HRA KG: What the Human Reference Atlas Knowledge Graph Maps, Enables, and Leaves Out

Researchers working with single-cell or spatial transcriptomics data run into the same obstacle again and again: the biology may be comparable, but the way it is represented is not.

One dataset labels a population “CD16+ monocyte.” Another calls the same cells “non-classical monocyte.” Neither tells you where in the kidney, the lung, or the lymph node those cells actually sit. Anatomy described against 2D references carries no consistent spatial coordinate. A marker used to define a cell population in one study is rarely tied to the anatomical structure that population occupies. Bringing two datasets together therefore means reconciling terminology, anatomy, cell types, and biomarkers by hand before any biological analysis can begin.

The Human Reference Atlas Knowledge Graph (HRA KG) was built to remove that reconciliation step. Instead of treating anatomy, cell types, biomarkers, and spatial position as four separate resources, it connects them inside a single knowledge graph describing the healthy adult human body.

Understanding both what it covers and what it deliberately excludes is the difference between using it well and being misled by it.

What Is the HRA Knowledge Graph?

The NIH-funded Human BioMolecular Atlas Program (HuBMAP) produced the Human Reference Atlas (HRA): a three-dimensional, multi-scale, expert-curated atlas of the healthy adult body that runs from whole organ down to single cell, with computational access to every level.

HRA KG v2.2, published in Scientific Data in July 2025, is the knowledge-graph formalization of that atlas. It is built as RDF, queryable through SPARQL, and deployed as Linked Open Data at lod.humanatlas.io.

A cross-scale architecture

The core architecture links three biological layers:

• Anatomical structures (AS) are referenced to the Uberon ontology.
• Cell types (CT) are referenced to the Cell Ontology (CL) and the Provisional Cell Ontology (PCL).
• Biomarkers (B) include genes, proteins, and lipids.

These layers are joined through ASCT+B tables — expert-curated records of which cell types occur within which anatomical structures, and which biomarkers characterize those cell types. What the knowledge graph adds on top of the tables is the ability to query those relationships programmatically across all 71 organs at once, rather than reading them off one organ at a time.

The graph also incorporates 13 types of Digital Objects, among them 3D reference organs, 2D functional tissue unit illustrations, organ mapping antibody panels (OMAPs), and registered tissue datasets. All of them are unified through the Common Coordinate Framework Ontology (CCFO), which standardizes concepts and relationships across every Digital Object type.

The numbers, as of the May 2025 release

What the HRA KG Enables

Querying across scales, from organ to cell type to biomarker

The graph’s central capability is traversing biological relationships across scales. Consider a question that sounds straightforward:

Which cell types occur in the kidney cortex, and which biomarkers define them?

Answering that conventionally means moving between an anatomy reference, a cell-type resource, and a marker database, then reconciling three vocabularies. The HRA KG holds those relationships in one graph, so a single SPARQL query can span the anatomical structure (AS) → cell type (CT) → biomarker (B) path. The same query shape works across all 71 organs represented in the current version.

Registering tissue datasets in spatial context

The graph also connects molecular datasets to the spatial organization of the body. Tissue blocks generated through single-cell sequencing or spatial transcriptomics can be registered within the HRA’s 3D Common Coordinate Framework (CCF), which gives every experimental sample a standard spatial address inside the broader anatomical atlas.

Once a block is registered, its cell types and biomarkers inherit that spatial and anatomical context. Cross-lab queries such as “show me every registered tissue block from the left kidney cortex with CD4+ T cell annotations” become answerable, because every registered dataset is normalized to the same spatial reference rather than to whatever internal convention the originating lab used.

For spatial transcriptomics in particular this matters, because spatial context is otherwise very difficult to compare consistently across independent experiments. A shared coordinate framework is the common reference those datasets can be interpreted against.

Cell type annotation with Azimuth

Researchers starting from unlabelled single-cell RNA-seq data have a far more accessible entry point: Azimuth. It maps single-cell RNA-seq data against reference cell-type hierarchies, assigning labels tied to the HRA’s ontology framework.

That link is what turns a dataset-specific annotation scheme into an interoperable one. Instead of keeping local labels, researchers connect their cell populations to HRA-supported cell type identifiers and hierarchies. This is probably the most heavily used practical entry point into the HRA ecosystem, because it requires no SPARQL knowledge and produces annotations that line up with every other HRA-registered dataset.

Improving the community ontologies

The HRA’s contribution is not confined to its own graph. As of May 2025, HRA expert editors had contributed 162 terms to Uberon, 155 to the Cell Ontology, and 468 to PCL — including 461 brain cell-type terms.

Those additions become part of the wider ontology ecosystem rather than staying inside HRA infrastructure. Other biomedical tools and databases built on the same community ontologies inherit the curation work directly.

A shared reference across consortia

The HRA KG functions as a common backbone for more than 25 consortia, including HuBMAP, KPMP (kidney), GTEx, and the BRAIN Initiative. Datasets produced by different groups can reference the same anatomical structures, cell types, and biomarkers instead of study-specific terminology, which makes cross-consortia integration tractable and cuts the manual harmonisation work that normally precedes it.

The Limitations of the HRA KG

“Healthy, adult body” is not a footnote on this resource. It is the design decision everything else follows from. Knowing which of your research questions sit inside that scope — and which do not — matters more than any other single fact about this graph.

Healthy adult does not mean disease atlas

Disease tissue, pediatric anatomy, and the fuller range of demographic variation sit outside the current scope. A tumour microenvironment researcher or a paediatric cardiologist cannot point the HRA KG at their tissue and expect a direct match.

What they can do is use it as the healthy baseline against which deviation is measured. That is a real and useful role, but it is not the same thing as a disease-state atlas.

SPARQL creates an access barrier

The graph’s programmatic flexibility comes with a practical cost. Custom queries are served primarily through the HRA’s SPARQL endpoint, so working directly with the graph requires familiarity with RDF, ontologies, and SPARQL query construction.

This bites hardest on multi-hop queries across AS → CT → B relationships: a poorly formed query tends to return an empty result set with almost no diagnostic signal explaining why. The companion site at cns-iu.github.io/hra-kg-supporting-information offers example queries, and the HRA API exposes pre-built endpoints, but as it stands the access model favours researchers who already have ontology engineering or semantic web experience.

PCL is changing rapidly

The Provisional Cell Ontology has expanded quickly since mid-2024, particularly with the addition of brain cell types. That growth reflects improving biological resolution, but it also introduces a reproducibility problem for computational workflows: cell-type identifiers may be deprecated, renamed, or reorganised between versions.

A pipeline that hardcodes PCL identifiers has to account for ontology versioning. For anyone building long-lived analysis pipelines, version awareness is not optional.

Organ coverage is extensive but not complete

Seventy-one organs are covered in v2.2, but tissues with especially complex or variable anatomy — lymph nodes and peripheral nervous system structures among them — are only partially mapped.

Literature linkage sits outside the graph

HRAlit, the literature database connecting HRA entities to published research, is currently served from a relational database rather than integrated into the RDF graph. Cross-querying publications alongside atlas data means a separate access step, not a single SPARQL call.

Demographic variation is not yet systematically encoded

The graph represents a healthy adult reference, but variation by sex, ancestry, and BMI is not yet built into the data structure in a queryable way.

How to Use the HRA KG Correctly

Where you enter the HRA ecosystem should match what you are trying to do.

Whatever access method you use, version pinning matters. Treat HRA KG v2.2 as a specific data release, not an interchangeable static reference. Record the exact HRA version in your methods, and pull versioned data and code from the companion repository for reproducibility — especially when PCL identifiers are involved.

For disease work, the intended workflow is comparative, not substitutive. Register experimental data into the CCF to assign spatial and anatomical context, then characterize cell type composition and biomarker expression against the healthy HRA reference. Using the healthy baseline to make disease deviation quantifiable is legitimate and increasingly common — but it is a different claim from treating the HRA KG as a description of diseased tissue.

When citing the resource, use the final Scientific Data version rather than the bioRxiv preprint:

Bueckle A, Herr BW, Hardi J, Quardokus EM, Musen MA, Börner K. “Construction, Deployment, and Usage of the Human Reference Atlas Knowledge Graph for Linked Open Data.” Scientific Data 12, 1100 (2025). DOI: 10.1038/s41597-025-05183-6

Specify the HRA version and your data access date in the methods section.

What Is Changing Next

Seventy-one organs is not a ceiling. Organs with complex or variable anatomy — lymph nodes, peripheral nervous system components, reproductive tissues — are in active development, and new ASCT+B tables and 3D reference objects are ingested as expert teams complete them. Researchers working in underrepresented organ systems should check the release notes before assuming their organ of interest is permanently absent.

HuBMAP’s stated ambition also extends past the healthy baseline, toward ageing, disease, and other biological perturbations. The healthy adult atlas is the foundation for those comparisons rather than the endpoint. Data contributions from the CZI-funded Human Cell Atlas and the NIH BRAIN Initiative extend the graph toward disease and developmental contexts as well.

HRAlit integration is on the roadmap. Folding the literature linkage database into the RDF graph as a full component would allow queries connecting cell types and anatomical structures directly to the publications that characterize them — collapsing a two-step data access process into one, much as PubMed Knowledge Graph 2.0 did for papers, patents, and trials.

Finally, the structured RDF format makes the graph a natural candidate for graph-augmented AI. A researcher should eventually be able to ask “which cell types occur in the kidney glomerulus, and what biomarkers define them?” in natural language, with a system translating that request into the appropriate structured query. That kind of interface removes much of the current SPARQL barrier without giving up the underlying structured representation.

Where This Reference Layer Leaves Researchers

The HRA KG is the most rigorously structured, ontology-grounded, cross-scale knowledge graph of the human body currently available as open data. For single-cell researchers, spatial transcriptomics practitioners, and anyone building tools that need a standardized human anatomical reference, it belongs in the workflow.

It is a healthy-adult reference, not a comprehensive disease atlas. It covers 71 organs, it is actively versioned, and its most flexible query capabilities sit behind SPARQL. Those constraints are not weaknesses so much as consequences of building a scientifically controlled reference instead of attempting to represent every biological state at once.

The HRA KG does not map the sick human body. It maps the healthy one with enough structure and precision to give researchers a reference against which disease-related change can be measured, quantified, and compared. That reference layer is what much of the single-cell and spatial biology ecosystem has been missing.

Call to Action

The value of a knowledge graph depends on how easily its structured information reaches the tools you already use. The HRA KG shows what a rigorously curated reference layer makes possible — and how much work still sits between a SPARQL endpoint and a working research workflow. Axy is built for that gap: a scientific knowledge graph where every link is auditable, conflicting evidence is preserved rather than flattened, and every change passes human review.

If your work involves single-cell datasets, spatial transcriptomics, anatomical annotation, or literature-scale discovery, apply to join below and help build the shared, connected map of science rather than just use another tool.