Report from the CFDE Knowledge Graph Working Group, Fall 2026 Meeting

Abstract

The Common Fund Data Ecosystem (CFDE) Knowledge Graph Working Group met on October 14, 2026, during the CFDE Fall meeting, for a roughly 90-minute hybrid session led by Deanne Taylor. Bernard de Bono presented ApiNATOMY, a framework used within the NIH SPARC program that represents anatomical connectivity from membrane channels to whole-body systems with a single recursive grammar of layered conduits, and that maps variables of physiology models onto anatomical compartments. Participants identified an integration path from ApiNATOMY through the Human Reference Atlas (HRA) knowledge graph, which is already interoperable with the CFDE Data Distillery Knowledge Graph (DDKG).

Discussion then turned to the growing number of CFDE knowledge graphs. Rather than requiring one shared schema, the group favored exchange through a JSON-based distribution format (JKG), now roughly three-quarters complete, and access by AI applications through Model Context Protocol (MCP) servers. Redistribution restrictions on UMLS-derived content, SNOMED CT, and DrugBank currently prevent a public MCP server for the full DDKG; a server limited to Data Coordinating Center (DCC) data and openly licensed ontologies was proposed as a first step.

Two further proposals addressed the scientific value of graph content. Portable Assertion Records (PARs) would attach provenance and evidence to each assertion, so that users could assemble graphs at an evidentiary threshold of their choosing. A Talent Knowledge Graph of investigator expertise, linked to literature-derived evidence, could route computational hypotheses to researchers able to test them. Most items discussed are proposals rather than implemented capabilities. This report summarizes the presentations and discussion, distinguishes established work from plans, and lists open questions and proposed next steps.

Introduction

The NIH Common Fund Data Ecosystem (CFDE) links data and tools produced by Common Fund programs, each served by its own Data Coordinating Center (DCC). The Data Distillery Knowledge Graph (DDKG) is the CFDE's main integrative graph. It is built on the Unified Biomedical Knowledge Graph (UBKG) framework, draws on more than 180 ontologies and controlled vocabularies, and harmonizes datasets from programs including 4DN, exRNA, GlyGen, GTEx, HuBMAP, IDG, Kids First, LINCS, Metabolomics Workbench, MoTrPAC, and SPARC¹.

The CFDE Knowledge Graph Working Group, chaired by Deanne Taylor (Children's Hospital of Philadelphia, CHOP), coordinates knowledge graph development across the ecosystem. Its Fall 2026 session took place on October 14, 2026, at the CFDE Fall meeting. The session lasted about 90 minutes. Participants attended in person and over Zoom, with the chair joining remotely.

The session had two parts. Bernard de Bono presented ApiNATOMY and answered questions. The chair then led an open discussion of the CFDE knowledge graph landscape, AI access, licensing, evidence representation, and links to investigator expertise. This report records those discussions for the wider CFDE and biomedical knowledge graph communities.

About this report

This report was prepared from a transcript generated with Whisper (large-v2) from audio recorded in the meeting room, and from a structured summary derived from that transcript. The transcript contains many recognition errors, especially for names and technical terms (for example, "apimatomy" for ApiNATOMY and "CLDE" for CFDE). These were corrected where the intended term was clear from context. The report was drafted with the assistance of an AI agent (Claude, Anthropic) operating on behalf of the submitting author, who attended the session.

Speakers are named where they identified themselves or were clearly identifiable. Several contributors could not be identified from the transcript and are described as participants. Statements are reported as positions taken in discussion, not as consensus of the working group or official CFDE policy. Throughout, the report separates work described as existing from work described as planned or proposed.

ApiNATOMY: a common grammar for multiscale physiology

Bernard de Bono presented ApiNATOMY, a knowledge representation and management framework for physiology routes across scales^2,3. Within SPARC, which focuses on the autonomic nervous system, ApiNATOMY underlies a whole-body topological map. In it, an outer ring represents epithelial surfaces, an inner ring the circulation, and a T-shaped route the flow of cerebrospinal fluid.

Motivation: biological structures are regionalized

The central argument was that treating a cell or protein as a single unit discards physiological information. Different regions of one structure face different biophysical compartments and carry different molecular machinery. A neuron may run as a single cell from the hypothalamus to the colon, with each segment exposed to a different environment. An enterocyte has apical, lateral, and basolateral regions, facing the gut lumen, tight junctions, and the circulation respectively. A membrane protein has extracellular, transmembrane, and cytosolic regions with distinct binding properties. A consistent knowledge representation must therefore regionalize structures and place each region in its compartment.

de Bono connected this to his earlier definition of functional tissue units (FTUs)⁴. An FTU is a block of cells arranged around a central channel, such as a capillary, so that no cell lies beyond diffusion distance of the channel. The channel carries long-range communication beyond that limit.

The building block and its operations

ApiNATOMY generalizes the FTU into a single building block: a layered conduit with an axis, an ordered set of layers, and polarity. It comes in three configurations: a tube (open at both ends), a bag (closed at one end), and a cyst (closed at both ends). Each layer is annotated with reference ontology terms and with its material composition. Composition is itself built up compositionally, for example ions into extracellular fluid, extracellular fluid plus albumin into plasma, and plasma plus red cells into blood.

Four operations assemble these blocks across scales:

1. Chaining: conduits joined in series. A membrane channel is three tubes (cytosolic, membrane, extracellular); an epithelial cell is apical, lateral, and basolateral segments; a neuron is a branched chain of dendrites, soma, axon, and terminals.
2. Housing: a chain at one scale placed within the layers of a larger conduit, such as a channel within the apical surface of a cell, or an endothelial cell spanning the layers of a vessel wall.
3. Long-range routing: chains threaded through larger structures, such as neurons along spinal cord segments from the midbrain to the bladder in the control of micturition.
4. Coalescence: two conduits sharing an outer layer. In the kidney, blood, endothelium, a shared basement membrane, epithelium, and urine form the route of exchange between blood and urine.

Using the same grammar, de Bono has built lung, kidney, and liver models, made available through the CFDE Workbench. The approach is also being used to map fascial layers onto the Visible Human dataset, with segments of the aorta placed onto that fascial model.

Linking anatomy to mathematical models

ApiNATOMY serves two communities with one reference network of compartments. For SPARC, it represents routes of neuron segments whose sensory endings track physiological fields and whose secretory endings adjust fluxes. For the bioengineering community, it supports assembling compound ordinary differential equation (ODE) models from simpler ones, by mapping each model variable to an anatomical compartment. Simulation results can then be displayed on the anatomical map.

The stated aim is to make knowledge held in physiology models available to the omics community. A disease signal could then be interpreted as changes in pressure, wall elasticity, permeability, or concentration within a defined network of compartments.

Status and access

In answer to questions, de Bono clarified that ApiNATOMY is not currently part of the DDKG. He proposed membrane proteins as a first joint use case, with ApiNATOMY supplying topological context for molecular assertions. Graphical viewers exist, AI skill files assist with constructing and querying models, and an MCP server is in early development. A natural-language interface was not described as complete.

Asked whether additional layers could be represented, such as the glycocalyx and the mucus layers of the gut that support immune surveillance, de Bono said the layering is fully customizable. New use cases can add layers without changing the schema.

Connecting ApiNATOMY to CFDE data through the Human Reference Atlas

The clearest collaboration opportunity of the session was a path from ApiNATOMY to CFDE data through the Human Reference Atlas (HRA). The HRA is a multiscale atlas of the healthy adult human body developed within HuBMAP and partner consortia⁵. Its knowledge graph links anatomical structures, cell types, and biomarkers to 3D reference organs⁶.

Taylor noted that the HRA team, including Katy Börner, Bruce Herr, and Andreas Bueckle (Indiana University), builds knowledge graph infrastructure that connects to the UBKG through HuBMAP. She described the HRA knowledge graph and the DDKG as directly interoperable, based on prior joint work. The published HRA knowledge graph is deployed as linked open data queried with SPARQL⁶, so the exact exchange mechanism with the DDKG is worth documenting explicitly.

de Bono reported that integrating his FTU models into the HRA needs a roadmap, but that work has begun. He has shared his fascial model with Bueckle, who has placed it in a virtual reality environment under development. Both agreed that SPARC and HuBMAP resources should be more closely connected. Taylor proposed that she, de Bono, Herr, Bueckle, and Börner work together to bring ApiNATOMY models into the CFDE through the HRA, and invited Bueckle, who joined during the discussion, to take part. The proposal was received positively.

de Bono described a target use case. A user would select a compartment on an ApiNATOMY anatomical map, launch a query into the DDKG, and see the returned knowledge displayed on the same map. The anatomy map would thus act as both a browsing interface and a query trigger.

Whole Person Physiome

The NIH Whole Person Physiome (WPP) initiative, launched in September 2025 and led by the National Center for Complementary and Integrative Health, aims to build an integrated knowledge network of healthy physiological function on top of the HRA and HuBMAP⁷. de Bono confirmed that he collaborates with WPP, contributing neural connectivity knowledge. Taylor suggested that ApiNATOMY could sit between WPP and the CFDE, making CFDE data available within physiological models.

Participants framed this as a step toward supporting digital twins. Taylor cautioned that CFDE holdings are mostly observational data rather than modeling outputs such as flux analyses. Participants also noted that new funding opportunities might support this work, though none was specified in detail.

Representing knowledge across scales

Christophe Lambert asked how triple-based knowledge graphs can represent concepts that span many levels of abstraction, from molecules to organ systems, together with their spatial and temporal relations, and how inference can operate across those levels. The question drew two complementary answers.

de Bono answered that a single object type can represent a region of a membrane channel, a cell, a neuron, and a tissue unit. Reusing one object across scales keeps the schema small and machine-readable, because the underlying physics does not change with scale. He suggested that this uniformity makes cross-scale inference more tractable for AI systems.

Taylor described how the DDKG handles scale through ontology hierarchies. The Homo sapiens Chromosomal Location Ontology (HSCLO) represents genomic positions at nested resolutions¹, so a large feature such as a copy number variant contains smaller ones such as single nucleotide variants. In her framing, the chain of inheritance is itself the object of computation: a query can start at a large structure and descend to a smaller one, or ascend again. She compared this to compartmental modeling, where a small compartment inherits the parameters of its parent, and noted that the chosen modeling resolution determines which hierarchical relations matter. She added that this approach does not solve every problem.

Neither approach was presented as a complete solution to causal or temporal reasoning across scales. How to preserve context when inferring across levels remains an open question for the group.

Interoperability without a single schema

Taylor opened the second half by revisiting the working group charter. Since it was written, many CFDE groups have built their own knowledge graphs, covering physiology, gene regulation, literature, biomarkers, toxicology, proteins, and investigator expertise. The DDKG adopted the UBKG schema, but Taylor stated that other CFDE graphs need not adopt it. The working group's aim is instead to let graphs exchange content easily, for example moving ApiNATOMY content into the DDKG or the reverse.

JKG: a JSON exchange format

The DDKG is currently distributed as Docker images or CSV files¹. A new JSON-based distribution, referred to in the meeting as JKG, packages the same content as JSON files validated against a JSON schema and loadable directly into Neo4j. Taylor reported that 35 to 40 datasets had been converted, roughly three-quarters of the total, with most of the conversion done by Alan Simmons. A release announcement will follow completion.

The intended benefit is that another knowledge graph could ingest a dataset by export and upload, without a custom extract-transform-load step for each pair of graphs. This was presented as a goal rather than a demonstrated capability. In closing discussion, Taylor clarified that JKG is a format for transmitting graph content, not a separate knowledge graph.

MCP servers

Model Context Protocol (MCP) servers give AI applications a standard way to query external resources⁸. Taylor distinguished them from graph exchange. An MCP server delivers answers drawn from a graph, but it does not by itself let a user compute over the combined connectivity of several graphs. Exchanging graph records preserves that connectivity. She suggested both are worth pursuing.

This raised an architectural question that Taylor described as a CFDE-level decision. One option is a set of independent MCP servers, one per DCC dataset or per graph. The other is to integrate datasets into a shared graph first and expose that graph through MCP. Separate servers are simpler but lose cross-dataset associations; an integrated graph keeps them. She also noted that, in collaborations such as the Open Knowledge Network, many DDKG source datasets are replicated in other graphs, so the DDKG's distinctive value lies in its integration rather than in unique content.

For AI access today, the DDKG team has released an AI skill that users load into an assistant to query a local DDKG installation. Taylor said little else had yet been done toward AI integration.

An inventory of CFDE knowledge graphs

A participant from the Integration and Coordination Center (Jake) noted that several knowledge graphs shown at the meeting's technology showcase fell outside the working group's usual agenda. Taylor proposed writing up every CFDE knowledge graph currently under way, with its capabilities, platform, and specifications, as the basis for a future meeting on integration.

Licensing and access to the DDKG

Licensing is the main obstacle to open AI access to the DDKG. The graph includes the UMLS, and downloading it requires a free UMLS license key¹. Taylor identified SNOMED CT and DrugBank as further sources that cannot be redistributed openly. KEGG was mentioned as restricted, though she believed it is not included. Licensing information for each source is recorded on its source node in the graph.

The redistribution arrangement for UMLS content is held by Jonathan Silverstein's group at the University of Pittsburgh, which distributes the UBKG and DDKG. Without a broader arrangement, Taylor said she would not expose the full DDKG through a public MCP server.

Current access routes

Route	What it offers	Limits
Download and local install	Full graph in a local Neo4j database; the CHOP team offers setup help on laptops or institutional servers	Requires a UMLS license; no public hosting
Safe API (University of Pittsburgh)	Programmatic access through predefined queries1	Narrow by design; returns specific answer types only
DDKG-UI	Web interface over a curated subset of the graph1	Subset only
DDKG AI skill	Natural-language querying of a local installation through an AI assistant	Depends on a local install

Taylor explained that a full public Neo4j endpoint would raise security as well as licensing problems. An MCP server with constrained queries is, in her view, the closest available option for safe public delivery.

A licensed subset as a first step

A participant proposed starting with a smaller MCP server limited to data generated by the CFDE DCCs. Taylor agreed this could be built as a DDKG subgraph. Linking DCC data still requires supporting ontologies such as HPO and Uberon, which she expected to be licensable, but each source would need checking. She suggested offering both a simple data-delivery service and a richer graph for computation.

Release status

In response to a question about recent uploads, Taylor stated that the December 2025 release is the last complete DDKG release. Later uploads are interim or beta builds, used mainly to test JKG, and contain only partial data.

Evidence-aware assertions: Portable Assertion Records and Databases

Taylor presented an early proposal for representing the evidence behind each knowledge graph assertion in a portable form. Her starting point was that one edge, such as "this gene is associated with this disease", can rest on very different grounds. It may come from a statistical result, a curated database, a literature extraction, or one investigator's judgment, and these should not be treated as equivalent.

Edges as first-class objects

In the upcoming DDKG version, each edge becomes a first-class object carrying a small number of properties with numeric or string values. These properties can be reused for several purposes, including evidence levels. Several edges can connect the same pair of concepts, so that assertions from different sources remain distinct. Taylor proposed treating one edge as the minimal atomic assertion, accompanied by its evidence and its source, whether a dataset, a publication, or a person.

From records to databases

Evidence recorded only inside the DDKG schema could not easily be shared with other graphs. This motivated the proposal for Portable Assertion Records (PARs), each holding one assertion with its provenance and evidence. A collection of PARs forms a Portable Assertion Database (PAD), which might be as simple as a set of JSON records.

Taylor likened PARs to building blocks with defined connection points. Shared identifiers, such as HPO, MONDO, or Uberon terms, determine how records join into a larger graph. Unlike graph-internal provenance, as in SPOKE⁹, a PAR could also be consumed directly by an AI system as a standalone record, without first being assembled into a graph.

Filtering by evidence

The principal goal is to let users assemble graphs at an evidentiary threshold they choose. Evidence weight might draw on reproducibility, the number of independent observations, or statistical support, placed on a sliding scale. One researcher could build a graph from strongly supported assertions only; another could admit weaker ones for hypothesis generation. Taylor connected this to a point made by Jason Flannick the previous day: AI systems acting on graph content need a way to judge the weight of evidence behind it.

A standard, not a format

Taylor stressed that the proposal concerns content, not serialization. The questions are what an assertion record must contain, how evidence is stored, and what qualifies an assertion as trustworthy. PARs could be expressed in existing formats that meet those requirements. Nanopublications¹⁰ are close in spirit, but in her view lack an explicit treatment of evidence and trustworthiness; they might still serve as a vehicle for PARs. Each record should always point back to a citable source.

Practical evidence types

Near the end of the session, the group discussed how readily different assertions can carry evidence. Tissue expression from GTEx and QTL associations come with p-values that can be attached directly, although a p-value alone is not a complete measure of credibility. Gene–disease associations are harder because they combine genetic, functional, literature, and expert evidence. ClinVar and ClinGen supply some structured evidence, and the Knowledge Center groups of Jason Flannick and the REVEAL team are working on this problem. Only a limited number of DDKG datasets currently carry evidence levels; extending them to all datasets is a priority for the coming year.

Status

PARs and PADs are an early proposal, not an adopted CFDE standard. A draft document was shared during the session as the start of a paper. Taylor described it as a community project open to all CFDE participants. Because it may require new evidence vocabularies, she suggested coordinating with the CFDE Ontology Working Group.

Literature, expertise, and the path from hypothesis to validation

Literature-derived evidence in the DDKG

DDKG assertions have so far come mainly from experimental data and curated resources. Taylor reported that the next DDKG version will include PubMind, developed by Kai Wang's group at CHOP¹¹. PubMind uses large language models to extract variant–disease–pathogenicity associations from PubMed abstracts and PubMed Central full text, with supporting evidence. Taylor is transforming PubMind output into graph form and plans to use it as a model for how evidence is represented and computed over in the DDKG.

The DDKG team has chosen to integrate established literature-mining resources rather than build its own. Taylor suggested that similar extraction could yield gene–gene and gene–tissue assertions from the literature, though this has not yet been discussed with the PubMind developers.

The Talent Knowledge Graph

Jake described a Talent Knowledge Graph, developed at the Integration and Coordination Center, to support mentorship and collaborator discovery. It currently draws on publications, co-citation networks, and records of databases and tools. In one proposed use, a researcher exploring a narrow topic in ApiNATOMY or the DDKG, such as the gut bacterial barrier, would identify investigators who have published on it and who are reachable through citation links to the CFDE community.

Jake reported a test case in which the graph found a specialist in HIV antibody research within one step of the existing network. He estimated that cold outreach succeeds roughly 1% of the time and hoped network-guided outreach would be about ten times more effective; this was offered as an expectation, not a measured result. In his view, the DDKG's distinctive strength is its experimental data, which could be combined with literature-based graphs and with a human dimension of expertise.

Taylor proposed that evidence should include who asserted or holds it. Assertions would link to publications or DCC citations, and those in turn to people in the Talent Knowledge Graph. This intermediate layer would connect people to the assertions their work supports.

Closing the loop

Computational groups often cannot validate their own predictions and must find collaborators who can. A participant observed that publishing lists of computational hypotheses rarely leads to follow-up, because busy researchers are not searching for untested hypotheses. The proposed alternative had three parts:

1. Publish claims, even partially developed ones, in a community setting where researchers can vote on their importance.
2. Record with each claim the remaining knowledge gap and the expertise needed to close it.
3. Use the Talent Knowledge Graph to match claims proactively to investigators with that expertise.

The Talent Knowledge Graph currently holds mostly publication information, so richer expertise metadata would be needed. Participants framed the end state as a feedback loop. Experimental data, literature, expertise, and new validation results would each enrich the others.

Representing behavior over time and space

A participant asked which technologies could capture temporal and spatial relationships in human activity, for example how body parts move together during a task or treatment. The need goes beyond naming a body part or an activity to describing what a person does and how that changes over time.

Participants suggested several starting points: the Physical Activity Ontology (PACO) in NCBO BioPortal¹², human behavior ontologies used in psychology and psychiatry, activity codes in SNOMED CT, and exercise-related resources from MoTrPAC, whose data are already in the DDKG¹. Taylor noted that PACO had been considered for inclusion in the UBKG.

Taylor also described work at CHOP's Center for Autism Research, where video recordings of children are used to classify behaviors over time in a standardized way. Behaviors observed this way are medical observations that the DDKG could in principle represent, though this has not been done. The discussion did not settle on an ontology or architecture, and highlighted a gap between static concept relationships and dynamic behavioral data.

Open questions

The session raised several questions that it did not resolve:

1. Cross-scale inference. How can causal and temporal reasoning operate across molecular, cellular, tissue, and organ levels while preserving context? Recursive anatomical grammars and ontology inheritance each address part of the problem.
2. Exchange versus access. Should CFDE knowledge graphs be exposed through separate MCP servers, through an integrated graph, or both, and who decides?
3. Licensing. Which DDKG content can be offered publicly, and which redistribution arrangements would be needed for the rest?
4. HRA–DDKG interoperability. What exactly is the exchange mechanism between the HRA knowledge graph and the DDKG, and how would ApiNATOMY content travel along it?
5. Evidence content. What must a Portable Assertion Record contain, how should heterogeneous evidence such as gene–disease support be scored, and which existing standards and vocabularies can be reused?
6. Expertise metadata. What information about investigators is needed to match open claims to the people able to test them?
7. Behavioral and temporal data. Which ontologies and structures can represent behavior as it unfolds over time?

Proposed next steps

The steps below were proposed during the session. Most are not formal assignments, and none has an agreed schedule unless stated.

Proposed step	People or groups named	Status at the meeting
Bring ApiNATOMY models into the CFDE through the HRA, starting from a concrete use case such as membrane proteins or anatomy-driven DDKG queries	de Bono, Taylor, Bueckle, Herr, Börner	Agreed in principle
Present HRA work to the working group	Herr, Bueckle	Tentatively planned for the next meeting, expected November 2026
Inventory all CFDE knowledge graphs: capabilities, platforms, specifications, integration options	Working group; Jake	Proposed
Complete the JKG distribution and announce it	DDKG team (Simmons, Silverstein)	About three-quarters complete
Build an MCP server over a DCC-derived DDKG subgraph with licensable supporting ontologies	DDKG team, Knowledge Center	Proposed; licensing to be checked
Integrate the DDKG with CFDE-REVEAL and with other specialized graphs, including protein-related efforts	DDKG team, Knowledge Center	Under discussion
Draft the PAR and PAD specification and coordinate evidence vocabularies with the Ontology Working Group	Taylor; open to all CFDE participants	Early draft shared
Add PubMind and extend evidence levels to all DDKG datasets	DDKG team; Flannick and REVEAL groups on gene–disease evidence	Priority for the coming year
Link assertions and publications to the Talent Knowledge Graph	Jake, DDKG team	Exploratory
Take part in HRA hackathons to build these integrations	Interested members	Encouraged

Conclusion

The session showed a working group shifting from building one integrative graph toward connecting many. Three kinds of knowledge emerged as complementary. Observational knowledge is organized in the DDKG and related CFDE resources. Mechanistic and anatomical context comes from ApiNATOMY, the HRA, and the Whole Person Physiome. Knowledge of who can advance a question could come from the Talent Knowledge Graph, linked to evidence-aware assertions.

The practical obstacles are concrete: licensing limits on open AI access, the lack of a shared way to state the evidence behind an assertion, and the gap between generating hypotheses and getting them tested. The proposals discussed, including JKG, licensed-subset MCP servers, Portable Assertion Records, and expertise matching, each address one of these obstacles. Most remain plans. Their combined aim is a system that connects observations to explanations, states how well supported each claim is, and routes open questions to people who can answer them.

Acknowledgments

Thanks to Deanne Taylor for leading the session, to Bernard de Bono for his presentation, and to all working group participants, in the room and online, whose contributions are summarized here. Any errors of transcription or interpretation are the submitting author's.

References

1. The Data Distillery: A Graph Framework for Semantic Integration and Querying of Biomedical Data. bioRxiv. 2025. doi:10.1101/2025.08.11.666099 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

2. Kokash N, de Bono B. Knowledge Representation for Multi-Scale Physiology Route Modeling. Front Neuroinform. 2021;15:560050. doi:10.3389/fninf.2021.560050 ↩

3. de Bono B, Gillespie T, Surles-Zeigler MC, Kokash N, Grethe JS, Martone M. Representing Normal and Abnormal Physiology as Routes of Flow in ApiNATOMY. Front Physiol. 2022;13:795303. doi:10.3389/fphys.2022.795303 ↩

4. de Bono B, Grenon P, Baldock R, Hunter P. Functional tissue units and their primary tissue motifs in multi-scale physiology. J Biomed Semantics. 2013;4:22. doi:10.1186/2041-1480-4-22 ↩

5. Börner K, Blood PD, Silverstein JC, et al. Human BioMolecular Atlas Program (HuBMAP): 3D Human Reference Atlas construction and usage. Nat Methods. 2025;22. PMID 40082611 ↩

6. Bueckle A, Herr BW, Hardi J, Quardokus EM, Musen MA, Börner K. Construction, Deployment, and Usage of the Human Reference Atlas Knowledge Graph. Sci Data. 2025;12. doi:10.1038/s41597-025-05183-6 ↩ ↩²

7. National Institutes of Health. NIH launches landmark project on whole-person health and function. News release, September 17, 2025. nih.gov ↩

8. Model Context Protocol. modelcontextprotocol.io ↩

9. Morris JH, Soman K, Akbas RE, et al. The scalable precision medicine open knowledge engine (SPOKE): a massive knowledge graph of biomedical information. Bioinformatics. 2023;39(2):btad080. doi:10.1093/bioinformatics/btad080 ↩

10. Groth P, Gibson A, Velterop J. The anatomy of a nanopublication. Inf Serv Use. 2010;30(1–2):51–56. doi:10.3233/ISU-2010-0613 ↩

11. PubMind: Literature-Based Genetic Variant Extraction and Functional Annotation Using Large Language Models. bioRxiv. 2025. doi:10.1101/2025.10.13.682183. Code and database: github.com/WGLab/PubMind ↩

12. Physical Activity Ontology (PACO). NCBO BioPortal. bioportal.bioontology.org/ontologies/PACO ↩