Designed specifically with machine learning and precision medicine applications in mind, PrimeKG integrates information across biological scales — from genes and proteins to pathways, phenotypes, anatomy, drugs, and diseases — into a single graph structure.
Its structure, relationship types, multimodal features, and accompanying rebuild scripts make it particularly suitable for graph neural networks (GNNs), link prediction, drug repurposing, and other machine learning workflows. But PrimeKG also has limitations that become important as soon as a model is expected to make clinically meaningful or time-sensitive inferences. Most notably, its edges carry no evidence dates and no confidence scores, and the commonly used release is effectively a static snapshot of a set of databases that continue to evolve.
What Is PrimeKG and Why Was It Built?
Before 2022, a researcher building a drug repurposing model barely had a pleasant choice. Hetionet, for instance, had drug–disease relationships but lacked the phenotype, pathway, and anatomical context needed to model disease heterogeneity. The alternative was stitching together six to eight databases by hand, resolving identifier mismatches, and hoping the result held together.
PrimeKG was developed by Chandak, Huang, and Zitnik to address this fragmentation. The goal was to create a holistic, multimodal, disease-centric knowledge graph that connects gene and protein biology up through anatomy, drug, and phenotype while remaining practical for machine learning. Rather than forcing every research group to rebuild the same integration pipeline from scratch, PrimeKG provides an integrated starting point.
The construction pipeline merges 20 source databases, including OMIM, DrugBank, MONDO, NCBI Gene, DisGeNET, STRING, Gene Ontology, Reactome, SemMedDB, HPO, SIDER, BGee, CTD, and UBERON. The integration process harmonised identifiers, constructed the graph, extracted its largest connected component, and added reverse edges. Python rebuild scripts were also released, allowing researchers to regenerate the graph as source databases change.
Another important feature is its multimodal design. Drug nodes include molecular fingerprints, while disease and drug descriptions derived from clinical guidelines provide textual information. This allows PrimeKG to serve as input for multimodal models without requiring researchers to construct every feature representation independently.
That ML-first philosophy is what separates PrimeKG from many traditional biomedical knowledge graphs. It was designed not merely to answer queries, but to provide a structured substrate on which computational models can learn.
What Can PrimeKG Enable?
1. Drug repurposing with GNNs
This is PrimeKG’s best-known application. By training link prediction models on the drug–disease subgraph, researchers can identify approved drugs whose structural graph position suggests efficacy against diseases outside their current indication. The graph’s explicit separation of indication, contraindication, and off-label use edge types gives these models more signal than graphs that collapse all drug–disease relationships into a single undifferentiated edge.
2. Drug–target interaction prediction
PrimeKG connects drugs to protein targets through DrugBank and STRING, those proteins to diseases through DisGeNET, and diseases to phenotypes through HPO. A GNN trained across these multi-hop paths can predict novel drug–target interactions from the neighbourhood structure of known ones. This is a task where PrimeKG’s cross-database integration offers a real advantage over single-source alternatives.
3. Multimodal biomedical AI
The combination of graph topology, molecular fingerprints, and clinical text descriptions makes PrimeKG a practical backbone for graph-augmented LLM systems. Models like BioMedKG and Fact Finder retrieve structured subgraphs from PrimeKG to ground biomedical reasoning, anchoring generation in typed, curated relationships rather than leaving the model to rely entirely on learned text patterns. It provides a meaningful lever against hallucination in clinical contexts.
4. Continual learning and evolving benchmarking (PrimeKG-CL)
The 2025 PrimeKG-CL paper rebuilt PrimeKG from source databases at two time points — June 2021 and July 2023 — producing the first biomedical KG benchmark with genuine temporal evolution. Between the snapshots, 5.83 million edges were added, 889,000 were removed, and 7.21 million remained persistent across the interval. For anyone building models meant to survive contact with an evolving knowledge base, this is the first benchmark that actually measures that capacity rather than assuming it.
5. Cross-disease network analysis
With more than 17,000 diseases sharing genes, pathways, phenotypes, and drugs in a single graph, PrimeKG’s disease-centric structure also makes it useful for analysing relationships across diseases. It supports comorbidity analysis and cross-indication discovery at a scale that spans over 100,000 nodes and 4 million relationships — something single-disease studies cannot reach.
The Limitations
The central issue with PrimeKG is not its size. It is what the edges do not tell you.
Zero evidence dates
PrimeKG contains 4,050,249 undirected relationships and zero evidence dates at the edge level. Query it for “which drugs treat disease X” and there is no way to tell a 1995 consensus from a 2024 guideline revision, a single case report from a large randomised trial, or a manually curated entry from an NLP co-occurrence extraction. Of everything in this review, this is the fact to internalise before building any clinical inference system on top of PrimeKG.
A static snapshot of dynamic databases
GO, HPO, MONDO, and CTD all update on independent release cycles, and PrimeKG does not track any of them automatically. The scale of that drift is not hypothetical: PrimeKG-CL’s own two-snapshot comparison found 5.83 million edges added and 889,000 removed in roughly two years. A model trained on the 2022 release is training on a graph that has already diverged materially from its own source databases. The rebuild scripts exist precisely to address this, but they require researchers to actively maintain and access the underlying databases.
No edge-level confidence scores
This matters because not all biomedical relationships are equally well established. A relationship supported by extensive experimental or clinical evidence should not necessarily carry the same evidentiary weight as one inferred from a single report. Yet without additional provenance handling, downstream models can treat both as graph relationships of comparable status.
SemMedDB-derived noise
These edges deserve separate scrutiny. They are extracted from the literature by NLP rather than curated by biologists, and Semantic MEDLINE’s extraction accuracy varies notably by relation type — inhibition and activation relationships in ambiguous text are a known weak point. Edges sourced this way should carry lower confidence than those drawn from DrugBank or OMIM curation.
One Technical Detail Researchers Should Not Miss
The directed/undirected edge count is a documented source of confusion worth stating plainly. The paper describes “30 types of undirected edges” and reports 4,050,249 relationships, but the underlying data file contains 8,100,498 rows — two directed rows per undirected relationship. Pipelines that ingest the file without accounting for this will silently double their training signal.
Researchers should decide explicitly whether their analysis uses directed or undirected edges, deduplicate where appropriate, and document the decision in the methods. This is not merely a documentation detail: graph construction choices can affect model behaviour and reproducibility.
How to Use PrimeKG Correctly
First, make the directed-versus-undirected decision explicit when loading kg.csv or edges.csv during data ingestion. Do not assume that the 8.1 million rows represent 8.1 million independent biological relationships.
Second, consider temporal sensitivity. If the research question involves recently approved drugs, emerging gene–disease associations, or changing clinical guidelines, the PrimeKG-CL snapshots can help determine whether conclusions are sensitive to graph version.
Third, pay attention to provenance. Where evidence quality matters, the rebuild pipeline can be used to identify source databases and distinguish manually curated relationships from NLP-derived ones.
Finally, for applications where currency matters, use the available Python rebuild scripts rather than assuming the original snapshot represents the current state of the source databases. When publishing results, specify the PrimeKG version and whether the original release or a rebuilt graph was used.
What Is Changing in the PrimeKG Ecosystem?
PrimeKG is increasingly becoming part of a broader effort to make biomedical knowledge graphs temporal, provenance-aware, and suitable for evolving AI systems.
PrimeKG-CL is particularly important because it turns the graph’s temporal limitations into an explicit continual-learning benchmark. Its results indicate that no single continual-learning strategy consistently dominates across tasks, with decoder selection and continual-learning strategy interacting substantially.
Other developments address limitations that PrimeKG leaves open. ChronoMedKG, for example, is designed around evidence provenance and reportedly provides publication-date traceability for 98.9% of its triples. For applications where evidence currency is central, this type of temporal representation is worth monitoring alongside PrimeKG.
At the same time, PrimeKG’s multimodal structure remains useful as graph-augmented LLM and multimodal GNN approaches develop. Its pre-integrated molecular and clinical-text features can reduce the feature-engineering burden associated with building these systems from separate resources.
The Bottom Line
PrimeKG is the most ML-ready disease-centric knowledge graph currently available as open data. For drug repurposing, drug–target prediction, and multimodal biomedical AI, it offers a starting point that would otherwise take months to assemble from scratch. The rebuild scripts mean it does not have to remain a static resource — though in practice, most research groups still work from the unmodified 2022 snapshot.
Use it with a clear view of what it lacks: no evidence dates, no confidence scores, no clinical trial or patent data, and a snapshot that has already drifted from its source databases. None of these are disqualifying. They are constraints, and they should shape how you interpret the outputs of anything trained on this graph.
PrimeKG did not build the perfect biomedical knowledge graph. It built the most practically useful one, along with the tools to improve it. The rest is up to the community.
Call to Action
PrimeKG shows how much an integrated, ML-ready graph unlocks — and how much still has to be checked by hand, because an edge with no date and no confidence score cannot tell you how much to trust it. Axy is built for that second half: folding scientific literature into a knowledge graph where every link is auditable, conflicting evidence is preserved rather than flattened, and every change passes human review.
If your work involves drug repurposing, precision medicine, or literature-scale discovery, apply to join below and help build the shared, connected map of science rather than just use another tool.
