructed to systematically model drug efficacy using 755 existing drug-disease treatments, and its predictions validated on two external sets of treatments while providing pharmacological insights on epilepsy. Hetionet remains widely used as a benchmark for drug repurposing research.

PrimeKG (Precision Medicine Knowledge Graph), developed at Harvard Medical School's Zitnik Lab, integrates more than 20 biological databases into a unified schema with approximately 129,000 nodes across 10 biomedical entity types and more than 8 million relationships. PrimeKG was designed around a disease-centred framework and includes drug indication, contraindication, and off-label use edges, which are often absent from other knowledge graphs. It is accompanied by text descriptions of clinical guidelines for drugs and diseases to enable multi-modal analyses, and it has been used to support network-based drug repurposing and graph machine learning applications.

SPOKE (Scalable Precision Medicine Oriented Knowledge Engine), developed at UCSF, is a large heterogeneous biomedical knowledge graph that integrates a wide variety of biomedical databases into a single graph, capturing relationships among genes, proteins, diseases, drugs, and clinical concepts, with a focus on broad biomedical discovery and precision medicine.

These graphs are not equivalent, and each has distinct strengths. Hetionet is static and well-characterised as a benchmark. PrimeKG is more comprehensive and disease-rich but requires periodic updating as upstream ontologies change. SPOKE has a broader clinical integration. Choosing the right graph for a given research question is not a trivial decision, and most serious applications involve evaluating multiple graphs or building task-specific subgraphs from primary sources.

What Knowledge Graphs Enable in Biomedical Research

Drug repurposing. The most developed application of biomedical knowledge graphs is drug repurposing: computationally predicting new therapeutic uses for existing drugs. The logic is structural: if a drug inhibits protein A, and protein A interacts with protein B, and protein B is associated with disease X through a pathway not originally considered in the drug's development, a multi-hop path through the knowledge graph can surface that connection. This type of integrative approach, which connects diverse biomedical and clinical entities to exploit network topology, has enabled computational analyses for drug-target prediction, network-based drug repurposing, adverse drug reaction prediction, disease mechanism discovery, and biomarker prioritisation.

Knowledge graph embeddings formalise this by learning low-dimensional vector representations of entities and relations that preserve the graph's topology. These embeddings support link prediction: given all known relationships in the graph, which new drug-disease or drug-target links are most likely to exist? This is the machine learning task underlying most computational drug repurposing pipelines built on knowledge graphs.

Drug target identification and prioritisation. Knowledge graphs support target prioritisation by enabling researchers to query which genes sit at network positions implicated in a disease of interest, which of those genes are druggable based on known protein structure and existing chemical matter, and which have supporting evidence from multiple independent data layers. Biomedical knowledge graphs provide a structured and traceable framework for target-oriented, AI-assisted drug discovery, with applications in target prioritisation, druggability assessment, mechanism interpretation, drug-target interaction prediction, molecular generation, lead optimisation, and ADMET and safety assessment.

Disease mechanism interpretation. A knowledge graph can serve as a substrate for mechanistic reasoning by tracing paths between a genetic variant and a clinical phenotype through intermediate molecular entities. Rather than asking which gene is associated with disease, researchers can ask which paths through the graph connecting the gene to the disease are best supported by existing evidence. This supports hypothesis generation that is grounded in the network structure of known biology, not just pairwise associations.

Adverse event prediction. Relationships between drugs and adverse events, between chemical structures and known toxicological profiles, and between targets and off-target effects can all be encoded as edges in a knowledge graph. Traversing these relationships allows computational prediction of safety liabilities before clinical testing.

PrimeKG and Temporal Completeness

One under-appreciated challenge in knowledge graph-based research is temporal completeness. A knowledge graph built from databases as they existed at a given point in time will not reflect drug approvals, newly characterised gene functions, or revised disease associations that occur after that cutoff. A 2026 preprint reporting PrimeKG-Plus, a refreshed version of PrimeKG, identified 55 molecular entities approved after the original PrimeKG June 2021 cutoff, 46 of which were absent from the original graph. This means that drug repurposing analyses using the original PrimeKG could not consider nearly 50 approved drugs as potential candidates for repositioning.

This is a structural challenge for the field. The upstream ontologies that knowledge graphs depend on (Gene Ontology, Human Phenotype Ontology, MONDO, Comparative Toxicogenomics Database) update on independent, asynchronous cycles, continuously introducing new entities and deprecating outdated associations. A model trained on a 2021 snapshot will face substantial new knowledge by 2023 or 2024. Maintaining temporal relevance requires either continuous graph updating, which is technically demanding, or explicit acknowledgment of the graph's temporal scope in any downstream analysis.

Limitations That Are Not Yet Resolved

The current limitations of biomedical knowledge graphs are significant and should be stated plainly.

Heterogeneous data quality. Knowledge graphs aggregate information from sources of very different quality and curation standards. A relationship curated by hand from primary literature and a relationship inferred computationally from sequence similarity have fundamentally different evidentiary weight, but in the graph they can appear as equivalent edges. Most knowledge graphs do not adequately distinguish these evidence levels in a way that downstream reasoning can use.

Incomplete causal evidence. The relationships encoded in biomedical knowledge graphs are predominantly associative, not causal. An edge between a gene and a disease may reflect a statistically detected association from a GWAS, an experimental finding from a specific cell line, a computational prediction, or a curated literature claim. The graph cannot, by itself, tell you which of these relationships are causal, which are correlational, or which are context-dependent. Incomplete causal evidence is among the current limitations explicitly noted in recent comprehensive reviews of the field.

Static representations. Most knowledge graphs represent biology as if it were a fixed network of stable relationships, when in reality gene regulation, protein interactions, and disease mechanisms are highly dynamic, context-dependent, and cell-type-specific. A relationship that holds in one tissue type may not hold in another. A drug-target interaction that is relevant in the presence of a specific mutation may be absent in wild-type biology. Static graphs cannot capture this context-dependence without additional layering.

Privacy constraints. Incorporating patient-level clinical data into knowledge graphs, which would substantially strengthen the clinical evidence for drug-disease relationships, is constrained by privacy regulations and the practical difficulties of data sharing across institutions.

Experimental and clinical validation gap. Computational predictions from knowledge graph reasoning remain incompletely validated against wet-lab and clinical evidence. A drug repurposing prediction generated from network topology needs prospective experimental validation before it can be treated as a credible hypothesis for clinical development. This validation loop is slow, expensive, and inconsistently applied across the field.

The Role of LLMs in Knowledge Graph Research

Large language models (LLMs) have emerged as both a tool for improving knowledge graph construction and a source of new challenges. LLMs can extract entities and relationships from unstructured literature at scale, potentially accelerating the incorporation of new biological knowledge into graphs. But LLMs also generate plausible-sounding but incorrect relationships, a problem known as hallucination, which can introduce false edges into a knowledge graph with no obvious way to detect them.

Medical knowledge graphs often exhibit incompleteness due to knowledge gaps and structural limitations in their source ontologies, and LLMs offer promising capabilities for augmenting incomplete medical knowledge graphs. At the same time, while LLMs offer these capabilities, evaluating the reliability of their generated outputs remains a critical challenge, particularly in the absence of sufficient expert resources for validation.

Several hybrid approaches combine LLM-based reasoning with the structured, verifiable data in knowledge graphs: the LLM handles natural language interpretation and generates queries, while the knowledge graph provides the structured data source that anchors the response in verified relationships. This architecture uses the strengths of both without allowing the LLM's tendency to hallucinate to contaminate the underlying data.

What Knowledge Graphs Cannot Do Alone

Knowledge graphs are infrastructure for connecting and querying existing knowledge. They are not a mechanism for generating new biological knowledge from first principles. A knowledge graph can surface the connection between a genetic variant and a disease through multiple mechanistic paths. It cannot determine which of those paths is active in a specific patient, in a specific tissue, under a specific set of conditions, without additional experimental data.

This distinction matters for how knowledge graphs should be used in research. They are most powerful when used to generate prioritised, mechanistically grounded hypotheses that are then tested experimentally, and when the experimental results are used to update the graph. The value of a biomedical knowledge graph scales with the quality and completeness of the biological knowledge that has been encoded in it, and with the rigour of the reasoning applied to it.

The field is moving toward more dynamic, causally informed, and experimentally validated knowledge graphs, but that transition is ongoing. In the meantime, the existing infrastructure represents a genuine advance in the ability to navigate and reason over the fragmented landscape of biomedical knowledge, and its applications in drug repurposing, target identification, and disease mechanism research have produced results that justify continued development.

---

References

1. Bonner S et al. A review of biomedical datasets relating to drug discovery: a knowledge graph perspective. Briefings in Bioinformatics. 2022;23(6). https://arxiv.org/abs/2102.10062

2. Chandak P et al. Building a knowledge graph to enable precision medicine. Scientific Data. 2023;10:67. https://zitniklab.hms.harvard.edu/projects/PrimeKG

3. Lu Y et al. Biomedical knowledge graph: a survey of domains, tasks, and real-world applications. arXiv preprint. 2025. https://arxiv.org/pdf/2501.11632v2

4. Ye W et al. Biomedical knowledge graphs driving new target identification and novel drug discovery. Targetome. 2026. https://journal.hep.com.cn/targetome/EN/10.48130/targetome-0026-0025

5. Himmelstein DS et al. Systematic integration of biomedical knowledge prioritizes drugs for repurposing. eLife. 2017. https://www.biorxiv.org/content/10.1101/087619v3

6. PrimeKG-Plus: a refreshed and rare-disease-enriched precision medicine knowledge graph. bioRxiv preprint. 2026. https://www.biorxiv.org/content/10.64898/2026.07.14.738415.full.pdf

7. Pusch L, Conrad TOF. Combining LLMs and knowledge graphs to reduce hallucinations in biomedical question answering. arXiv preprint. 2024. https://arxiv.org/pdf/2409.04181

8. Task-driven knowledge graph filtering improves prioritizing drugs for repurposing. BMC Bioinformatics. 2022. https://push-zb.helmholtz-munich.de/deliver.php?id=33054

9. Can LLMs support medical knowledge imputation? arXiv preprint. 2025. https://arxiv.org/html/2503.22954v1