DGLink: Automated knowledge graph construction from biomedical data repositories

While portals supporting user-submitted biomedical data hold substantial scientific value, the re-use of this data at scale is limited by a lack of standardization and annotation. DGLink addresses this gap by automatically assembling portal content into a knowledge graph linking data within a portal and mapping it to external knowledge. This knowledge graph creates a shared interoperability layer that enables knowledge-driven data interpretation and cross-portal discovery. Crucially, the DGLink system not only aggregates existing metadata for projects but also automatically parses and annotates experimental data files directly, greatly increasing annotation depth. For more information on the system please feel free to read the abstract of our talk recently presented at ISMB 2026 or take a look at our GitHub repository.

This page provides access to a DGLink-constructed knowledge graph combining content from multiple National Cancer Institute Cancer Research Data Commons (NCI-CRDC) data portals, namely:

Additionally, we provide summary statistics for the merged NCI-CRDC knowledge graph and example applications including a structured semantic interface. Finally, we explore how DGLink can integrate into an agentic workflow via an MCP-based knowledge-graph server, providing both example MCP queries and benchmarking MCP performance.

Coverage

What was processed

DGLink extracts from the data files a portal makes downloadable, so how much of each portal was processed differs. The General Commons was covered in full at the open-access level; the Genomic and Proteomic Data Commons each only had a few cases/studies processed as a proof of concept.

gc
General Commons
Full open-access coverage

Every open-access study in the NCI General Commons was processed end to end — DGLink extracted content from all of their data files. Controlled-access studies appear as metadata only.

52
studies (total)
2
open-access studies
2
studies with extracted content
gdc
Genomic Data Commons
Proof-of-concept sample

A proof-of-concept subset of studies and cases was processed to demonstrate genomic coverage — not the entire commons.

50,270
cases (total)
150
cases with extracted content
pdc
Proteomic Data Commons
Proof-of-concept sample

A proof-of-concept subset of studies and cases was processed to demonstrate proteomic coverage — not the entire commons.

6,640
cases (total)
41
studies with extracted content
Subgraph of the EGFR gene linked to studies and cases across the GC, GDC and PDC portals.
Cross-portal

One gene, three portals

EGFR — a single gene node — linked to studies and cases drawn from all three CRDC portals. This triangulation across data modalities is exactly what DGLink's merge produces, and what no single portal can show on its own.

  • Gene (EGFR)
  • General Commons (GC)
  • Genomic Data Commons (GDC)
  • Proteomic Data Commons (PDC)
Knowledge graph

Merged CRDC graph (NCI GC · GDC · PDC)

2,748
Genes in all 3 portals
103,120
Nodes
869,897
Edges
16
Entity types
5
Relation types
27,110
Cross-portal nodes

Entities by type · 16 types

Entity type GCGDCPDC Total Cross-portal Share of all nodes
Case 1,366 50,270 6,640 58,276 –
56.5%
Gene 2,862 32,048 23,700 32,476 23,386
31.5%
RNAProduct 11 7,044 4,043 7,699 3,390
7.5%
SmallMolecule 405 20 2,074 2,378 118
2.3%
Disease 167 299 170 545 76
0.5%
Study 64 91 288 443 –
0.4%
Publication 310 – – 310 –
0.3%
NamedThing 160 14 158 288 39
0.3%
MacromolecularComplex 80 22 90 148 39
0.1%
OrganismTaxon 22 – 122 139 5
0.1%
GrossAnatomicalStructure 75 66 48 138 38
0.1%
BiologicalProcess 68 2 44 103 10
0.1%
StudyPopulation 89 – – 89 –
0.1%
Agent 50 – – 50 –
0.0%
Protein 9 8 12 21 8
0.0%
CellularComponent 10 1 7 17 1
0.0%

Portal columns count a node once for each portal it belongs to, so cross-portal entities add to more than one — Cross-portal is how many nodes of that type are shared across two or more portals.

Relationships by type · 5 types

Relationship type GCGDCPDC Total Share of all edges
related_to 4,083 397,679 336,600 738,362
84.9%
has_part 1,510 50,270 19,465 71,245
8.2%
associated_with 1,375 51,406 7,149 59,930
6.9%
mentions 310 – – 310
0.0%
contributes_to 50 – – 50
0.0%

Coverage by portal

gc
General Commons
5,748 nodes · 7,328 edges
gdc
Genomic Data Commons
89,885 nodes · 499,355 edges
pdc
Proteomic Data Commons
37,396 nodes · 363,214 edges

Nodes are counted once per portal they appear in, so cross-portal entities (e.g. a shared gene) contribute to more than one column — that overlap is exactly what merging the portals produces.

Interoperability

Inside the merge

Every node and edge records the portal(s) it came from. When DGLink recognizes that an entity seen in different portals is the same real-world thing, it collapses them into one node whose portal list keeps every origin. Those shared nodes are the bridges that make cross-portal questions answerable.

2,748
genes observed across all three CRDC portals (GC · GDC · PDC) — the same gene independently surfacing in genomic, general and proteomic commons. That triangulation across data modalities is something no single portal can produce.

Extracted vs. exposed metadata, by portal membership

Each row is a set of portals a node belongs to; rows spanning more than one portal are the cross-portal bridges. Extracted = content DGLink pulled from the data files; metadata = what the portals already expose.

Portal membership Extracted Metadata Total Split (% extracted)
GDC 12,443 50,573 63,016
19.7%
GDC + PDCbridge 24,019 4 24,023
100.0%
PDC 3,390 6,943 10,333
32.8%
GC + GDC + PDCbridge 2,799 0 2,799
100.0%
GC 779 1,882 2,661
29.3%
GC + PDCbridge 241 0 241
100.0%
GC + GDCbridge 45 2 47
95.7%
All portals 43,716 59,404 103,120
42.4%
Extracted from files Exposed metadata

The bridge rows are almost entirely extracted content — cross-portal links come from what DGLink mined out of the data files, not from portal-exposed metadata.

Explore

Open the graph in Neo4j Browser

Prefer to write Cypher and see rendered subgraphs? The full knowledge graph is served by Neo4j Browser. Authentication is disabled for this deployment, so no login is required just enter bolt+s://dglink.indra.bio:7676 in the connection URL box and click Connect.

Open Neo4j Browser ↗

Paste this to render the cross-portal gene set:

// EGFR linked to the first 2 studies/cases from each CRDC portal (GC + GDC + PDC)
MATCH (a)-[r:`biolink:related_to`]->(g:`biolink:Gene` {name: 'EGFR'})
WITH a, g, r, a.portal[0] AS portal
WITH portal, collect(r)[0..2] AS rels
UNWIND rels AS r
RETURN startNode(r) AS a, r, endNode(r) AS g
Application

Ask the CRDC knowledge graph questions directly

The DGLink Model Context Protocol (MCP) server lets you put a question to the merged CRDC knowledge graph in plain English. It gives LLM agents direct access to the graph, so a single question can reason across the GDC, GC, and PDC — with no Cypher to write.

  • MCP integration: direct, tool-based access to Neo4j via the mcp-neo4j-cypher server.
  • Model-agnostic: works with any MCP-capable client (Claude Desktop, Cursor, or a custom agent).
  • Cross-portal reasoning: a single question can span GDC, GC, and PDC data.
Try it

Example MCP queries

Click a question to read the agent’s response. Connect your own LLM to the MCP server (below) to run them yourself.

How many genes are observed across all three CRDC portals (GDC, GC, and PDC) at once — and are landmark cancer genes like TP53, EGFR, and PTEN among them?
Agent response

claude-sonnet-5 · 7 agent turns · 1,058 output tokens · 30.4 s · $0.19

Trace TP53 across the three portals: which studies and cases reference it, and in what role, based on the source data-file column each mention was extracted from?
Agent response

claude-sonnet-5 · 8 agent turns · 3,591 output tokens · 59.3 s · $0.28

Which genes are called copy-number drivers in PDC proteogenomic studies and are also present in the Genomic Data Commons? Name a few, with the studies they came from.
Agent response

claude-sonnet-5 · 11 agent turns · 4,866 output tokens · 88.4 s · $0.38

For glioblastoma, how many GDC genomic cases and PDC proteomic studies converge on the same disease concept?
Agent response

claude-sonnet-5 · 11 agent turns · 2,019 output tokens · 45.5 s · $0.13

Which cancer genes appear in BOTH a General Commons (GC) study's data files AND a PDC proteomic study? List a few of the shared genes.
Agent response

claude-sonnet-5 · 8 agent turns · 2,781 output tokens · 54.6 s · $0.28

Which genes have acetylation or glycosylation site data in PDC data files AND are also called copy-number drivers? List a few, with the columns the evidence came from.
Agent response

claude-sonnet-5 · 9 agent turns · 4,002 output tokens · 60.2 s · $0.33

For TP53, show the exact source data file and column that each mention was extracted from, per portal.
Agent response

claude-sonnet-5 · 8 agent turns · 2,904 output tokens · 55.9 s · $0.20

Connect

Add the MCP server to your own LLM

DGLink serves the merged graph over a streamable-HTTP MCP endpoint, so any MCP-capable client — Claude Desktop, Claude Code, Cursor, or a custom agent — can connect with just a URL. No authentication is required, and there is nothing to install locally: no uv, no Python, no Neo4j driver.

For example, to add it to Claude Desktop, merge the following into claude_desktop_config.json:

{
  "mcpServers": {
    "dglink-crdc": {
      "url": "https://dglink.indra.bio/crdc/mcp/"
    }
  }
}

Or from the command line with Claude Code:

claude mcp add --transport http dglink-crdc https://dglink.indra.bio/crdc/mcp/

Your client then gets tools to inspect the graph schema and run read Cypher against it. The endpoint is read-only, so queries can explore the graph but never modify it. Older clients without native remote-MCP support can bridge to the same URL with npx mcp-remote https://dglink.indra.bio/crdc/mcp/.

Benchmark

Does the knowledge graph help? Raw API vs. knowledge graph agent

One immediate follow-up question to this type of approach might be, “Is the knowledge graph actually necessary, or could a state-of-the-art LLM get the same result if just given access to a portal’s API?” We can test exactly this by giving the same agent the same question twice: once with only API access, and once with access to the knowledge graph. Everything else is held constant — same model, system prompt, and turn budget — so the only variable is the MCP toolset.

All trials were run using Claude Sonnet 5 headless via claude -p, with the built-in file and shell tools disabled so an agent’s only capability is its own MCP server.

  • Raw-API agent: Agent was given access to each portal's GraphQL client as well as tools for previewing and downloading portal data files.
  • Knowledge graph (KG) agent (DGLink): Agent was given direct MCP access to the merged CRDC KG.

Both agents can reach the same files — the comparison is pre-indexed graph vs. search-at-query-time, not “can read files” vs. “can’t”.

Query 1 · cross-portal

Is TP53 present in all three CRDC portals?

“Is the gene TP53 present in all three CRDC portals in the ingested data — a GDC case, a GC (General Commons) study, and a PDC proteomic study? Name the specific study/case in each portal.”

Raw-API agent
✕2 of 3 portals confirmed

Confirmed TP53 in GDC and PDC, but could not verify the GC link — no gene-level field exists in the GC metadata API.

DGLink KG agent
✓3 of 3 portals confirmed

One merged gene node named the specific GDC case, GC study and PDC study in a single traversal.

Effort to get there

Metric Raw-API agent DGLink KG agent Relative DGLink advantage
Agent turns 26 7
3.7× fewer
Output tokens 11,739 1,573
7.5× fewer
Wall time 169.6 s 31.1 s
5.5× faster
Cost (USD) $0.63 $0.17
3.7× cheaper
Raw-API agent DGLink KG agent Shorter teal bar = less effort for the same question.
Raw-API agent response
DGLink KG-agent response
Query 2 · cross-portal

Which cancer genes appear in both caNanoLab and a PDC proteomic study?

“Which cancer genes appear in BOTH the caNanoLab (GC) study data files AND a PDC proteomic study? List a few shared genes.”

Raw-API agent
7 shared genes surfaced

Limited to text search over file names, it could only spot-check a handful of gene symbols — not compute the intersection.

DGLink KG agent
2,586 shared genes, exactly

Computed the full overlap — 2,586 of 2,652 caNanoLab-linked genes also appear in a PDC study — as one graph traversal.

Effort to get there

Metric Raw-API agent DGLink KG agent Relative DGLink advantage
Agent turns 44 9
4.9× fewer
Output tokens 11,149 2,963
3.8× fewer
Wall time 166.4 s 48.4 s
3.4× faster
Cost (USD) $1.15 $0.27
4.3× cheaper
Raw-API agent DGLink KG agent A set intersection across portals is one traversal in the graph, but expensive over raw APIs.
Raw-API agent response
DGLink KG-agent response
Gene panel · scored vs. independent gold

Which of a 20-gene panel appear in caNanoLab data files?

“For each gene in this 20-gene panel, determine whether it appears in any data file of the caNanoLab study (NCI General Commons, phs_accession 10.17917).”

Unlike the open-ended queries above, this task is scored automatically against an independent gold standard: 14 of the 20 panel genes are truly present. Neither agent produced a false positive, so the whole difference is in recall — how many of the 14 each one actually found.

Raw-API agent
4 of 14 genes found

Recall 0.29 · F1 0.44. Limited to filename/metadata text search, it cannot see genes buried inside the study’s data tables.

DGLink KG agent
12 of 14 genes found

Recall 0.86 · F1 0.92. File-content gene mentions are pre-indexed on the graph, so it reads them straight off the study node.

Scored against the gold standard

Metric Raw-API agent DGLink KG agent Relative DGLink advantage
Precision 1.00 1.00
tie
Recall 0.29 0.86
3.0× higher
F1 0.44 0.92
2.1× higher
Accuracy 0.50 0.90
1.8× higher
True positives 4 / 14 12 / 14
3.0× more
Raw-API agent DGLink KG agent Longer teal bar = better score (higher is better for every metric above).

Gold-positive: 14/20. KG agent missed IDH1, TP53. Raw-API agent missed BRCA2, IDH1, IGF1R, MAPK3, MTOR, PTEN, RB1, SMAD4, TP53, VHL.

Effort to get there

Metric Raw-API agent DGLink KG agent Relative DGLink advantage
Agent turns 51 12
4.3× fewer
Output tokens 18,575 3,575
5.2× fewer
Wall time 222.2 s 71.3 s
3.1× faster
Cost (USD) $1.37 $0.19
7.2× cheaper
Raw-API agent DGLink KG agent Shorter teal bar = less effort — here, for 3× the recall.
Raw-API agent response
DGLink KG-agent response