Biomedical research has never lacked information. It frequently fails to connect it.
Every day, thousands of new papers describe novel genes, therapeutic targets, disease mechanisms, and experimental findings. Those results live in PubMed, in Google Patents, and in ClinicalTrials.gov — different stages of the same translational journey, held in systems with incompatible identifiers, varying metadata standards, and almost no interoperability. Connecting them by hand is slow, error-prone, and increasingly impractical as the literature keeps growing.
PubMed Knowledge Graph 1.0, released in 2020, was the first serious attempt to link papers, authors, affiliations, biomedical entities, and NIH funding. It was useful for bibliometrics and impact tracking, but it stayed confined to a single document type. PKG 2.0 breaks that boundary: patents and clinical trials now join the graph, linked to papers through shared biomedical entities, citation links, and NIH project identifiers. The result is infrastructure for studying how biomedical discoveries move from laboratory research to clinical application — and, at the same time, a probabilistic construction pipeline whose limitations you need to understand before drawing analytical conclusions from it.
This review covers what PKG 2.0 contains, what it enables, where it falls short, and how to use it responsibly.
What PKG 2.0 Is and Why It Was Built
PKG 2.0 first appeared on arXiv in October 2024, then in peer-reviewed form in Scientific Data (Nature) in June 2025. The dataset is available at pubmedkg.github.io, with a PKG25S4 preview update already announced.
It is currently one of the largest biomedical knowledge graphs available, containing more than 36 million PubMed papers, 1.3 million patents, over 480,000 clinical trials, 482 million biomedical entity linkages, 19 million citation links, and 7 million NIH project linkages.
The construction pipeline relies on state-of-the-art natural language processing: the BERN2 biomedical entity extraction system, a high-performance author disambiguation model achieving 98.09% F1 against the PKG 1.0 baseline, multi-source citation integration, and NIH ExPORTER for funding and project linkages.
For biomedical researchers, data scientists, and AI developers, PKG 2.0 is a significant step toward machine-readable biomedical knowledge.
What the Three-Domain Integration Enables
For researchers working in translational medicine, drug discovery, scientometrics, and biomedical AI, a unified graph opens analytical possibilities that were previously impractical or impossible.
Translational research mapping. This is PKG 2.0’s core value proposition: tracing the path from scientific finding to patent claim to clinical application. A gene-disease association published in 2012, patented by a pharmaceutical company as a drug target in 2015, and carried into a Phase II trial by 2019 is now queryable as a single trajectory. That infrastructure simply did not exist before.
The capability matters most in drug repurposing, target validation, competitive intelligence, and technology landscape analysis — fields where understanding how a laboratory discovery becomes a clinical application is often as important as the discovery itself.
Large-scale author and collaboration network analysis. Author disambiguation is a chronic problem in biomedical bibliometrics. Researchers publish under common surnames, use different naming conventions across journals, and change institutions mid-career, which produces both false merges and false splits in citation networks and undermines collaboration analyses.
PKG 2.0’s disambiguation model achieved an F1 score of 98.09% on the PKG 1.0 benchmark and links authors to ORCID wherever coverage exists. That makes it meaningfully more reliable than raw PubMed metadata for tracking research communities — mapping international collaboration networks, identifying emerging research communities, tracking institutional partnerships, and studying knowledge diffusion across disciplines.
Funding-to-output linkage via NIH ExPORTER. The 7 million project linkages connecting NIH grants to papers, patents, and trials make PKG 2.0 distinctly useful for research policy analysis, in ways that extend well beyond traditional literature mining.
Funding agencies can examine how specific grant programs translate into scientific outputs. Universities can evaluate how publicly funded research contributes to intellectual property and clinical innovation. Researchers can ask which NIH initiatives generate the highest clinical impact, how long funded discoveries typically take to reach a patent, and which funding mechanisms produce the largest collaboration networks. None of that is answerable from PubMed or patent databases alone, and it is arguably the graph’s most distinctive feature for institutional research offices and science policy researchers.
Biomedical entity co-occurrence and relationship mining. Across papers, patents, and trials, PKG 2.0 links 482 million entities spanning nine types: gene/protein, disease, drug/chemical, species, mutation, cell line, cell type, DNA, and RNA. That makes it a rich substrate for cross-domain questions — which genes co-occur in patents but not in papers, or which diseases have active trials with no corresponding patent activity.
Because entities are standardized rather than treated as plain text, these relationships can be queried efficiently across millions of records, which is what makes the graph useful for hypothesis generation, literature exploration, and systems-level biomedical research.
LLM grounding for biomedical AI. Traditional retrieval-augmented generation retrieves passages of text that may carry no explicit relationships between concepts. A structured, typed entity-relationship graph is precisely the kind of verified knowledge base that addresses hallucination risk in biomedical LLM applications.
A model generating a drug-mechanism claim can be checked against PKG 2.0’s linkages, and a RAG system can pull structured subgraphs instead of unstructured text chunks. For AI-assisted literature synthesis, biomedical question answering, hypothesis generation, and translational research workflows, that is a real step up in reliability.
The Limitations
The authors acknowledge directly that biases and inaccuracies are unavoidable, because PKG 2.0 relies on probabilistic and statistical models for entity extraction, author disambiguation, and institutional disambiguation. Treat it as a powerful analytical resource, not as infallible ground truth — entity-level precision matters.
The most consequential structural limit is abstract-only extraction. BERN2 pulls entities from titles and abstracts, not full text. If a paper’s key finding sits in a table, a figure legend, or the supplementary methods, that entity does not appear in the graph at all. Co-occurrence analyses built on PKG 2.0 are analyses of what abstracts say, not of what papers actually report. That does not disqualify the graph, but it has to shape how you interpret co-occurrence patterns.
A few other limits are worth flagging directly:
- No launch-time user interface. Access requires SQL or programmatic handling, which puts the dataset out of reach for researchers without database or scripting skills.
- PubMed-scope only. There is no coverage of ecology, materials science, physics, or other fields outside PubMed’s index.
- Citation data lags the preprint era. arXiv, bioRxiv, and medRxiv are incompletely indexed, so fast-moving fields show citation networks that trail actual knowledge flow.
- Patent coverage is partial. At roughly 1.3 million patents, non-US and non-English patents are underrepresented relative to the global patent literature.
- Snapshot staleness. PKG 2.0 is versioned by update cycle, so fast-moving areas — COVID-19 therapeutics, LLMs in biomedicine, cancer immunotherapy — can lag meaningfully between releases. The PKG25S4 preview signals active maintenance, but check the version date against your analysis window regardless.
How to Use PKG 2.0 Correctly
The dataset lives at pubmedkg.github.io in structured table format, parsed from PubMed XML with additional tables for NIH ExPORTER, ORCID, and other integrated sources. There is no web query interface at launch; SQL or programmatic access is required, though the authors have flagged a friendlier interface as a planned improvement.
Four practical rules follow directly from the limitations above.
Match your analysis to the extraction scope. If your question depends on relationships buried in methods or results sections rather than abstracts, PKG 2.0 will systematically miss a subset of them. Run full-text mining tools such as PubTator 3 or the Europe PMC annotator alongside it for precision-sensitive work.
For author-level analysis, filter to high-confidence disambiguation. PKG 2.0 ships disambiguation confidence scores — apply a threshold, and validate against ORCID where coverage exists, especially for individual-researcher or small-network studies.
For cross-domain queries, verify edge provenance. A linkage connecting a paper to a patent is a shared entity co-occurrence, not a citation. A paper and a patent sharing a gene name are linked in the graph; that does not mean the patent cites or builds on the paper. Keep co-occurrence linkages and citation linkages analytically distinct.
Cite the version you used. The graph changes between snapshots. Cite the specific version (for example PKG25S4) and the access date in your methods — reproducibility depends on pinning to a version.
What Is Changing
The PKG25S4 update is already in preview, promising updated data synchronization and enhanced cross-database entity links. If you are building a pipeline on top of PKG 2.0, it is worth tracking. Several developments beyond it are shaping how researchers will use biomedical knowledge graphs over the next few years.
Full-text extraction is the obvious next frontier. The most significant improvement PKG could make is extending entity extraction beyond titles and abstracts. As open-access mandates expand PubMed Central’s full-text coverage and NLP models get more efficient at processing long documents, abstract-only extraction becomes an avoidable limitation rather than a structural one. The PMC Open Access subset already covers a substantial share of recent papers, so the groundwork exists. For researchers studying complex molecular pathways, drug mechanisms, or experimental methodologies, that transition would substantially improve both graph completeness and analytical precision.
LLM integration is a natural next step. Large language models excel at synthesizing information but struggle with factual consistency, particularly in specialized scientific domains. Structured knowledge graphs supply a layer of verified relationships that models can ground their responses against. PKG 2.0’s typed, entity-linked structure suits graph-augmented retrieval (GraphRAG), where a model pulls structured subgraphs instead of raw text chunks. As GraphRAG architectures mature in biomedical AI, PKG 2.0 is positioned to serve as a knowledge substrate for scientific question answering, automated literature synthesis, hypothesis generation, and AI-assisted drug discovery.
Cross-graph integration is the more ambitious frontier. No single biomedical knowledge graph captures every aspect of biological knowledge. Connecting PKG 2.0 to the Monarch Initiative (gene-phenotype associations), PrimeKG (drug-disease-gene triples), and GBIF for One Health research linking human disease to ecological systems would enable queries no single graph currently supports. The prerequisite is robust entity identifier alignment — mapping PKG 2.0’s BERN2-normalized entities to Monarch’s HPO terms and PrimeKG’s DrugBank IDs. Those integrations would let researchers move beyond literature-centric analyses into relationships spanning molecular biology, clinical medicine, biodiversity, environmental health, and translational science.
Conclusion
PKG 2.0 is the most ambitious attempt so far to connect the three pillars of biomedical knowledge — published research, patent claims, and clinical evidence — into a single queryable graph. For translational research mapping, bibliometric analysis, and funding-to-output tracing, it is infrastructure that did not previously exist.
Use it with the precision its pipeline warrants: understand the abstract-only scope, apply disambiguation confidence thresholds, keep co-occurrence and citation linkages separate, and cite the version. Used carefully, it is a genuinely powerful substrate for biomedical AI, GraphRAG systems, cross-domain knowledge graphs, and research synthesis. Calibrated use beats both uncritical adoption and unnecessary avoidance.
PKG 2.0 does not make biomedical knowledge management a solved problem. It makes it, for the first time, a tractable one — and that is a meaningful advance.
Call to Action
Biomedical knowledge is only as useful as the connections you can trace through it. PKG 2.0 shows how much that unlocks, and also how much still has to be checked by hand — provenance verified edge by edge, co-occurrence separated from citation, versions pinned. Axy is built for that second half: folding literature into a scientific knowledge graph where every link is auditable, conflicting evidence is preserved rather than flattened, and every change passes human review.
If you want to help build the shared, connected map of science rather than just use another tool, apply to join below.
