{"paper_id":"2166a4ed-4f80-4600-89a1-cb97f52f55f6","body_text":"Construction, Deployment, and Usage of the\nHuman Reference Atlas Knowledge Graph for\nLinked Open Data\nAndreas Bueckle 1*, Bruce W. Herr II 1*, Josef Hardi 2, Ellen M. Quardokus 1, Mark A. Musen 2, Katy Börner 1*\n1 Department of Intelligent Systems Engineering, Luddy School of Informatics, Computing, and Engineering,\nIndiana University, Bloomington, IN, USA\n2 Stanford Center for Biomedical Informatics Research, Stanford University, Stanford, CA, USA\n* Corresponding authors\nAbstract\nThe Human Reference Atlas (HRA) for the healthy, adult body is developed by a team of international,\ninterdisciplinary experts across 20+ consortia. It provides standard terminologies and data structures for\ndescribing specimens, biological structures, and spatial positions of experimental datasets and ontology-linked\nreference anatomical structures (AS), cell types (CT), and biomarkers (B). We introduce the HRA Knowledge\nGraph (KG) as central data resource for HRA v2.2, supporting cross-scale, biological queries to Resource\nDescription Framework graphs using SPARQL. In December 2024, the HRA KG covered 71 organs with 5,800\nAS, 2,268 CT s, 2,531 Bs; it had 10,064,033 nodes, 171,250,177 edges, and a size of 125.84 GB. The HRA KG\ncomprises 13 types of Digital Objects (DOs) using the Common Coordinate Framework Ontology to\nstandardize core concepts and relationships across DOs. We (1) provide data and code for HRA KG\nconstruction; (2) detail HRA KG deployment by Linked Open Data principles; and (3) illustrate HRA KG usage\nvia application programming interfaces, user interfaces, and data products. A companion website is at\nhttps://cns-iu.github.io/hra-kg-supporting-information.\nIntroduction\nThe multimodal, three-dimensional (3D) Human Reference Atlas (HRA) 1 aims to map the healthy, adult human\nbody across scales—from the whole body to the single cell and biomarker levels. Data from different sources\n(organs, technologies, and labs), many built with standard operating procedures (SOPs,\nhttps://humanatlas.io/standard-operating-procedures), need to be integrated. The HRA Knowledge Graph (KG)\ndefines and provides the core data structures that are used to store, link, and query HRA data.\nKGs are widely used to interlink data about relevant entities within a specific domain or task. The Google\nKnowledge Graph (https://blog.google/products/search/introducing-knowledge-graph-things-not) supports\nGoogle Search with its billions of searches processed daily. Major online shopping retailers such as Amazon 2\nuse knowledge graphs to organize products, searches, and media items. KGs across domains are structured\nusing vocabularies, e.g., Friend of a Friend (FOAF , http://xmlns.com/foaf/spec), Simple Knowledge\nOrganization System (SKOS, https://www.w3.org/2004/02/skos), and Music Ontology\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\n(http://musicontology.com). Plus, there exist collaborative efforts for publishing structured data on the web. For\nexample, https://schema.org promotes the structured representation for data on the web and is used in\napplications from Google, Microsoft, and Pinterest to create data-driven web experiences. An overview of other\ncommonly used vocabularies is available on\nhttps://www.w3.org/wiki/T askForces/CommunityProjects/LinkingOpenData/CommonVocabularies.\nIn the biomedical domain, ontologies are widely used to structure data, which is of high relevance to HRA KG\nconstruction. For example, the National Center for Biomedical Ontology (NCBO) BioPortal provides easy\naccess to 1,168 ontologies and the EMBL-EBI Ontology Lookup Service (OLS) supports 267 ontologies. The\nUber-anatomy ontology (Uberon, https://www.ebi.ac.uk/ols4/ontologies/uberon)3 is a cross-species ontology\nrepresenting body parts, organs, and tissues, primarily focused in vertebrates. The Cell Ontology (CL,\nhttps://obofoundry.org/ontology/cl.html)4 is also a cross-species ontology, but it focuses on classifying and\ndescribing cells. These two ontologies are linked data, so we can determine assertions such as kidney nerve\ncell (CL:1000606) has location (RO:0002100) kidney (UBERON:0002113) from the knowledge represented in\nthe ontologies.\nOntologies are an indispensable part of generating, using, and maintaining KGs as they enable unifying\nnomenclature across assay types, organs, donors, teams, and consortia. A recent publication by He et al. 5,\nfeaturing the HRA, shows how ontologies can be used to model, integrate, and reason over previously siloed\nclinical, pathological, and molecular kidney data for precision medicine. It highlights the development of the\nprecision medicine metadata ontology (PMMO) to integrate dozens of variables between the Kidney Precision\nMedicine Project (KPMP ,https://www.kpmp.org)6,7 and Chan Zuckerberg Initiative (CZI) CELLxGENE (CxG)\ndata (https://cellxgene.cziscience.com). It then shows specific use cases in detecting healthy vs. acute kidney\ninfection (AKI)/chronic kidney disease (CKD) disease states in cells supported by PMMO, Kidney Tissue Atlas\nOntology (KTAO), and the HRA’s CCF Ontology, described in a related publication 8.\nBiomedical KGs use and interlink multiple ontologies to store and query data. For example, the Unified Medical\nLanguage System (UMLS) 9 “metathesaurus” contains approximately 3.4 million biomedical concepts, updated\nevery 6 months in May and November and is derived from other biomedical terminologies and ontologies. The\nPetagraph KG 10 uses the UMLS metathesaurus to integrate biomolecular datasets and connects them to\napproximately 200 cross-referenced ontologies to support exploration of gene variant epistasis as well as\nbiological assertions with reduced dimensionality, and to link relevant features to chromosome position and\nchromosomal neighborhoods. The Human BioMolecular Atlas Program (HuBMAP ,\nhttps://hubmapconsortium.org)11,12 Unified Biomedical Knowledge Graph 1 (UBKG) connects HuBMAP\nexperimental data to ontologies. The Scalable Precision Medicine Open Knowledge Engine\n(SPOKE,https://spoke.ucsf.edu)13,14 processes 41 databases (53 million edges) and 11 ontologies to create an\nintegrated graph with user access via a Representational State Transfer (REST) application programming\ninterface (API). Petagraph, HuBMAP , and SPOKE use the Neo4J graph platform (https://neo4j.com ). Efforts\nlike BioCypher (https://biocypher.org) 15 enable the rapid construction and maintenance of KGs at lower cost.\nThis also addresses the lack of reusability and integrability, where KGs are built manually for a specific task,\nand, as a result, do not adhere to Findable, Accessible, Interoperable and Reusable (FAIR) 16 principles. KGs\ncan be used to extract knowledge across constantly evolving ontologies and data in various states of\naccessibility (private and public).\nIn this paper, we present the HRA KG v2.2, which uses 10 ontologies to interlink 34 anatomical structures\n(AS), cell types (CT), plus biomarkers (B) tables (see Box 1), 71 3D Reference Objects for organs,, 22\nFunctional Tissue Units (FTUs) 17, 11,698 single-cell (sc) datasets, and other HRA Digital Objects (DOs, see\nResults > HRA Digital Objects), together with the HRA API that supports efficient queries using REST ,\nJavaScript, TypeScript, Angular 17+, or Python 3.6+ code plus several interactive user interfaces (UIs).\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nSpecifically, we present open code and infrastructure to construct the HRA KG out of disparate data across\ntabular/non-tabular and nested/flat HRA DOs while ensuring processed data conforms to 5-star Linked Open\nData (LOD) 18 principles. The HRA KG data can be accessed via content negotiation from\nhttps://lod.humanatlas.io and dynamically queried via its SPARQL (see Box 1) endpoint at\nhttps://lod.humanatlas.io/sparql.\nBox 1. Key technologies used for constructing, deploying, and using the HRA KG.\n● 3D Reference Objects: Mesh-based 3D models describing organs in the male or female body of the HRA. They\nare used in HRA applications to register and explore tissue blocks and associated datasets 19. All 71 3D Reference\nObjects of HRA v2.2 have crosswalks that link individual 3D AS to ontology terms in Uberon or Foundational\nModel of Anatomy (FMA) 20,21.\n● ASCT+B tables: Anatomical structures (AS), cell types (CT), plus biomarkers (B) form ASCT+B tables, which are\nauthored by multiple experts across many consortia. They capture the relationship between AS (and the AS\nlocated in them), CT s found inside these AS, and the Bs (genes, proteins) used to characterize the CT , see details\nin related publications 1,22.\n● Cell Type Annotation (CTann): Azimuth23, CellTypist24,25, and popV 26 are used to assign cell types to cells from\nsc/snRNA-seq studies. Manually compiled crosswalks are used to assign ontology IDs to CT ann cell types, see\ndetails in a related publication 1.\n● Common Coordinate Framework (CCF) Ontology: The CCF ontology 8 provides the main vocabulary for\nconstructing atlases of the human body, including the HRA. Critically, the CCF provides the framework for\nconstructing atlases, but is not an atlas itself. It includes concepts and properties needed to describe the human\nbody, from organs down to cell types and biomarkers, for organizing spatial data, and for capturing donor, sample\nand dataset metadata published in the KG.\n● Content negotiation: This mechanism is used by HTTP servers to serve different versions of a resource at the\nsame Uniform Resource Identifier (URI) based on the parameters given in the HTTP request\n(https://developer.mozilla.org/en-US/docs/Web/HTTP/Content_negotiation). HTTP requests can specify headers\nwhich provide additional information for the server to act on. Using the Accept header, an agent can specify what\nformat (or a ranked list of acceptable formats) they would like the response returned in. A web browser will\ntypically request text/html, but machines or programmers may request other formats like application/json or any of\nthe RDF formats supported by the HRA KG (see subsection on Deployment in Methods > Data Processing\nPipeline).\n● Crosswalks: An ontological mapping of terms in HRA DOs to ontology terms in the ASCT+B tables 22. Crosswalks\ncan link, e.g., 2/3D Reference Objects of organs to AS and CT , and OMAPs 27 to CT and B. This definition is\nadapted from a related publication 1.\n● HRA Digital Objects (DOs): HRA DOs are the data components for generating the HRA KG. They are explained\nin detail in the Results > HRA Digital Objects section. ASCT+B tables, 3D Reference Objects, and OMAPs are\nexamples of HRA DOs that are processed to become part of the HRA KG.\n● Linked Open Data (LOD): A common data sharing pattern 28, developed for the Semantic Web\n(https://www.w3.org/2001/sw/wiki/Main_Page) that describes how to structure and share semantically rich data\nthat allows for maximum reuse and utility. T o be LOD, the data should have an open license, use URIs in the data\nto name entities whose URIs resolve (i.e., can be queried either directly via web request or via SPARQL) to\nretrieve structured data in RDF about that URI and link to other resources via URIs.\n● Linked Data Modeling Language (LinkML): LinkML (https://linkml.io/linkml/) 29 is a flexible linked data modeling\nlanguage that allows us to author schemas in YAML (https://yaml.org) which describe the structure of one’s data.\nAdditionally, it is a framework for working with and validating data in a variety of formats (JavaScript Object\nNotation [JSON], RDF , tab-separated values [TSV]), with generators for compiling LinkML schemas to other\nframeworks.\n● Organ Mapping Antibody Panels (OMAPs): T abular data structures with panels of experiment-derived and\ntested antibodies to target proteins for identifying AS, CT s, cell states, or cell membrane staining in organs.\n● Persistent Uniform Resource Locator (PURL): A PURL is a type of URL pointing to a resolution service rather\nthan to a website. This enables the resolution service to use content negotiation to determine what content is\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nneeded (e.g.,HTML for humans, structured data for machines) and to redirect or directly return the relevant data.\nPURLs are used in LOD to provide persistent, resolvable URIs for entities so that they can be referenced without\nworry of the URIs changing.\n● Resource Description Framework (RDF): RDF is a standard to represent connected data on the web. It defines\nrelationships between data objects, enabling exchange of structured information through triples consisting of a\nsubject, predicate, and object (https://www.w3.org/RDF/).\n● SPARQL: A query language for RDF graphs (https://www.w3.org/TR/sparql11-query), SPARQL can be used to\nwrite declarative code to retrieve triples that describe two entities and their relationship in an RDF graph. Triples\nhave a subject-predicate-object relationship.\n● Subject matter experts (SMEs): Individuals who possess specialized training in areas related to HRA\nconstruction, such as anatomists, surgeons, clinicians, and physicians. SMEs may have valuable knowledge\nabout individual organs or organ systems or parts thereof, such as their cellular or molecular make-up, or be\nexperts on experimental procedures.\nRelated Work\nThe goal of the HRA KG is to make the increasing amount of expertise and biomedical knowledge inside HRA\nDOs (see Box 1 and Results > HRA Digital Objects) programmatically accessible. Similarly, with the growing\ndiversity, volume, and complexity of biomedical data (including scientific literature) in recent years, KGs have\nbeen used extensively to extract, represent, and make programmatically available knowledge otherwise\nobscured in scientific papers. Recently, SemOpenAlex 30 provided more than 26 billion RDF triples describing\nscholarly knowledge. The SemOpenAlex Explorer (https://semopenalex.org) enables non-programmer access\nto the SemOpenAlex KG. The Leibniz Information Centre for Science and T echnology University Library in\nHannover (Germany) publishes the Open Research Knowledge Graph 31 (ORKG, https://orkg.org) alongside an\nAI tool called ORKG Ask (https://ask.orkg.org). The tool allows researchers without programming expertise to\nbrowse data-driven comparisons and thematic paper collections. KGs have also been used to facilitate mining\nbiological entities out of 29,000,000 PubMed abstracts 32. The creators performed author name disambiguating,\nadded funding data via the NIH ExPORTER (https://reporter.nih.gov/exporter), and collected educational and\nprofessional backgrounds of author via Open Researcher and Contributor ID (ORCID, https://orcid.org) and\nMapAffil33. Various biological areas have been captured in KGs: RNA-KG 34 federates entities related to\nribonucleic acid (RNA) from 60 public resources with the goal of aiding RNA therapy development. The authors\npreprocessed each disparate data source, built a metagraph using ontology terms for biomolecular entities and\ntheir interactions, then made the graph queryable with SPARQL endpoint and allowed the user to download\ndata products in different formats. The open-source RTX-KG2 35 allows building a KG across 70 data sources,\nincl. UMLS, the Semantic MEDLINE (https://pubmed.ncbi.nlm.nih.gov) Database (SemMedDB) 36, ChEMBL 37,\nReactome38,39, DrugBank 40, and the Small Molecule Pathway Database (SMPDB). It includes provenance and\nscientific references where applicable. The aforementioned UBKG (https://ubkg.docs.xconsortia.org) connects\nrelated concepts from biomedical ontologies and vocabularies. It combines the concept of assertions, i.e., the\nchaining of two entities and a relationship into a subject-predicate-object relationship, with the UMLS to build\nconnections between sets of assertions. Petagraph 10 (https://github.com/T aylorResearchLab/Petagraph) is built\nusing UBKG to integrate over 20 large multi-omics human and mouse genomic datasets. It leverages 180+\nontologies, and annotation resources that support a diversity of genomics data types, like STRING\n(https://string-db.org), Reactome (https://reactome.org) 38, and GENCODE (https://www.gencodegenes.org).\nThe goal of Petagraph is to harmonize omics data for rapid feature selection and hypothesis generation. It\ncontains 32 million nodes and 118 million relationships. Like UBKG, Petagraph uses Neo4j and does not allow\nqueries via SPARQL. For custom queries, the maintainers need to add an endpoint. Note also that a license is\nneeded to use UMLS (https://www.nlm.nih.gov/databases/umls.html).\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nAn essential part of building the HRA KG (and others) is to link entities in it to ontologies (see Results > HRA\nKG Construction, Deployment, and Usage) to make the data in it FAIR 16 and enable connections to other units\nof biomedical knowledge. In a study 41 of the the reuse of study metadata in ClinicalTrials.gov, values in\n302,091 trial records were tested for their adherence to expected data types and use terms from biomedical\nontologies. Among other issues, the authors found that ontologies are not the only vocabulary used for\nsearch-related fields, e.g., condition and intervention. Further, almost 50% of the conditions were not annotated\nwith Medical Subject Headings (MeSH, https://www.ncbi.nlm.nih.gov/mesh) terms. Ontologies can be used to\nprovide controlled, unified vocabulary across entities in a KG. This facilitates the aggregation of research\nresources, such as results from clinical trials from disparate sources, terms for human anatomy, and cell type\ntypologies. The Center for Expanded Data Annotation and Retrieval (CEDAR) 42,43 and its Workbench aims to\nunify how metadata for biomedical datasets are created, annotated, and shared to facilitate data sharing and\nacceleration of discovery. An important element increasingly used to support standardization efforts are\nCommon Data Elements (CDEs), which define specific questions and their permissible answers. Although\nCDEs offer a solid conceptual basis for interoperability, there is no broadly accepted format for their\nserialization or exchange.\nThe application, extension, and validation of ontologies is a separate research area. T o make annotation with\nontology-aligned terms easier for researchers and developers, validators are used for building graphical\nrepresentations of ontology views and user-provided hierarchies, e.g., for the HRA and the Human\nDevelopmental Cell Atlas (https://www.humancellatlas.org/dca) 44. The Stimulating Peripheral Activity to Relieve\nConditions (SPARC, https://sparc.science) effort constructs its own vocabulary from ontologies in relevant\ndomains, e.g., anatomy and physiology 45. This SPARC Vocabulary is applied in segmentation and annotation\ntools so that ontological terms can be applied to research data. T o address the issue that these annotations\nhappen in experimental data, where not all terms and relationships may be represented in community\nontologies, researchers can expand SPARC Vocabulary by managing terms via InterLex\n(https://scicrunch.org/resolver/RRID:SCR_016178), a lexicon for biomedical terms used by the National\nInstitute of Diabetes and Digestive and Kidney Diseases (NIDDK) Information Network (https://dknet.org),\namong others. Relatedly, the SPARC Connectivity Knowledgebase of the Autonomic Nervous System\n(SCKAN)46 provides SPARC data on the autonomic nervous system (ANS) and provides origins, terminations,\nand routing of ANS projections for improved neuromodulation devices and bioelectronic medicine for nervous\nsystem diseases. In this case, SCKAN presents distilled connectivity knowledge by experts, published\nliterature, textbooks, and SPARC scientific data. SCKAN is used to automatically generate anatomical and\nfunctional connectivity maps on the SPARC portal (https://sparc.science/apps/maps?type=ac). Like SCKAN 46,\nthe Data Distillery Knowledge Graph (DDKG, https://dd-kg-ui.cfde.cloud/), on the other hand, focuses on\nextracting essential knowledge out of disparate data sources in the Common Fund Data Ecosystem (CFDE,\nhttps://info.cfde.cloud/). It aims to facilitate better integration and reuse of CF data to accelerate discoveries in\nbiomedical research by extending Petagraph's schema to integrate over 30 large genomics datasets in\ncollaboration with NIH Common Fund efforts, such as HuBMAP , the Cellular Senescence Network (SenNet,\nhttps://data.sennetconsortium.org)47, Gabriella Miller Kids First Pediatric Research (Kids First,\nhttps://kidsfirstdrc.org), 4D Nucleome (https://4dnucleome.org) 48, the Genotype-Tissue Expression (GTEx,\nhttps://gtexportal.org/home)49 Project, and the Library of Integrated Network-Based Cellular Signatures (LINCS,\nhttps://lincsproject.org/LINCS). Like Petagraph, it is scalable for new genomics data types and cross-dataset\nanalyses and includes development of user-friendly interfaces (API, UI) for secure and efficient querying. It\ncontains 40 million nodes and 300 million relationships. A set of use cases is available at\nhttps://dd-kg-ui.cfde.cloud/use_cases.\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nResults\nHRA Digital Objects\nHRA DOs come in diverse formats, such as ASCT+B tables 22, 3D Reference Objects, and OMAPs 27. Each DO\nhas a type, name, and version. For example, asct-b/kidney/v1.5 indicates asct-b as the type, kidney as the\nname, and v1.5 as the version. These DOs were provided by SMEs (see Box 1), then reviewed and validated\nthroughout the data acquisition process.\nAs an example, ASCT+B tables are complex data structures. When constructing a table, SMEs are asked to\ncrosswalk AS, CT , and B terms to ontology terms so they can be mapped to the HRA. ASCT+B tables make it\npossible for anatomists, surgeons, and other experts to digitize knowledge about the cells and biomarkers in\nhealthy tissue; however, parsing of unprocessed ASCT+B tables is not advisable, as detailed validation and\nadditional enrichment are required before the data can be used in HRA construction, see details in the Result >\nHRA KG Construction section.\nA complete list of DO types in HRA v2.2 is provided in Table 1. DOs are available in a variety of formats on the\nLOD server at https://lod.humanatlas.io (see Methods > HRA KG Construction and Deployment > Data\nProcessing Pipeline). Figure 1 illustrates high-level relationships among the 13 DO types.\nTable 1. Different DO types used in the HRA KG, describing their purposes and the data they contain plus SOPs detailing\nthe construction of these DOs.\nDigital Object Description\nReference data\nasct-b Represents an ASCT+B table, see Box 1. Contains detailed knowledge about human body parts in\na hierarchical order, explaining the organization of AS, the CT in each AS, and the Bs that\ndistinguish each CT 22. An SOP is available 50.\n2d-ftu Provides 2D illustrations of FTU structures in an organ, with image assets and cell annotations that\nassign proper labels and identifiers based on CL for each image segment. An SOP is available 51.\nref-organ Represents a 3D Reference Object, see Box 1. Provides 3D models of human organ structures,\ncomplete with accurate size and position data, to support the creation of a comprehensive 3D model\nof the human body, with each 3D model object carefully annotated with a proper label and an\nidentifier from the Uberon and FMA ontologies. Multiple SOPs are available 52,53.\nlandmark Provides 3D model shapes representing features near organs of interest (e.g., an artery or pelvis\nbone near a kidney) to help experts accurately orient themselves when registering tissue blocks into\na 3D Reference Object.\nvascular-geometry Provides detailed geometry information on the human blood vascular system which captures key\nattributes of different vessels such as diameter and length, population, sample size, and reference\nto the source of data. Multiple SOPs are available 54–57.\nmillitome Provides data about cutting tissue samples using a millitome device. An SOP is available 58.\nctann Represents a crosswalk, see Box 1. Translates cell type labels or abbreviations from sc/snRNAseq\nanalysis tools, specifically Azimuth 23, CellTypist24,25, and popV 26 into standardized terms in CL. The\ntranslation quality is measured using standard predicates, such as exact match and narrow match to\nensure consistent data harmonization across sc/snRNAseq analyses.\nomap Represents an OMAP , see Box 1. Reduces the costs of conducting cell imaging experiments.\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nOMAPs27 contain a panel of antibodies designed to target specific proteins for identifying CT s, AS,\ncell states, or cell membrane staining within organs, based on actual experimental projects. An SOP\nis available 59.\nschema Describes the structure, i.e., the schema, of the normalized form of a single DO type, its metadata,\nor shared concepts between DOs.\nvocab Contains various reference ontologies and vocabularies that hold standard concepts and\nrelationships used to construct the DOs. These DOs are typically external biomedical ontologies like\nCL and Uberon and provide a convenient mechanism for querying reference ontologies alongside\nHRA-curated DOs.\ncollection Combines multiple DOs to create a collection of data. The HRA itself is defined in the HRA KG as a\ncurated collection of DOs in each release version.\nExperiment data\nds-graph Provides sample registration information submitted by consortium members in HuBMAP or other\nefforts, including accurate sample sizes and positions. When combined with ref-organ data, this\ninformation helps create 3D visual tissue sample placements. Additionally, the sample information is\nlinked to datasets from researchers' assay analyses that offer deeper insights into the tissue\nsamples.\ngraph Contains externally created RDF (see Box 1) graph data that are useful for HRA use cases.\nFigure 1. The 13 DO types in the HRA KG and how they relate to each other. Note that we replaced underscores in edge\nlabels with blank spaces for legibility. For class names, we added a blank space between CamelCased class names.\nEntity-relationship diagrams are provided on the companion website at\nhttps://cns-iu.github.io/hra-kg-supporting-information.\nThe high-level relationships between DO types are as follows:\n2D Illustrations (green): 2d-ftu DOs illustrate both AS (because FTUs are AS) and the CT in them, based on\nexperimental data. 2d-ftu DOs can be downloaded in their processed form or as scalable vector graphics\n(SVG), Portable Network Graphics (PNG), or Adobe Illustrator (AI) files.\n3D Spatial Reference (yellow): ref-organ DOs get anatomical context from landmark DOs, and millitome DOs\nprovide extraction sites for an entire organ that a ref-organ DO represents. ref-organ and landmark DOs can be\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\ndownloaded as GLB files, the binary format of glTF (Graphics Library Transmission Format or GL Transmission\nFormat, see https://www.khronos.org/gltf). millitome DOs can be downloaded in the JavaScript Object Notation\nfor Linked Data format (JSON-LD, https://json-ld.org).\nBiological Structure (warm pink): The asct-b DO type plays a central role for multiple other DO types, see Box\n1 and Table 1. vascular-geometry DOs provide vascular metadata for asct-b DOs. Both can be downloaded in\ntheir raw distributions as comma-separated values (CSV) files.\nExperimental Data (muted pink): ds-graph DOs describe experimental datasets mapped into a ref-organ. They\ncan be downloaded as JSON-LD files.\nExperiment Settings (bright blue): omap DOs enable detection of proteins and cell types and are thus\nconnected to AS and Bs in asct-b DOs. omap DOs can be downloaded as CSV or Microsoft Excel files (XLSX).\nOther DO types (dark blue): CT ann crosswalks (ctann) map manual and machine learning annotations for cell\ntypes from different cell type annotation tools and workflows to the ASCT+B tables. ctann and omap DOs allow\nus to map experimental datasets, represented as ds-graphs, into the HRA. ctann DOs can be downloaded as\nCSV files. vocab DOs are referenced by asct-b, omap, ctann, and ref-organ DOs to annotate AS, CT , and Bs\nwith ontology terms and can be downloaded as Web Ontology Language files (OWL,\nhttps://www.w3.org/OWL). graph DOs are ad-hoc graphs that can reference any other DO type as needed,\ndepending on their function and scope, and can thus have any download format available for the referenced\nDOs. All current graph DOs in the HRA KG are listed in Supplemental Table 1. collection DOs aggregate\nmultiple DOs and can be downloaded as YAML files. The collection type combines other DOs to allow end\nusers to create customized configurations for their particular needs. Importantly, the HRA itself is a collection\nDO (https://lod.humanatlas.io/collection/hra/latest/). All current collections in the HRA KG are listed in\nSupplemental Table 2. Finally, the schema DO type describes the structure of all 12 other DO types plus their\nmetadata and can be downloaded in a variety of formats, including YAML, PNG, and SVG.\nHRA KG Construction, Deployment, and Usage\nThe HRA KG represents major DOs of the HRA v2.2, including 36 ASCT+B tables, 23 OMAPs, 22 2D FTUs,\n71 3D Reference Objects (plus two whole body models with all organs for male/female and a crosswalk from\n3D AS to ontology terms), see https://apps.humanatlas.io/dashboard/data. In December 2024, the HRA KG\nhad 10,064,033 nodes, 171,250,177 edges, and a size of 125.84 GB. The size of the 71 3D Reference Objects\n(GLB files, https://lod.humanatlas.io/ref-organ) in HRA v2.2 is 301 MB. In addition, the data covers anatomical\nlandmarks which are used in the HRA Registration User Interface (RUI) 19 to facilitate tissue block placement in\n3D Reference Objects; these are available at https://lod.humanatlas.io/landmark. As of HRA v2.2, there are\nlandmarks for 59 out of 65 3D Reference Objects. T ogether, they are 261 MBs large.\nT o make raw as well as processed HRA DOs available in a programmatic manner as RDF graphs, the HRA\nKG sits at the center of the HRA data ecosystem and serves as the primary database for the HRA, see Figure\n2.\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nFigure 2. DOs (A), experimental datasets (B), HRA applications (C), data portals (D), and external services (E) form the\necosystem around the HRA KG. Note that the HRA KG is able to serve all existing versions of the HRA and HRA DOs.\n(A) 13 types of HRA DOs fully describe the HRA. They are represented as a HRA KG collection. 703 individual\nDOs go through a 3-step process of normalization, enrichment, and deployment via the hra-do-processor\n(https://github.com/hubmapconsortium/hra-do-processor), where they are transformed from raw data in\nmiscellaneous file formats to an RDF graph. Thus, the HRA KG integrates the knowledge from these DOs.\nT ogether, the HRA DOs form the HRA collection, which is the graph representation of the HRA. All 13 DO\ntypes are described in detail in the HRA DOs section above. The normalization, enrichment, and deployment\nare described in the HRA Construction subsection and Methods > HRA KG Construction and Deployment >\nData Processing Pipeline below.\n(B) Graph representations of external experimental datasets from various sources with mean B expression\nvalues and cell population data, resulting in the ds-graph DO type. Experimental data from various portals is\nmapped into the HRA via one or a combination of multiple methods, such as 3D tissue registration 19, CL 4\naligned Azimuth 23, CellTypist24,25, and popV 26 annotations, OMAPs 27 for spatial proteomic data, or via ontology\ncrosswalks to ASCT+B tables. These portals include the CZI CxG portal (https://cellxgene.cziscience.com ), the\nGene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo), the GTEx 49 Portal\n(https://gtexportal.org/home), the GenitoUrinary Developmental Molecular Anatomy Project (GUDMAP ,\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nhttps://www.atlas-d2k.org/gudmap)60, the Human Cell Atlas (HCA,\nhttps://www.humancellatlas.org/data-portal)61, the Human Tumor Atlas Network (HTAN,\nhttps://humantumoratlas.org/explore)62, the HuBMAP 11,12 Data Portal (https://portal.hubmapconsortium.org), the\nKPMP6,7 Kidney Tissue Atlas (https://atlas.kpmp.org), and the SenNet 47 Data Portal\n(https://data.sennetconsortium.org/search).\n(C) HRA applications and services use the HRA KG as their main data backend through the HRA API or via a\nSPARQL endpoint (https://lod.humanatlas.io/sparql). The HRA Portal (https://humanatlas.io) 1 and Dashboard\n(https://apps.humanatlas.io/dashboard) provide usage and data statistics about the HRA by querying the HRA\nKG. The HRA API provides the HRA UIs with access to Uberon and CL IDs for AS, CT , and B as well as\nspatial entities for tissue blocks and organs 8 for the RUI and Exploration User Interface (EUI) 19. For example,\nwhen using the EUI to select AS, CT , and B terms (on the left side of the UI), counts are retrieved from those\nrelationships that are curated from multiple graphs. Since all graphs are in the RDF format, it is feasible to\nquery across multiple graphs seamlessly without modifying the source graphs. HRA Cell Type Populations\n(HRApop) provide cell types and mean biomarker expressions for experimental datasets mapped to the HRA.\nThe FTU Explorer (https://apps.humanatlas.io/ftu-explorer) 63 accesses CL IDs for cells and HUGO Gene\nNomenclature Committee (HGNC, https://www.genenames.org)64 IDs for biomarkers via the HRA KG. Ad-hoc\nqueries to retrieve counts and access DOs from the HRA KG are easily possible via the SPARQL endpoint.\nHRAlit data 65 (https://github.com/x-atlas-consortia/hra-lit) connects 136 DOs from HRA v1.4 to 583,117 experts,\n7,103,180 publications, 896,680 funded projects, and 1,816 experimental datasets. The HRA Organ Gallery in\nvirtual reality (VR) 66 utilizes the HRA KG to show predicted cell types in tissue blocks in immersive, 3D space.\nThe primary API at https://apps.humanatlas.io/api has programming language-specific client libraries in\nJavaScript, TypeScript, Angular 17+, and Python 3.6+. These client libraries are published to common code\npackage managers, including NPM (https://www.npmjs.com ) and PyPi (https://pypi.org), and they wrap the API\ncalls into simple function calls to use from code, making HRA data easy to use from software development\nenvironments. A full list of client libraries is available at https://humanatlas.io/api. A set of example Python\nNotebooks are provided at https://github.com/x-atlas-consortia/hra-api/tree/main/notebooks.\n(D) The HRA KG is used in several data portals, including some from which ds-graph DOs are being\nextracted: HuBMAP , SenNet, GTEx, and KPMP . For example, the HRA-UBKG Exporter\n(https://github.com/x-atlas-consortia/hra-ubkg-exporter) is used to make HRA data available for HuBMAP Data\nPortal services (https://portal.hubmapconsortium.org), such as Uberon 3 aligned organ pages (e.g.,\nhttps://portal.hubmapconsortium.org/organ/lung), and AS search and filter functionality. Links to publicly\naccessible instances of UIs using the HRA KG on data portals are provided in Supplemental Table 3.\n(E) Several external services serve and/or use the HRA KG data: NCBO Bioportal and Ontobee host the HRA\nCCF Ontology 8 (see Box 1) at https://bioportal.bioontology.org/ontologies/CCF and\nhttps://ontobee.org/ontology/CCFO. The OLS provides a collection of HRA DOs for validation at\nhttps://www.ebi.ac.uk/ols4/ontologies/hra. CellGuide (https://cellxgene.cziscience.com/cellguide) utilizes the\nASCT+B tables to identify and present to their users canonical biomarkers for cell types. Finally, the NIH3D\nplatform by the National Institute of Allergy and Infectious Diseases (NIAID) hosts all 71 3D Reference Objects\nfor organs in the HRA v2.2 (https://3d.nih.gov/collections), plus two United files with all organs combined. 22\n2D FTU illustrations are at https://bioart.niaid.nih.gov/discover?collection=2.\nMetagraph\nHRA DO types can be aggregated into five thematic subgraphs: Spatial Reference, FTU Illustrations,\nExperiment Settings, Biological Structure, and Experimental Data. Figure 3 presents the HRA KG metagraph,\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nwhich depicts the higher-order relationships among these interconnected subgraphs. The Spatial Reference\nsubgraph (yellow) specifically is presented and explored in detail in a prior publication 8.\nFigure 3. This metagraph illustrates high-level relationships within the HRA KG.\nThe Biological Structure subgraph anchors all other components. It contains the AS, CT s, and Bs together\nwith their ontological relationships. AS and CT can have self-loops, where an AS can be part of another AS,\ncreating a partonomy, and a CT can be a subclass of another CT in a typology.\nThe Spatial Reference subgraph represents the 3D CCF used to accurately position (3D) Reference Organs,\nanatomical landmarks, and millitomes (see Table 1) within the human body. This subgraph links to the\nBiological Structure to provide the 3D anatomical context for the AS.\nSimilarly, the FTU Illustrations subgraph uses the Biological Structure subgraph to enrich its 2D FTU\nIllustrations and the FTU Illustration Nodes within it with the ontology-aligned naming for FTUs and CT .\nThe Experimental Data subgraph focuses on experimental Datasets generated from assay analyses\nperformed on Donor Tissue Blocks (Samples). These are assigned an Extraction Site with the RUI 19 based on\ntheir anatomical origins to provide a location within the CCF . Since HRA v1.2, extraction sites are placed\nrelative to Reference Organs; note that this is also a change in terminology, which we used to call\nhas_placement8. All possible alternative locations of an extraction site given its intersection(s) with one or\nmultiple 3D AS are captured in a Corridor. Systematic whole-organ registration is available as a Millitome,\nwhich defines a set of connected extraction sites placed relative to a Reference Organ. This subgraph also\naccommodates derived data computed from the assay results and extraction sites, such as (a) Cell\nSummaries, which provide cell type populations and mean gene expression values for specific cell types and\ntheir associated datasets and 3D extraction sites, and (b) Collision Summaries, which identify AS that overlap\nwith the registered tissue blocks inside a 3D extraction site.\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nFinally, the Experimental Settings subgraph catalogs Antibody Panels via OMAPs 27, capturing details of\nspecific Antibodies used to detect particular biomarkers.\nHRA KG Construction\nThe 13 HRA DO types described above come from SMEs, who contribute their knowledge of anatomy,\nantibodies, pathology, and experimental procedures via disparate data sources. Manually curated and\nexperimental datasets from diverse sources need to be mapped to the HRA and standard ontologies,\nnormalized to a standard format (e.g., unification of term labels), and enriched (e.g., linked to existing\nontologies in support of causal reasoning). We developed a software tool called hra-do-processor\n(https://github.com/hubmapconsortium/hra-do-processor) to normalize and enrich these DOs, then deploy\nthem as RDF graphs. A catalog of these graphs is available on the HRA KG LOD server at\nhttps://lod.humanatlas.io. A complete list of code pieces used for HRA KG construction, deployment, and\nusage is shown in Supplemental Table 4. The HRA KG is constructed twice a year, coinciding with the HRA\nrelease cycle 1 (see release notes at https://humanatlas.io/overview-training-outreach#release-notes).\nThe data processing pipeline uses a sequence of five steps to convert different raw datasets into the HRA KG:\nnormalization, enrichment, deployment, finalization, and serving. Implementation details for the data\nprocessing pipeline are provided in the Methods > HRA KG Construction and Deployment > Data Processing\nPipeline section.\nNormalization: This initial step ensures that all incoming data is transformed into a consistent format that\naligns with the predefined schema. For example, unnormalized, raw ASCT+B tables are available via the\nASCT+B Reporter (https://apps.humanatlas.io/asctb-reporter). A normalized example is provided in the form of\nan ASCT+B record in our format in Supplemental Figure 1.\nEnrichment: This step converts the normalized data to RDF graph format and enriches it with related\nrelationships, entities, and metadata from external resources, either from ontologies like Uberon or CL or via\nAPIs like HGNC.\nDeployment: Once the data is enriched, it is prepared for use in downstream applications or for access by end\nusers. This stage involves organizing the data into its final distribution formats and setting up the correct file\nsystem directory structure.\nFinalization: This step involves generating the necessary metadata and landing pages for web publication,\ne.g., https://lod.humanatlas.io/asct-b/eye/latest leads to the most recently published ASCT+B table for the eye.\nIn addition, this stage includes building the SPARQL database that will be uploaded to the web for users to\naccess at https://lod.humanatlas.io/sparql.\nServing: Data processed in the previous steps, including raw DO data, processed data products, HTML\npages, metadata, and the SPARQL database, are made available online at https://lod.humanatlas.io. The data\nis regularly updated and synchronized, either during scheduled releases or when updates occur, to ensure that\nthe most current version is always available.\nHRA KG Deployment\nThe HRA KG LOD server at https://lod.humanatlas.io provides metadata for processed HRA DOs as Data\nCatalog Vocabulary (DCAT) datasets (https://www.w3.org/TR/vocab-dcat-3) in five different graph formats,\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nincluding JSON-LD. Every DO has a PURL (see Box 1), such as\nhttps://purl.humanatlas.io/asct-b/lymph-node/v1.4. Raw data files are hosted on a global content delivery\nnetwork (CDN), e.g.,\nhttps://cdn.humanatlas.io/digital-objects/ref-organ/knee-female-right/v1.3/assets/3d-vh-f-knee-r.glb.\nThe next section details how HRA DOs, e.g., ASCT+B tables, can be accessed and used in the JSON-LD\nformat via Python. We deliver the graph data via Amazon Web Services (AWS, https://aws.amazon.com).\nHRA KG Usage\nUsers access the HRA KG via UIs, APIs, and data products on https://lod.humanatlas.io to answer biomedical\nquestions. A list of all HRA applications that use the HRA KG is provided in Supplemental Table 3. A list of\npublications and aliases used throughout HRA applications per HRA DO is provided in Supplemental Table 5.\nThe HRA KG makes it possible to access harmonized, high-quality reference and experimental data in\nstandard data formats. Three widely used queries are detailed here: (1) retrieve AS-CT-B records from the\nASCT+B tables, (2) compute mean biomarker expression values for cell types across datasets in HRApop, and\n(3) query the HRA KG to achieve two types of predictions: predict cell type populations for extraction sites, and\npredict 3D registration corridors for a given cell type population, see HRA user stories 1-2 in a related\npublication1.\nT o simplify HRA KG usage, we use the https://grlc.io service to make a set of canned SPARQL queries\navailable to execute as simple web requests. Internally, the service creates an OpenAPI specification\n(https://swagger.io/specification) that advertises the available queries. We provide a user-friendly interface to\nthese queries at https://apps.humanatlas.io/api/grlc. This deployment was inspired by the PubMed MeSH\nSPARQL Explorer at https://id.nlm.nih.gov/mesh/query.\nGet records from ASCT+B tables\nThe HRA KG API makes it easy to retrieve AS-CT-B records, properties, or counts for one or multiple organs.\nExemplary queries and links to resulting data are provided here. Documentation of the OpenAPI specification\nis at https://apps.humanatlas.io/api/grlc/hra.html. Annotated screenshots of the query interface with instructions\non how to run the queries or download resulting data can be found at the companion website at\nhttps://cns-iu.github.io/hra-kg-supporting-information/#how-to-run-queries-via-our-openapi-spec.\nGet ASCT+B counts for all tables: Retrieve the number of unique AS, CT , and B terms across all ASCT+B\ntables via this query for the latest HRA release:\n● https://apps.humanatlas.io/api/grlc/hra.html#get-/as-ct-b-counts\nReturns: A table with the number of unique AS, CT , B in the latest version of the HRA.\nRetrieve the number of unique AS, CT , and B terms across all ASCT+B tables for all HRA releases:\n● https://apps.humanatlas.io/api/grlc/hra.html#get-/as-ct-b-counts-all-versions\nReturns: A table with the number of unique AS, CT , B across all versions of the HRA (not just the latest one).\nNote that the only difference between the two queries is the added FROM HRA: statement, which limits the\nSPARQL search pattern of the first query to the latest HRA collection at\nhttps://purl.humanatlas.io/collection/hra.\nGet ASCT+B records for one organ : Given the PURL of an asct-b DO, retrieve ASCT-B records from the\nASCT+B table:\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\n● https://apps.humanatlas.io/api/grlc/hra.html#get-/asctb-in-table.\nReturns: A table with one row per AS-CT-B record in the specified table, in the format: AS, CT , B labels, and\nthe AS, CT , B ontology ID (if crosswalked, otherwise it returns a temporary ID).\nExemplary Python code that uses this endpoint is provided on the companion website at\nhttps://cns-iu.github.io/hra-kg-supporting-information#basic-usage. The user can choose from common data\nformats (CSV or JSON) for the response via the Accept header.\nGet ASCT+B records for all organs: Retrieve individual records for all ASCT+B tables (rather than just\ncounts or records for one organ):\n● https://github.com/x-atlas-consortia/hra-pop/blob/main/queries/hra/asctb-records.rq\nReturns: A table with one row per AS-CT-B record for all ASCT+B tables, in the format: AS, CT , B labels, and\nthe AS, CT , B ontology ID (if crosswalked, otherwise it returns a temporary ID). Because the result of this query\nis large (1,048,576 rows and a total of 325 MB) and takes longer to run, it has not been deployed via\nhttps://grlc.io. Rather, the query response has been preprocessed and is available for download as a zipped\nCSV file on GitHub:\nhttps://github.com/x-atlas-consortia/hra-pop/blob/main/output-data/v0.11.1/reports/hra/asctb-records.csv.zip.\nAn exemplary Jupyter Notebook on how to run a SPARQL query against the HRA KG is at\nhttps://cns-iu.github.io/hra-kg-supporting-information/#notebook-to-query-the-hra-knowledge-graph-kg.\nRetrieve Mean Biomarker Expression Values for Cell Type(s)\nThe hra-pop graph (https://lod.humanatlas.io/graph/hra-pop/latest) in the HRA KG contains mean B expression\nvalues for CT s inside datasets. T o compute mean B expressions, scanpy 67, numpy 68, and anndata 69 are used.\nConcretely, scanpy’s to rank_gene_groups() method assigns mean biomarker expressions. We normalize gene\nnames with a lookup table from Ensembl Release 111 64 (https://www.ensembl.org/index.html) to HGNC\nv2023-09-1870 (https://www.genenames.org). Both the code to compute mean B expressions and the look-up\ntable from Ensembl to HGNC are linked in Supplemental Table 1 under the entry for the hra-pop graph.\nRetrieve mean biomarker expression values for a given cell type across organs: Retrieve all\nexperimental datasets from HRApop that contain a canonical CT plus biomarker expression values with this\nquery:\n● https://apps.humanatlas.io/api/grlc/hra-pop.html#get-/datasets-with-ct\nReturns: A table with atlas datasets that have the given CT . There is one CT-BM expression per row. The\nresponse includes the dataset source (which portal it was downloaded from), the dataset ID (must be an\nInternationalized Resource Identifier [IRI, https://www.w3.org/International/O-URL-and-ident.html]), the organ,\nthe donor sex, the tool that assigned the CT , the CT ontology ID (same for every row as provided by user), a\nhuman-readable CT label, the number of cells of the given type in the dataset, a B ontology ID, and finally, the\nmean expression value for all Bs that characterize this CT computed via CT ann tools (see Box 1) in HRApop 1 .\nIf multiple CT ann tools were used for the same CT and dataset, multiple rows are provided to show the\ndifferent mean Bs.\nPredict Cell Type Populations and Locations\nSeven user stories 1 have been identified for the HRA based on interviews with more than 40 atlas architects\nworking on human atlases, see Supplemental Table 6. While all user stories are directly supported by HRA\nUIs and the HRA API for users with little to no programming experience, the HRA KG can also be queried\ndirectly in support of these user stories. Supplemental Table 6 also lists sample queries that help experts\nretrieve knowledge from DOs, identify processed data of interest, and run analyses to answer biomedical\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nquestions. Here, we detail the queries that support the US#1 (Predict cell type populations) and #2 (Predict\nspatial origin of tissue samples).\nPredict cell type populations (US#1): The HRA KG is used access the HRApop graph\n(https://lod.humanatlas.io/graph/hra-pop/latest) with cell type populations for AS, datasets, and extraction sites\nto improve the accuracy of annotations for sc-transcriptomics and sc-proteomics datasets. A demonstration\napplication is available at https://apps.humanatlas.io/us1.\nThe user provides an extraction site using the RUI. The extraction site is then posted to the HRA API\n(https://apps.humanatlas.io/api/#post-/hra-pop/rui-location-cell-summary), which returns a predicted cell\nsummary given the 3D collisions of the extraction site with AS inside the 3D Reference Object for which the\nextraction site was defined. Under the hood, the API queries a graph inside the HRA KG\n(https://cdn.humanatlas.io/digital-objects/graph/hra-pop/v0.11.1/assets/atlas-as-cell-summaries.jsonld) to\nretrieve a predicted cell summary for the extraction site using the SPARQL query at\nhttps://github.com/x-atlas-consortia/hra-api/blob/main/src/library/hra-pop/queries/as-weighted-cell-summaries.r\nq.\nPredict spatial origin of tissue samples (US#2): For the inverse case, the cell type populations from the\nHRApop graph mentioned above can be used to predict the 3D location of datasets with unknown spatial\norigin. A demonstration application is available at https://apps.humanatlas.io/us2.\nThe user provides a cell summary for a given dataset whose origin is uncertain or unknown (beyond basic\nmetadata such as the organ and the tool used to assign different CT). The application features an optional\ndropdown menu to select the organ and tool. Supported organs and tools for the dropdown menu are available\nvia the HRA API endpoints at https://apps.humanatlas.io/api/hra-pop/supported-organs and\nhttps://apps.humanatlas.io/api/#get-/hra-pop/supported-tools, respectively. These endpoints run SPARQL\nqueries under the hood:\nhttps://github.com/x-atlas-consortia/hra-api/blob/main/src/library/hra-pop/queries/supported-reference-organs.r\nq (to get the reference organs supported by HRApop) and\nhttps://github.com/x-atlas-consortia/hra-api/blob/main/src/library/hra-pop/queries/supported-tools.rq (to get the\ntools).\nOnce the user provides the cell type population of the dataset of unknown spatial origin, the application posts\nthis population to an endpoint in the HRA API (https://apps.humanatlas.io/api/hra-pop/cell-summary-report),\nwhich takes a minimum of a CT ID and a column for the percentage of that CT in the dataset.\nUnder the hood, the HRA API runs the SPARQL query at\nhttps://github.com/x-atlas-consortia/hra-api/blob/main/src/library/hra-pop/queries/select-cell-summaries.rq,\nwhich returns a listing of the most similar AS, datasets, and extraction sites (by cosine similarity).\nDiscussion\nThis paper detailed the construction, deployment, and usage of the HRA KG to make HRA DOs available as\nLOD (see Box 1). The HRA provides a CCF (see Box 1) with standard terminologies and data structures for\ndescribing specimens, biological structures, and spatial positions of experimental datasets and ontology-linked\nreference AS, cell types, and biomarkers. It makes it possible to map and integrate anatomical, cellular, and\nmolecular data of the human body in 3D. It also enables researchers and clinicians to consistently annotate,\ncompare, and understand biological structures and analysis results across tissues. The HRA KG offers a\nstructured, interconnected data representation of the HRA, incorporating dozens of highly curated reference\ndatasets and supporting different HRA UIs.\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nThe HRA KG makes it possible to access HRA data efficiently and to ask biological questions via\nprogrammatic queries. Researchers can leverage the KG to enrich their assay data for deeper insights, while\nclinicians can use it to explore biological questions, such as identifying the types and populations of epithelial\ncells in the human eye. The HRA KG is composed of multiple named graphs, each focusing on a specific part\nof the atlas, such as biological structures or spatial references. The HRA collection is a collection of DOI’d DOs\n(ASCT+B T ables, 3D Reference Objects, OMAPs, etc.) that make up the core of the HRA at each release.\nWhen processed, it compiles to a (large) RDF graph and is hosted by the HRA KG at\nhttps://purl.humanatlas.io/collection/hra.\nThe HRA uses a KG (as opposed to a relational database) to ensure (1) Flexibility. The schema of the HRA\nKG can be extended as needed when new organs or HRA DO data types become available (as opposed to a\nrigid schema that would need to be chosen for a relational database). (2) Easy extension. The HRA KG can\nbe easily extended with DOs for new organs. Existing DOs can easily be updated. In a relational database, one\nwould need a set of new tables. Many HRA DOs, such as 3D Reference Objects, are non-tabular and highly\nnested, which is highly challenging to model in relational databases. (3) Support for disparate data. HRA\nDOs take on many forms. For example, ASCT+B tables capture AS, CT s in those structures, and the Bs that\ncharacterize them; they are linked to OMAP and Antibody Validation Report (AVR) 27 tables (tabular data), 2D\nimages and 3D models (graphic assets), as well as dataset graphs such as HRAlit 65 and HRApop\n(highly-nested). A relational database would make it necessary to choose a schema for each of these DO\ntypes. A KG enables integration of many different DO types. (4) Answering biological questions across\nHRA DO types. The KG structure makes it easier to programmatically answer questions across multiple DOs\nfor one entire organ via graph queries (e.g., the ASCT+B table for the kidney and the 3D Reference Object for\nthe female, left kidney). In a relational database, this would necessitate a set of new tables that would need to\nbe carefully created with foreign keys and relationships to support the kind of dynamic graph-based queries\nreadily available in SPARQL. (5) Deployment as 5-Star LOD. RDF graphs can be imported into triple stores in\ntheir native format and easily be queried together with connected biomedical ontologies (genes, proteins, cells,\nanatomy), which are also published as RDF . Existing HRA KG queries bridge CL 4, Provisional Cell Ontology\n(PCL)71, Uberon 3, HGNC 64, and HRA nodes, properties, and relationships stored or imported from their\nrespective graphs.\nLimitations & Future Work\nThe current HRA KG has a number of known limitations that will be addressed in future HRA releases:\nAutomation: While many parts of the HRA KG construction process are automated, collecting and providing\nDOs in their original form (CSVs, GLBs, SVG, etc.) is still a manual process involving human labor. In future\nreleases, we aim to employ machine learning algorithms 17,72–80 to speed up data segmentation and annotation,\nusing human expertise to review (not hand-compile) DOs.\nBuild time: At present, building the HRA KG from unprocessed DOs using code in\nhttps://github.com/hubmapconsortium/hra-kg takes about 13 hours on a local Linux server with 256GB RAM\nand 20 cores. As new DO types are added and HRApop and HRAlit grow, we will need to optimize the\nconstruction process by optimizing KG structures, using parallelization, and optimizing normalization and\nenrichment code (some libraries are particularly slow for certain DO types). Preliminary results from one\nexperiment showed a nearly one-third reduction in execution time, demonstrating the potential of parallelization\nin optimizing KG construction.\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nReduce ASCTB-TEMP terms: As of HRA v2.2, 221 CT s across 36 ASCT+B tables do not yet exist in CL or\nPCL; instead they have an ASCTB-TEMP expert provided label. GitHub issues have been submitted for all and\nthe EBI team is adding these terms to existing ontologies. As of December 2024, a total of 141 CT were added\nto CL 4, and 461 CT were added to PCL 71 by ASCT+B table editors. Additionally, 126 AS terms were added to\nUberon3.\nData modeling: T o be most useful to the HRA KG, each new DO type must have a LinkML schema (see Box\n1), normalization code, and enrichment code to transform the raw data into useful, queryable information. As\nnew use cases are identified, the HRA structure and canned queries will be revised and expanded. Currently,\nHRAlit is being served via a relational database. Knowledge modeling is underway to create a HRAlit KG and\nto properly connect it to the HRA KG, which will allow users to query peer-reviewed literature and funded\nawards for entities in the HRA KG.\nEase of use: Retrieving data from KGs requires experience writing SPARQL queries, which few clinicians and\nbiomedical researchers possess. The HRA KG comes with canned queries at\nhttps://apps.humanatlas.io/api/grlc/ as well as Jupyter Notebooks (see companion website at\nhttps://cns-iu.github.io/hra-kg-supporting-information). Going forward, we will create an HRA Developer Portal\nto help train and provide resources to users learning how to use the HRA KG. Additionally, since KGs offer\naccess to structured data, we are exploring the possibilities of utilizing modern large language models (LLMs)\nto allow users to ask questions in prose. An LLM, enhanced by retrieval-augmented generation (RAG), could\nbe used to build a chatbot for easy natural language queries that are informed by the knowledge in the HRA\nKG. With the research space around LLMs changing and the possibilities surrounding them expanding, we are\nmonitoring this space for use cases and opportunities. Further, we will continue to grow the HRA KG by\nfederating datasets, assays, samples, and donors from an ever increasing number of portals and consortia.\nFinally, we are working on an HRA KG Explorer UI, which will allow users to browse the KG via the web to\nquickly identify, select, and download HRA DOs of interest in all available graph formats. This will enable easy\naccess to the HRA KG to users without experience writing code, making API requests via https://grlc.io, or\nrunning SPARQL queries.\nMethods\nAll data and code needed to construct, deploy, and use the HRA KG are available at\nhttps://github.com/hubmapconsortium/hra-kg. A full list of all data and code is available in Supplemental Table\n4.\nHRA KG Construction and Deployment\nThe hra-do-processor (https://github.com/hubmapconsortium/hra-do-processor) is built around three main\ncomponents: the schema, the data processing pipeline, and the web infrastructure. The following sections\ndetail each of these components.\nSchema\nWell-defined data schemas are crucial for ensuring data consistency, interoperability, and validation in data\nmanagement and analysis. SOPs that define the schema make it easier for others to contribute to and use the\nHRA data. LinkML 29is a flexible and user-friendly schema language designed to create effective data models\nand validation tools to ensure input data adheres to a defined schema. Entity relationship diagrams of core\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nHRA schemas explain the relationships between different HRA DOs, see examples on the companion website\nat https://cns-iu.github.io/hra-kg-supporting-information/#mermaid-diagrams.\nData Processing Pipeline\nThe HRA KG data processing pipeline has five key steps:\nNormalization: The hra-do-processor loads and parses the disparate source data produced by diverse SMEs\nand transforms it into a standardized linked-data representation. For example, in the case of asct-b DOs, the\nsource data comes from Google Sheets exported as CSV files; during normalization, the hra-do-processor\nreads this tabular structure and converts it into the tree structure shown in Supplemental Figure 1. An\nexemplary YAML file is provided at\nhttps://github.com/cns-iu/hra-kg-supporting-information/blob/main/docs/intermediary_format.yaml. We chose\nYAML as the standard file format for the normalized data due to its simplicity, readability, interoperability with\nJSON, and high level of support in LinkML. By converting the LinkML schema into a JSON-Schema file, we\ncan easily validate the translated YAML data to be sure it adheres to the defined schema and our automated\ningestion code is correctly implemented.\nEnrichment: After the source data is translated into YAML and validated, enrichment begins by converting the\nvalidated data into OWL-based statements. The LinkML framework offers tools that facilitate the transformation\nof tree-structured data into OWL constructs, including class and property declarations, as well as class\ninstances (or individuals) (see Supplemental Figure 2). We chose OWL as the data representation for the\nenrichment step due to its robust capabilities for rich data expression, its ability to embed semantic meaning,\nand its seamless integration with LinkML. LinkML provides direct support for OWL by allowing schema\nelements to include the type of OWL constructs, making it easy to map the data into a semantically rich\nontology structure.\nThe enrichment process continues by integrating additional information from reference ontologies such as\nUberon, FMA, and CL, as well as authoritative databases like Research Resource Identifier (RRID,\nhttps://www.rrids.org) and the Antibody Registry API (https://www.antibodyregistry.org) to retrieve metadata\n(label, description) about antibodies, for which there is only an RRID in the raw HRA DO for OMAPs.\nThe goal is to enhance the initial data gathered from SMEs with more detailed, authoritative information. In\nasct-b DOs, many data points reference Uberon and CL terms. We enrich these terms by retrieving\nsupplementary information from the corresponding ontologies, including class hierarchies, labels, definitions,\nsynonyms, database references, and visual depictions. For example, we identified the standard label for a\nCL:0002306 as \"epithelial cell of proximal tubule,\" which is categorized under the broader class\n\"meso-epithelial cell.\" These details, which were absent from the original dataset, add valuable context. The\nend result is a semantically enriched dataset that not only preserves the original data but also extends it with\nadditional context, relationships, and meaning.\nDeployment: HRA data is used by many different tools and user communities. HRA UIs like the EUI and RUI 19\nuse JSON-LD, which is best when using the data directly and imperatively (i.e., in a programming language\nusing for loops), e.g., Python and JavaScript have native support for handling JSON and has semantics built\nin.\nThe Blazegraph (https://blazegraph.com ) SPARQL server uses T erse RDF Triple Language (Turtle,\nhttps://www.w3.org/TR/rdf12-turtle); additionally, the Turtle format helps developers write SPARQL queries to\nthe HRA KG by making its triple structure explicit and showing possible subjects, predicates, and objects.\nOlder semantic web tools use RDF/Extensible Markup Language (XML,\nhttps://www.w3.org/TR/rdf-syntax-grammar), N-Triples (https://www.w3.org/TR/n-triples/), and N-Quads\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\n(https://www.w3.org/TR/n-quads). Additionally, Robot (http://robot.obolibrary.org/convert.html), Apache Jena\n(https://jena.apache.org/documentation/io/), and RDF I/O technology (RIOT) use XML for reifying graphs. HRA\nKG data is preprocessed in those formats for it to be readily usable by others. Publishing all these formats\nstreamlines the content negotiation process later (see Box 1) when different applications access the published\nHRA KG on the LOD server at https://lod.humanatlas.io, which can then immediately deliver the HRA data in\nthe correct format. The tool also prepares the metadata that accompanies the graph data.\nThe deployment process includes setting up the file system by creating directories to organize and store output\nfiles for deployment. The hra-do-processor then copies files and data assets into their designated folders as a\npreparation for the next step. Data and metadata for each DO are pre-converted into the formats previously\nmentioned.\nFinalization: In the deployment step, data and metadata for each DO are converted and exported. Finalization\nderives additional files across all DOs, including metadata catalogs, latest versions of each DO, HTML landing\npages to navigate the HRA KG and view DOs, and an indexed and optimized database file for the Blazegraph\nSPARQL server accessible at https://lod.humanatlas.io/sparql. The database contains the latest version of\nevery DO, every version of the HRA collection, and a metadata catalog that contains metadata for every\nversion of every DO in the HRA KG.\nServing: T o make the processed data widely accessible, AWS is used to serve the HRA KG as linked open\ndata, employing three of its core services: S3, ECS and CloudFront for data storage, computation, and content\ndelivery, respectively. Implementation details are provided in the next section.\nWeb infrastructure\nAmazon S3 (Simple Storage Service) is a highly scalable data storage service to store and retrieve data. The\nHRA KG uses S3 to store the content from the local deployment directories, including the Blazegraph database\nfile. By syncing these local directories with an S3 storage, the data is securely stored and readily available for\ncontent delivery.\nAmazon ECS (Elastic Container Service) is a fully managed container service to run applications in Docker\ncontainers for a highly scalable and reliable environment for our computation needs (https://www.docker.com ).\nFor HRA, a Blazegraph instance is run within an ECS container. The ECS container periodically checks the S3\nstorage for an updated Blazegraph database file. When a newly built Blazegraph file is detected, ECS will\nseamlessly update the Blazegraph server to ensure that the latest data is available for querying.\nAmazon CloudFront is a global Content Delivery Network (CDN) designed to accelerate the distribution of\ncontent by caching copies at multiple serving locations around the world. The HRA KG uses CloudFront to\ncreate a URL fabric that caches and serves content from S3 storage to ensure fast and reliable access for\nusers, regardless of their geographical location. The content stored in S3 is made publicly available through\nURLs like https://cdn.humanatlas.io/digital-objects and https://lod.humanatlas.io. Additionally, CloudFront\nprovides advanced content negotiation features through Amazon CloudFront functions, enabling the dynamic\nhandling of URLs starting with https://purl.humanatlas.io and https://lod.humanatlas.io. Content negotiation\nallows the web infrastructure to serve data in different formats based on user needs, whether a user requires\nRDF , XML, JSON, or another format. The PURL returns HRA DO data based on the Accept header of the\nrequest: human users get redirected to the LOD Server, machines to JSON or RDF versions. The LOD Server\nreturns HRA DO metadata based on the Accept header: Human users get HTML, machines get JSON or RDF .\nMoreover, CloudFront also acts as an intermediary for the SPARQL endpoint hosted by Blazegraph within ECS\nby making it accessible at https://lod.humanatlas.io/sparql.\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nUpdates since the CCF .OWL paper (HRA v1.2 / CCF .OWL v2.0.1)\nThe specimen, biological structure, and spatial ontologies in support of a HRA (v1.2) using CCF v2.0.1 were\nintroduced in a prior publication 8. Starting with HRA v2.0, published in December 2023, the CCF Ontology\n(v3.0) is separated from the HRA collection. CCF v3.0 is a DO of type (vocab) and the HRA collection is a DO\nof type (collection). Before then, it was a graph, essentially the HRA collection plus the CCF Ontology. Now the\nHRA collection references the CCF and is compiled from a collection of curated HRA DOs and is hosted by the\nHRA KG at (https://purl.humanatlas.io/collection/hra/v2.2). This change was necessary to establish a boundary\nbetween the framework for creating atlases, the CCF , and a specific atlas, the HRA.\nOther Ontologies\nThe HRA KG includes other reference ontologies at https://lod.humanatlas.io/vocab (e.g., Uberon and CL) so\nthey can be queried together in an efficient manner. Table 2 lists all ontologies that are included in the HRA KG\ntogether with their version numbers.\nTable 2. Ontologies used in the HRA KG as of HRA v2.2.\nName Description Version\nNumber\nURL Main Website\nCCF Common Coordinate\nFramework Ontology 8\n3.0 https://purl.humanatlas.io/voc\nab/ccf\nhttps://humanatlas.io/ccf-ontolog\ny\nCL Cell Ontology 4 2024-09-26 https://purl.humanatlas.io/voc\nab/cl\nhttps://obophenotype.github.io/c\nell-ontology/\nFMA Foundational Model of\nAnatomy20,21\n5.0.0 https://purl.humanatlas.io/voc\nab/fma\nhttp://si.washington.edu/projects\n/fma\nHGNC HUGO Gene\nNomenclature\nCommittee64\n2024-03-04 https://purl.humanatlas.io/voc\nab/hgnc\nhttps://www.genenames.org/\nHRAVS HuBMAP Research\nAttributes Value Set\n2.5.3 https://purl.humanatlas.io/voc\nab/hravs\nhttps://bioportal.bioontology.org/\nontologies/HRAVS\nLMHA Cell Ontology for Human\nLung Maturation\n(LungMAP Human\nAnatomy)81\n1.4 https://purl.humanatlas.io/voc\nab/lmha\nhttps://bioportal.bioontology.org/\nontologies/LUNGMAP_H_CELL\nPCL Provisional Cell\nOntology71,82\n2024-07-11 https://purl.humanatlas.io/voc\nab/pcl\nhttps://github.com/obophenotyp\ne/provisional_cell_ontology\nRO OBO Relation Ontology\n(https://doi.org/10.5281/z\nenodo.32899)\n2024-04-24 https://purl.humanatlas.io/voc\nab/ro\nhttps://github.com/oborel/obo-rel\nations\nUberon Uberon Multi-species\nAnatomy Ontology 3\n2024-11-25 https://purl.humanatlas.io/voc\nab/uberon\nhttps://obophenotype.github.io/u\nberon/\nVCCF Vasculature Common\nCoordinate\nFramework22,83\n2024-02-23 https://purl.humanatlas.io/voc\nab/vccf\nhttps://github.com/hubmapconso\nrtium/hra-vccf\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nData Availability\nThe primary server for the HRA KG v2.2 is at https://lod.humanatlas.io. The SPARQL endpoint to query the\nHRA KG is at https://lod.humanatlas.io/sparql. The HRA API (https://apps.humanatlas.io/api) supports\nprogrammatic access to the HRA KG and exemplary queries are available via the companion website at\nhttps://cns-iu.github.io/hra-kg-supporting-information/.\nBioPortal hosts (1) the HRA at https://bioportal.bioontology.org/ontologies/HRA (mirror of\nhttps://purl.humanatlas.io/collection/hra/v2.2) and (2) the CCF Ontology at\nhttps://bioportal.bioontology.org/ontologies/CCF (mirror of https://lod.humanatlas.io/vocab/ccf).\nOLS hosts the latest versions of both the HRA and CCF at https://www.ebi.ac.uk/ols4/ontologies/hra and\nhttps://www.ebi.ac.uk/ols4/ontologies/ccf, respectively. OLS provides both a web-based GUI for users and\nprogrammatic access via the OLS REST API (https://www.ebi.ac.uk/ols4/help), enabling the HRA and CCF to\nbe accessed using the same standard interface as other ontologies.\nThe NIH3D platform by NIAID hosts all 71 3D Reference Objects for organs in the HRA v2.2 alongside two\nUnited files with all organs combined (https://3d.nih.gov/collections). 22 2D FTU illustrations are at\nhttps://bioart.niaid.nih.gov/discover?collection=2.\nWeekly run term and relationship validation reports of ASCT+B T ables are available at\nhttps://github.com/hubmapconsortium/ccf-validation-tools/tree/master/reports.\nAll data and SOPs are released under Creative Commons Attribution 4.0 International (CC BY 4.0).\nCode Availability\nAll the code used to construct and deploy the HRA KG v2.2 is available on GitHub at\nhttps://github.com/hubmapconsortium/hra-kg and URLs are provided in Supplemental Table 4.\nDocumentation, including an additional overview of HRA KG construction code is provided in the supporting\ninformation repository for this paper at\nhttps://github.com/cns-iu/hra-kg-supporting-information/?tab=readme-ov-file#github-repositories-used-for-hra-k\ng-construction. We also created documentation with annotated screenshots to show how to run pre-made\nSPARQL queries against the HRA KG via https://grlc.io, see\nhttps://cns-iu.github.io/hra-kg-supporting-information/#how-to-run-queries-via-our-openapi-spec.\nAll code was released under the MIT License.\nAcknowledgments\nThanks go to Nancy Ruschman, James McLaughlin, and Yongqun (Oliver) He for their expert comments and\nsuggestions on an earlier version of this paper. Libby Maier supported the design of figures.\nThe HRA is under active development by HuBMAP , SenNet, KPMP , GUDMAP , and the National Institute of\nDiabetes and Digestive and Kidney Diseases (NIDDK) with expert input by the HRA Editorial Board and in\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nclose collaboration with experts from more than 15 other consortia. K.B. is a co-director of and is funded by the\nCIFAR MacMillan Multiscale Human program.\nThis research has been supported by the NIH Common Fund through the Office of Strategic\nCoordination/Office of the NIH Director under awards:\n● OT2OD033756 and OT2OD026671 (A.B., B.W.H., J.H., E.M.Q., M.M., K.B.);\n● OT2OD026675 and OT2OD033759 (A.B.)\n● OT2OD030545 (A.B., B.W.H., K.B.)\nFurther, this work was supported by:\n● the SenNet Consortium Organization and Data Coordinating Center (CODCC) under award number\nU24CA268108-01 (A.B., B.W.H., E.M.Q., K.B.);\n● by the NIDDK under award U24DK135157 (B.W.H.);\n● by the KPMP grant U2CDK114886 (A.B., B.W.H.);\n● and the NIH National Institute of Allergy and Infectious Diseases (NIAID), Department of Health and\nHuman Services under BCBB Support Services Contract HHSN316201300006W/HHSN27200002.\nThis research was supported in part by the Intramural Research Program of the U.S. National Institutes of\nHealth. The funders had no role in study design, data collection and analysis, decision to publish, or\npreparation of the manuscript. The content is solely the responsibility of the authors and does not necessarily\nrepresent the official views of the National Institutes of Health.\nAuthor Contributions\nA.B. led the writing of the paper, built the companion website at\nhttps://cns-iu.github.io/hra-kg-supporting-information, and leads the HRApop effort. He shares corresponding\nauthorship with B.W.H., K.B. B.W.H and J.H. built the hra-do-processor and engineered the CCF Ontology and\nHRA Data data structures. B.W.H. leads the development of the HRA user interfaces and the generation and\npublication of the HRA KG. A.B. and B.W.H. compiled a notebook to show basic usage of the HRA KG on the\ncompanion website. E.M.Q. compiled cell type annotation crosswalks used in and published with the HRA KG.\nM.M. oversees HRA ontology engineering work. K.B. leads the HRA effort and specified HRA KG usage to\nhelp focus HRA KG development and documentation. A.B., B.W.H., J.H., E.M.Q., and K.B. wrote the paper. All\nother authors reviewed and commented on the paper.\nCompeting Interests\nThe authors declare no competing interests.\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nReferences\n1. Börner, K. et al. Human BioMolecular Atlas Program (HuBMAP): 3D Human Reference Atlas Construction\nand Usage. bioRxiv 2024.03.27.587041 (2024) doi:10.1101/2024.03.27.587041.\n2. Yu, C. & Li. Building commonsense knowledge graphs to aid product recommendation. Amazon Science\nhttps://www.amazon.science/blog/building-commonsense-knowledge-graphs-to-aid-product-recommendati\non (2024).\n3. Mungall, C. J., T orniai, C., Gkoutos, G. V., Lewis, S. E. & Haendel, M. A. Uberon, an integrative\nmulti-species anatomy ontology. Genome Biol 13, R5 (2012).\n4. Diehl, A. D. et al. The Cell Ontology 2016: enhanced content, modularization, and ontology interoperability.\nJ. Biomed. Semant. 7, 44 (2016).\n5. He, Y . O. et al. Ontology-based modeling, integration, and analysis of heterogeneous clinical, pathological,\nand molecular kidney data for precision medicine. 2024.04.01.587658 Preprint at\nhttps://doi.org/10.1101/2024.04.01.587658 (2024).\n6. Himmelstein, D. S. et al. Systematic integration of biomedical knowledge prioritizes drugs for repurposing.\neLife 6, e26726 (2017).\n7. El-Achkar, T . M. et al. A multimodal and integrated approach to interrogate human kidney biopsies with\nrigor and reproducibility: guidelines from the Kidney Precision Medicine Project. Physiol. Genomics 53,\n1–11 (2021).\n8. Herr, B. W. et al. Specimen, biological structure, and spatial ontologies in support of a Human Reference\nAtlas. Sci. Data 10, 171 (2023).\n9. Bodenreider, O. The Unified Medical Language System (UMLS): integrating biomedical terminology.\nNucleic Acids Res. 32, D267 (2004).\n10. Benjamin J. Stear et al. Petagraph: A large-scale unifying knowledge graph framework for integrating\nbiomolecular and biomedical data. bioRxiv 2023.02.11.528088 (2023) doi:10.1101/2023.02.11.528088.\n11. Jain, S. et al. Advances and prospects for the Human BioMolecular Atlas Program (HuBMAP). Nat. Cell\nBiol. 25, 1089–1100 (2023).\n12. Snyder, M. P . et al. The human body at cellular resolution: the NIH Human Biomolecular Atlas Program.\nNature 574, 187–192 (2019).\n13. Morris, J. H. et al. The scalable precision medicine open knowledge engine (SPOKE): a massive\nknowledge graph of biomedical information. Bioinforma. Oxf. Engl. 39, btad080 (2023).\n14. Baranzini, S. E. et al. A biomedical open knowledge network harnesses the power of AI to understand\ndeep human biology. AI Mag. 43, 46–58 (2022).\n15. Lobentanzer, S. et al. Democratizing knowledge representation with BioCypher. Nat. Biotechnol. 41,\n1056–1059 (2023).\n16. Wilkinson, M. D. et al. The FAIR Guiding Principles for scientific data management and stewardship. Sci.\nData 3, 160018 (2016).\n17. Jain, Y . et al. Segmenting functional tissue units across human organs using community-driven\ndevelopment of generalizable machine learning algorithms. Nat. Commun. 14, 4656 (2023).\n18. Michael Hausenblas. 5-star Open Data. http://5stardata.info/en/ (2024).\n19. Börner, K. et al. Tissue registration and exploration user interfaces in support of a human reference atlas.\nCommun. Biol. 5, 1369 (2022).\n20. Golbreich, C., Grosjean, J. & Darmoni, S. J. The Foundational Model of Anatomy in OWL 2 and its use.\nArtif. Intell. Med. 57, 119–132 (2013).\n21. Rosse, C. & Mejino, J. L. V. A reference ontology for biomedical informatics: the Foundational Model of\nAnatomy. J. Biomed. Inform. 36, 478–500 (2003).\n22. Börner, K. et al. Anatomical structures, cell types and biomarkers of the Human Reference Atlas. Nat. Cell\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nBiol. 23, 1117–1128 (2021).\n23. Hao, Y . et al. Integrated analysis of multimodal single-cell data. Cell 184, 3573-3587.e29 (2021).\n24. Domínguez Conde, C. et al. Cross-tissue immune cell analysis reveals tissue-specific features in humans.\nScience 376, eabl5197 (2022).\n25. Xu, C. et al. Automatic cell-type harmonization and integration across Human Cell Atlas datasets. Cell 186,\n5876-5891.e20 (2023).\n26. Ergen, C. et al. Consensus prediction of cell type labels in single-cell data with popV. Nat. Genet. (2024)\ndoi:10.1038/s41588-024-01993-3.\n27. Quardokus, E. M. et al. Organ Mapping Antibody Panels: a community resource for standardized\nmultiplexed tissue imaging. Nat. Methods 20, 1174–1178 (2023).\n28. Berners-Lee, T . Linked Data - Design Issues. https://www.w3.org/DesignIssues/LinkedData.html (2006).\n29. Moxon, S. et al. The Linked Data Modeling Language (LinkML): A General-Purpose Data Modeling\nFramework Grounded in Machine-Readable Semantics. in CEUR Workshop Proceedings vol. 3073\n148–151 (2021).\n30. Färber, M., Lamprecht, D., Krause, J., Aung, L. & Haase, P . SemOpenAlex: The Scientific Landscape in 26\nBillion RDF Triples. in The Semantic Web – ISWC 2023 (eds. Payne, T . R. et al.) vol. 14266 94–112\n(Springer Nature Switzerland, Cham, 2023).\n31. Jaradeh, M. Y ., Oelen, A., Prinz, M., Stocker, M. & Auer, S. Open Research Knowledge Graph: A System\nWalkthrough. in Digital Libraries for Open Knowledge: 23rd International Conference on Theory and\nPractice of Digital Libraries, TPDL 2019, Oslo, Norway, September 9-12, 2019, Proceedings 348–351\n(Springer-Verlag, Berlin, Heidelberg, 2019). doi:10.1007/978-3-030-30760-8_31.\n32. Xu, J. et al. Building a PubMed knowledge graph. Sci. Data 7, 205 (2020).\n33. T orvik, V. I. MapAffil: A Bibliographic T ool for Mapping Author Affiliation Strings to Cities and Their\nGeocodes Worldwide. -Lib Mag. Mag. Digit. Libr. Forum 21, (2015).\n34. Cavalleri, E. et al. An ontology-based knowledge graph for representing interactions involving RNA\nmolecules. Sci. Data 11, 906 (2024).\n35. Wood, E. C. et al. RTX-KG2: a system for building a semantically standardized knowledge graph for\ntranslational biomedicine. BMC Bioinformatics 23, 400 (2022).\n36. Kilicoglu, H., Shin, D., Fiszman, M., Rosemblat, G. & Rindflesch, T . C. SemMedDB: a PubMed-scale\nrepository of biomedical semantic predications. Bioinformatics 28, 3158–3160 (2012).\n37. Mendez, D. et al. ChEMBL: towards direct deposition of bioassay data. Nucleic Acids Res. 47, D930–D940\n(2019).\n38. Milacic, M. et al. The Reactome Pathway Knowledgebase 2024. Nucleic Acids Res. 52, D672–D678\n(2024).\n39. Fabregat, A. et al. Reactome graph database: Efficient access to complex pathway data. PLOS Comput.\nBiol. 14, e1005968 (2018).\n40. Wishart, D. S. et al. DrugBank: a comprehensive resource for in silico drug discovery and exploration.\nNucleic Acids Res. 34, D668–D672 (2006).\n41. Miron, L., Gonçalves, R. S. & Musen, M. A. Obstacles to the reuse of study metadata in ClinicalTrials.gov.\nSci. Data 7, 443 (2020).\n42. Musen, M. A. et al. The center for expanded data annotation and retrieval. J. Am. Med. Inform. Assoc. 22,\n1148–1152 (2015).\n43. O’Connor, M. J. et al. Unleashing the value of Common Data Elements through the CEDAR Workbench.\nAMIA. Annu. Symp. Proc. 2019, 681–690 (2020).\n44. Caron, A. R. et al. A general strategy for generating expert-guided, simplified views of ontologies.\n2024.12.13.628309 Preprint at https://doi.org/10.1101/2024.12.13.628309 (2024).\n45. Surles-Zeigler, M. C. et al. Extending and using anatomical vocabularies in the stimulating peripheral\nactivity to relieve conditions project. Front. Neuroinformatics 16, (2022).\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\n46. Imam, F . T .et al. Developing a Multiscale Neural Connectivity Knowledgebase of the Autonomic Nervous\nSystem. 2024.10.25.620360 Preprint at https://doi.org/10.1101/2024.10.25.620360 (2024).\n47. SenNet Consortium et al. NIH SenNet Consortium to map senescent cells throughout the human lifespan\nto understand physiological health. Nat. Aging 2, 1090–1100 (2022).\n48. Dekker, J. et al. The 4D nucleome project. Nature 549, 219–226 (2017).\n49. Lonsdale, J. et al. The Genotype-Tissue Expression (GTEx) project. Nat. Genet. 45, 580–585 (2013).\n50. Quardokus, E. M., Record, E. & Herr II, B. W. SOP: Authoring Anatomical Structures, Cell Types and\nBiomarkers (ASCT+B) T ables. (2022) doi:10.5281/ZENODO.5746152.\n51. Bajema, R. Creating 2D Illustrations for Functional Tissue Units (FTUs). (2022)\ndoi:10.5281/zenodo.7409575.\n52. Schlehlein, H. & Quardokus, E. M. SOP: Creating 3D Models from Datasets. (2022)\ndoi:10.5281/zenodo.7384276.\n53. Quardokus, E. M., Bueckle, A., Borner, K., Record, E. & Browne, K. SOP: 3D Reference Object Approval.\n(2022) doi:10.5281/zenodo.5944197.\n54. Ju, Y . & Jain, Y . SOP: Computing Cell Type to Vasculature Distance Distributions. (2023)\ndoi:10.5281/ZENODO.10371473.\n55. Weber, G. & Gustilo, K. Authoring the Blood Vasculature Geometry T able. (2024)\ndoi:10.5281/zenodo.11623223.\n56. Weber, G. & Gustilo, K. Authoring the Pathway Organ Crosswalk T ables. (2024)\ndoi:10.5281/zenodo.11623745.\n57. Weber, G. & Gustilo, K. Constructing Blood Vasculature-Organ Crosswalk Diagrams. (2024)\ndoi:10.5281/zenodo.11623898.\n58. Kienle, P ., Quardokus, E. M. & Bueckle, A. Constructing a Millitome and Generating Virtual Tissue Blocks.\n(2023) doi:10.5281/ZENODO.7901004.\n59. Radtke, A. J. & Quardokus, E. M. SOP: Construction of Organ Mapping Antibody Panels for Multiplexed\nAntibody-Based Imaging of Human Tissues. Preprint at https://doi.org/10.5281/zenodo.5749883 (2021).\n60. McMahon, A. P . et al. GUDMAP: The Genitourinary Developmental Molecular Anatomy Project. J. Am.\nSoc. Nephrol. 19, 667 (2008).\n61. Rozenblatt-Rosen, O., Stubbington, M. J. T ., Regev, A. & T eichmann, S. A. The Human Cell Atlas: from\nvision to reality. Nature 550, 451–453 (2017).\n62. The Human Tumor Atlas Network (HTAN): exploring tumor evolution in time and space. Nature\nhttps://www.nature.com/collections/fihchcjehc (2024).\n63. Bidanta, S. et al. Functional Tissue Units in the Human Reference Atlas. bioRxiv 2023.10.16.562593\n(2023) doi:10.1101/2023.10.16.562593.\n64. Seal, R. L. et al. Genenames.org: the HGNC resources in 2023. Nucleic Acids Res. 51, D1003–D1009\n(2023).\n65. Kong, Y . & Börner, K. Publication, funding, and experimental data in support of Human Reference Atlas\nconstruction and usage. Sci. Data 11, 574 (2024).\n66. Bueckle, A. et al. The HRA Organ Gallery affords immersive superpowers for building and exploring the\nHuman Reference Atlas with virtual reality. Front. Bioinforma. 3, (2023).\n67. Wolf, F . A., Angerer, P . & Theis, F . J. SCANPY: large-scale single-cell gene expression data analysis.\nGenome Biol. 19, 15 (2018).\n68. Harris, C. R. et al. Array programming with NumPy. Nature 585, 357–362 (2020).\n69. Isaac Virshup, Sergei Rybakov, Fabian J. Theis, Philipp Angerer, & F . Alexander Wolf. anndata: Annotated\ndata. bioRxiv 2021.12.16.473007 (2021) doi:10.1101/2021.12.16.473007.\n70. Martin, F . J. et al. Ensembl 2023. Nucleic Acids Res. 51, D933–D941 (2023).\n71. T an, S. Z. K. et al. Brain Data Standards - A method for building data-driven cell-type ontologies. Sci. Data\n10, 50 (2023).\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\n72. Ma, J. et al. Segment Anything in Medical Images and Videos: Benchmark and Deployment. Preprint at\nhttps://doi.org/10.48550/arXiv.2408.03322 (2024).\n73. Ma, J. et al. Segment anything in medical images. Nat. Commun. 15, 654 (2024).\n74. Jain, Y . et al. Segmentation of human functional tissue units in support of a Human Reference Atlas.\nCommun. Biol. 6, 717 (2023).\n75. Wang, X. (Julie) et al. Generalized cell phenotyping for spatial proteomics with language-informed vision\nmodels. 2024.11.02.621624 Preprint at https://doi.org/10.1101/2024.11.02.621624 (2024).\n76. Israel, U. et al. A Foundation Model for Cell Segmentation. Preprint at\nhttps://doi.org/10.48550/arXiv.2311.11004 (2023).\n77. Jain, Y . et al. Vasculature segmentation in 3D hierarchical phase-contrast tomography images of human\nkidneys. 2024.08.25.609595 Preprint at https://doi.org/10.1101/2024.08.25.609595 (2024).\n78. Yagis, E. et al. Deep learning for 3D vascular segmentation in hierarchical phase contrast tomography: a\ncase study on kidney. Sci. Rep. 14, 27258 (2024).\n79. Brbić, M. et al. Annotation of spatially resolved single-cell data with STELLAR. Nat. Methods 19,\n1411–1418 (2022).\n80. Greenwald, N. F . et al. Whole-cell segmentation of tissue images with human-level performance using\nlarge-scale data annotation and deep learning. Nat. Biotechnol. 40, 555–565 (2022).\n81. Pan, H. et al. Comprehensive anatomic ontologies for lung development: A comparison of alveolar\nformation and maturation within mouse and human lung. J. Biomed. Semant. 10, 18 (2019).\n82. Ontology Lookup Service. Provisional Cell Ontology. https://www.ebi.ac.uk/ols4/ontologies/pcl.\n83. Weber, G. M., Ju, Y . & Börner, K. Considerations for Using the Vasculature as a Coordinate System to Map\nAll the Cells in the Human Body. Front. Cardiovasc. Med. 7, (2020).\n84. Cyberinfrastructure for Network Science Center. HRA Organ Gallery on Horizon Store. Oculus\nhttps://www.meta.com/experiences/quest/5696814507101529/ (2024).\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nSupplemental Figures\nSupplemental Figure 1. An excerpt of a normalized structure from the ASCT+B table for the kidney, with a focus on\ncell_marker_descriptors, where primary cell type, primary anatomical location, and associated characterizing biomarkers\nare detailed, and where references to support the claims regarding the cell type and its biomarkers are provided. An\nexample in YAML is provided at\nhttps://github.com/cns-iu/hra-kg-supporting-information/blob/main/docs/intermediary_format.yaml.\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nSupplemental Figure 2. This diagram shows an excerpt of a LinkML schema (represented by the two boxes on the left)\nand the corresponding acquired data (the box on the right). Arrows illustrate the transformation of the input text into OWL\nconstructs at the bottom. In the schema boxes, classes and slots are directly mapped to OWL classes and OWL\nproperties as specified in the annotations field. In the data box, each data item pair is translated into an OWL assertion\nstatement. For example, the data pair type_of: CellMarkerDescriptor generates a class assertion that indicates the data\nobject belongs to the CellMarkerDescriptor class. Similarly, the data pair primary_cell_type: CL:0002306 produces an\nannotation assertion that tells the same data object identifies the epithelial cell in the proximal tubule of the kidney\n(CL:0002306) as the primary cell type.\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nSupplemental T ables\nSupplemental Table 1. Graphs of the HRA KG (https://lod.humanatlas.io/graph).\nName Description Link\n2d-ftu-illustrations Contains 2D assets for FTUs. https://lod.humanatlas.io/graph/2d\n-ftu-illustrations/\nccf Graph representation of the CCF .OWL dataset. https://lod.humanatlas.io/graph/ccf\n/\nctann-crosswalks Graph representation of cell type annotation tool\ncrosswalks for HRApop (Azimuth 23, CellTypist24,25\npopV26).\nhttps://lod.humanatlas.io/graph/cta\nnn-crosswalks/\nds-graphs-enrichments Enriched graph of all public dataset graphs for the\nHRA.\nhttps://lod.humanatlas.io/graph/ds-\ngraphs-enrichments/\nhra-ccf-body Graph representation of the 3D Reference Objects\nfor human male and female bodies.\nhttps://lod.humanatlas.io/graph/hr\na-ccf-body/\nhra-ccf-patches Graph that provides placements (patches) to convert\nold 3D Reference Object registrations to newer\nversions for reference organs that had a different\nshape, placement, or size in earlier versions.\nhttps://lod.humanatlas.io/graph/hr\na-ccf-patches/\nhra-pop Contains donor and sample metadata as well as 3D\nextraction sites and cell summaries for 619 ATLAS\ndatasets used in HRA construction (as of v0.11.1).\nCode to compute mean B expression values per cell\nper dataset is at\nhttps://github.com/hubmapconsortium/hra-workflows/\nblob/main/containers/gene-expression/context/main.\npy#L108-L151\nExemplarily, the lookup table between Ensembl\nRelease 11164 (https://www.ensembl.org/index.html)\nand HGNC v2023-09-18 70\n(https://www.genenames.org) for CellTypist is at\nhttps://github.com/hubmapconsortium/hra-workflows/\nblame/main/containers/celltypist/context/ensemble-lo\nokup.csv\nhttps://lod.humanatlas.io/graph/hr\na-pop/\nHRAlit (currently\nserved as a relational\ndatabase)\nLinks the HRA DOs to publications, experts,\nexperimental datasets, and funded projects 65 to\nidentify relevant papers, SMEs, alignment with\nexisting ontologies, and funding trends to HRA\nconstruction and usage.\nN/A\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nSupplemental Table 2. Collections of the HRA KG (https://lod.humanatlas.io/collection). All structures are built using\nstandardized terms for labels, which are stored in the specialized CCF vocabulary, see CCF Ontology in Box 1.\nName Description Link\nds-graphs Consolidates experiment-result datasets (ds-graph) into a single graph\ndata structure. Contains processed graphs of datasets including donor\nand sample metadata from different consortia and tissue providers\nregistered to HRA via 3D extraction sites, available at\nhttps://lod.humanatlas.io/ds-graph. Also present are HRA graph DOs\n(https://lod.humanatlas.io/graph), which are externally processed RDF\ngraphs, such as crosswalks between cell types assigned by cell type\nannotation tools and cell types in ASCT+B tables.\nhttps://lod.humanatlas.io/co\nraphs\nhra Includes HRA DO types with a DOI. As of HRA v2.2, those are 2d-ftu,\nasct-b, ctann, omap, ref-organ, and vascular-geometry DO types. Other\nDO types can be added in the future once they have DOIs.\nhttps://lod.humanatlas.io/\ncollection/hra\nhra-api Centers on asct-b and associated 3D models (ref-organ, landmark) used\nin various HRA applications for visualizing and organizing reference data.\nhttps://lod.humanatlas.io/\ncollection/hra-api\nhra-ols HRA collection subset that focuses solely on AS, CT , and B (asct-b only).\nIt is used by OLS (https://www.ebi.ac.uk/ols4) for validation. The source\nfor the files on the LOD server are available at\nhttps://github.com/hubmapconsortium/3d-hra-ref-object-validation.\nhttps://lod.humanatlas.io/\ncollection/hra-ols\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nSupplemental Table 3. Most HRA applications use the HRA KG as their backend database.\nName Description URL\nHRA Portal Provides the landing page and additional web pages for exploring\nHRA-related resources. Some data statistics on the page are aggregated\nfrom the HRA KG via SPARQL queries.\nhttps://humanatlas.io\nRUI Allows users to register new tissue samples and related datasets into the\nhuman reference atlas 19. The “parts” hierarchy on the navigation panel and\nthe 3D human model at the center of the page are generated from the HRA\nKG via HRA API queries.\nhttps://apps.humanatla\ns.io/rui\nEUI Enables users to explore and query the HRA KG interactively 19. All\nvisualizations, data facets, statistics, and data browsers are generated\nthrough real-time querying of the knowledge graph.\nhttps://apps.humanatla\ns.io/eui\nASCT+B\nReporter\nEnables users to inspect ASCT+B T ables as tree visualizations 19 and\nsupports OMAP comparison to ASCT+B tables. Uses the HRA KG to serve\nprocessed and published ASCT+B T ables to the website.\nhttps://apps.humanatla\ns.io/asctb-reporter\nFTU\nExplorer\nFacilitates interactive exploration of FTUs within the HRA. The HRA KG\nprovides the 2d-ftu DOs that the FTU Explorer displays. Datasets and cell\nsummaries registered to these FTUs are currently separate, but will be\navailable in the HRA KG as ds-graph DOs in the future.\nhttps://humanatlas.io/2\nd-ftu-illustrations\nHRA\nDashboard\nOffers a summary and visual representation of data statistics and trends\nwithin the HRA knowledge graph. Data statistics and charts are generated by\naggregating data from the HRA KG.\nhttps://apps.humanatla\ns.io/dashboard\nHRA API\nendpoints\nProvides programmatic access to the HRA KG, enabling integration with\nother applications and services. A SPARQL endpoint for the HRA API allows\nusers to write their own SPARQL queries. Canned queries are available at\nhttps://apps.humanatlas.io/api/grlc.\nAPI documentation:\nhttps://humanatlas.io/a\npi\nHRA Organ\nGallery\nEnables an immersive view of the HRA by showing 71 reference organs and\n1,100+ tissue blocks alongside cell type populations in VR. 66,84\nhttps://humanatlas.io/h\nra-organ-gallery\nCell\nDistance\nExplorer\nAllows distance visualizations between cells in 2D and 3D. Features\nhistograms and violin graphs for vascular distances by CT . Does not use\nHRA KG but will use it for crosswalking CT s to CL and PCL in the future.\nhttps://apps.humanatla\ns.io/cde\nIn Data Portals\nHuBMAP\nData Portal\nSame as EUI above but only showing HuBMAP data https://portal.hubmapc\nonsortium.org/ccf-eui\nSenNet\nData Portal\nSame as EUI above but only showing SenNet data https://data.sennetcon\nsortium.org/ccf-eui\nGTEx Portal Same as EUI above but only showing GTEx data https://gtexportal.org/h\nome/eui\nKPMP\nTissue Atlas\nSame as EUI above but only showing KPMP data https://apps.humanatla\ns.io/eui/kpmp.html\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nSupplemental Table 4. An overview of all GitHub repositories used to construct, deploy, and use the HRA KG.\nName Description URL\nCode\nhra-kg HRA Knowledge Graph https://github.com/hubmapconsortium/hra-kg\nhra-api HRA API https://github.com/x-atlas-consortia/hra-api\nhra-do-processor HRA DO Processor https://github.com/hubmapconsortium/hra-do-\nprocessor\nccf-grlc Repository of canned SPARQL\nqueries that can be run like a REST\nAPI endpoint via https://grlc.io\nhttps://github.com/hubmapconsortium/ccf-grlc\nhra-ubkg-exporter A CLI for exporting a subset of the\nHRA KG to Unified Biomedical\nKnowledge Graph (UBKG) format\nhttps://github.com/x-atlas-consortia/hra-ubkg-\nexporter\nAPIs and Documentation\nHRA KG SPARQL endpoint SPARQL endpoint of the HRA KG https://lod.humanatlas.io/sparql\nHRA API UI with documentation for all HRA\nAPI endpoints and routes\nhttps://apps.humanatlas.io/api\nHRA API SPARQL queries Canned SPARQL queries via grlc.io\non HRA API\nhttps://apps.humanatlas.io/api/grlc\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nSupplemental Table 5. Publications on HRA DOs plus aliases used throughout HRA publications and applications.\nHRA DO Type Aliases Publication Title DOI\n2d-ftu - Functional Tissue Units\n- Functional Tissue Units (FTUs)\n- FTUs\nFunctional Tissue Units in the Human\nReference Atlas\nhttps://doi.org/10.11\n01/2023.10.16.562\n593\nasct-b - Anatomical Structures, Cell\nTypes, and Biomarker T ables\n- Anatomical Structures, Cell\nTypes, and Biomarker T ables\n(ASCT+B) T ables\n- ASCT+B T ables\nAnatomical structures, cell types and\nbiomarkers of the Human Reference Atlas\nhttps://www.nature.\ncom/articles/s4155\n6-021-00788-6\nAnatomical structures, cell types, and\nbiomarkers of the healthy human blood\nvasculature\nhttps://www.nature.\ncom/articles/s4159\n7-023-02018-0\nctann - Cell Type Annotations\n- Ctann\n- CTAnn\nlandmark - Landmark Organs\n- Landmark Anatomical\nStructures\nmillitome - Millitome\nomap - Organ Mapping Antibody Panels\n- Organ Mapping Antibody Panels\n(OMAPs)\n- OMAPs\n- Organ Mapping Antibody Panel\nT ables\nOrgan Mapping Antibody Panels: a\ncommunity resource for standardized\nmultiplexed tissue imaging\nhttps://www.nature.\ncom/articles/s4159\n2-023-01846-7\nref-organ - 3D Reference Object\n- 3D Reference Organ\nTissue registration and exploration user\ninterfaces in support of a human reference\natlas\nhttps://www.nature.\ncom/articles/s4200\n3-022-03644-x\nschema - Schema\nvascular-geometr\ny\nAnatomical structures, cell types, and\nbiomarkers of the healthy human blood\nvasculature\nhttps://www.nature.\ncom/articles/s4159\n7-023-02018-0\nConsiderations for Using the Vasculature\nas a Coordinate System to Map All the\nCells in the Human Body\nhttps://doi.org/10.33\n89/fcvm.2020.0002\n9\nvocab Specimen, biological structure, and spatial\nontologies in support of a Human\nReference Atlas\nhttps://www.nature.\ncom/articles/s4159\n7-023-01993-8\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\nSupplemental Table 6. HRA User Stories and how the HRA KG supports them.\nHRA User Story Value Added by HRA KG HRA KG Sample Queries\nUS#1. Predict cell type\npopulations\nUse the HRA KG to access the hra-pop\ngraph\n(https://lod.humanatlas.io/graph/hra-pop/lat\nest) with cell type populations for AS,\ndatasets, and extraction sites to improve\nthe accuracy of annotations for\nsc-transcriptomics and sc-proteomics\ndatasets. In the future, the HRA KG could\nbe used to store metadata for millions of\ncells individually.\n/as-weighted-cell-summaries:\nhttps://github.com/x-atlas-consortia/hra-api/b\nlob/main/src/library/hra-pop/queries/as-weig\nhted-cell-summaries.rq\nAccessible via HRA API:\nhttps://apps.humanatlas.io/api/hra-pop/rui-lo\ncation-cell-summary\nUS#2. Predict spatial\norigin of tissue samples\nUse the cell type populations from the\nhra-pop graph mentioned above to predict\nthe 3D location of datasets with unknown\nspatial origin.\n/select-cell-summaries:\nhttps://github.com/x-atlas-consortia/hra-api/b\nlob/main/src/library/hra-pop/queries/select-c\nell-summaries.rq\nAccessible via HRA API:\nhttps://apps.humanatlas.io/api/hra-pop/cell-s\nummary-report\n/supported-organs:\nhttps://github.com/x-atlas-consortia/hra-api/b\nlob/main/src/library/hra-pop/queries/support\ned-organs.rq\n/supported-reference-organs:\nhttps://github.com/x-atlas-consortia/hra-api/b\nlob/main/src/library/hra-pop/queries/support\ned-reference-organs.rq\n/supported-tools:\nhttps://github.com/x-atlas-consortia/hra-api/b\nlob/main/src/library/hra-pop/queries/support\ned-tools.rq\nUS#3. Compare\nreference tissue with\naging/diseased tissue\nUse the EUI to examine AS, CT , and B. Via\nthe HRA API, the RUI runs SPARQL\nqueries of the HRA KG to retrieve the AS\npartonomy, CT typology, and Bs.\n/tissue-blocks:\nhttps://apps.humanatlas.io/api/#get-/v1/tissu\ne-blocks (then user parameters to get\ndatasets by age (range), BMI (range), sex\nAccessible via HRA API:\nhttps://apps.humanatlas.io/api/#get-/v1/tissu\ne-blocks (then user parameters to get\ndatasets by age (range), BMI (range), sex\n/scene:\nhttps://github.com/x-atlas-consortia/hra-api/b\nlob/main/src/library/v1/queries/scene.rq\nNote: #{{FILTER}} gets replaced with filters\ndocumented on HRA API endpoint at\nhttps://apps.humanatlas.io/api/#get-/v1/scen\ne to get datasets by age (range), BMI\n(range), sex, etc.\nUS#4. Compare\nreference FTUs with\naging/diseased FTUs\nUse the HRA KG to retrieve 2D illustrations\nof 22 FTUs in the FTU Explorer\n(https://apps.humanatlas.io/ftu-explorer/#/)\n/ftu-parts:\nhttps://apps.humanatlas.io/api/grlc/hra.html#\nget-/ftu-parts\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint \n\n/2d-ftu-illustrations (DO)\nhttps://lod.humanatlas.io/graph/2d-ftu-illustra\ntions/\nUS#5. Provide cell\ndistance distribution\nvisualizations\nUse the Cell Distance Explorer to visualize\ndistance distributions between different\ncells and cell types in 2D or 3D\nsc-proteomics tissue. In the future, the\nHRA KG will be used to crosswalk cell type\nlabels to CL\n(https://obofoundry.org/ontology/cl.html)4 to\nadd automated grouping features. Cell\ndistances could also be stored in the HRA\nKG.\nNo sample queries yet, but support is\npossible for adding comparisons to other\ndatasets or simpler color scheme via CL\nonce crosswalks exist\nUS#6. Develop\nlightweight atlas\ncomponents\nUse the HRA KG to enable HRA web\ncomponents to improve data access and\nanalysis outside of the HRA applications\necosystem\n(https://apps.humanatlas.io/us6/). More\ninformation is provided under Results >\nUsing the HRA KG > HRA Applications.\nComponents use HRA KG, e.g., EUI\ncomponent uses HRA KG via HRA API,\nsame for RUI and FTU Explorer (see US#3\nand 4)\nUS#7. Implement\ndashboard for HRA\nExplore usage statistics of atlas data and\ncode to check HRA growth and outreach\nover time. SPARQL queries to the HRA KG\nserve stats on experimental data and all\nHRA DO types\n/digital-objects-per-organ:\nhttps://github.com/x-atlas-consortia/hra-dash\nboard-data/blob/main/queries/sparql/data/di\ngital-objects-per-organ.rq\n/as-per-organ:\nhttps://github.com/x-atlas-consortia/hra-dash\nboard-data/blob/main/queries/sparql/human\natlas.io/as-per-organ.rq\nhra-growth.atlas.rq\nhttps://github.com/x-atlas-consortia/hra-dash\nboard-data/blob/main/queries/sparql/data/hr\na-growth.atlas.rq\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted December 23, 2024. ; https://doi.org/10.1101/2024.12.22.630006doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}