henhaobio
Sign in

bioRxivBioinformaticsPreprint

KRAKEN: A provenance-tracked knowledge graph for multiomic and wellness research

Amy K. GleniD, Drew Witherington, Trent LeslieiD, Andrew M. BaumgartneriD, Bharana FernandoiD, Bhargav VemuriiD, Ornit NahmaniD, Gustavo GlusmaniD, Leroy HoodiD, Lance T. PfliegeriD, Noa RappaportiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Existing general-purpose biomedical knowledge graphs tend to focus on disease mechanisms and drug repurposing, leaving multiomic and wellness-relevant content underrepresented. KRAKEN (Knowledge Research & Analysis Kit for Evidence Networks) addresses this gap by integrating existing graphs (including Translator KG Open, RTX-KG2, and ROBOKOP) with specialized sources such as RefMet, LIPID MAPS, NIH Common Data Elements, Polygenic Score Catalog, and derived wellness measures including biological age and biological BMI. The resulting graph spans ~26M nodes and ~121M edges across 62 entity types. KRAKEN adopts the Biolink Model as its semantic layer, ensuring compatibility with standardized resources emerging from the NIH NCATS Biomedical Data Translator program. A lightweight, modular build system rebuilds the full graph (including entity resolution), with peak memory consumption under 48 GB, and supports flexible inclusion or exclusion of sources, allowing the user to scope the graph to a domain of interest. Built-in analytical tools include multi-hop reasoning, subgraph extraction, text, vector and hybrid entity search, and enrichment analyses, all accessible through an interactive web interface, a REST API, and a Model Context Protocol server, the last enabling direct consumption by agentic and LLM-based systems. KRAKEN is freely available at https://app.krakenkg.com.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (9 Oct 2026) · CC BY 4.0

A Biolink-compliant graph (∼15 million nodes, ∼113 million edges) adds RefMet and LIPID MAPS metabolite and lipid classifications, plus polygenic score and clinical data element entities, to Translator content.

Deep phenotyping cohorts yield mixed omic, polygenic and questionnaire features, yet existing knowledge graphs link metabolites and lipids mostly via drug and disease edges, rarely treat scores, questionnaires or derived measures as entities, and expose query endpoints specified before agents were a design target. KRAKEN adds RefMet, LIPID MAPS, polygenic score and data element content under Biolink, keeps source predicates and provenance on edges, and serves agent-oriented tools.

Key results

  1. ShownRelease v2.1.0 holds ∼15 million nodes and ∼113 million edges over 62 Biolink node types and 114 edge types from over 100 primary knowledge sources, and most schema-level links join different entity families.

    15,099 distinct meta-triples, 82% crossing node type families; edges mostly computational predictions (∼53 million) or manual curation/validation (∼31 million), plus statistical association, text-mining and logical-entailment edges; Multiomics KG, HMDB, GO-Plus, DISEASES and Uberon act as stronger cross-domain bridges between sources. · Content and Scale; Data Sources

  2. ShownNode search mapped most platform-reported labels in a biological body mass index (bBMI) signature to canonical entities with little manual curation; nearly all failures were unnamed metabolite spectral features.

    482 of 543 analytes (356 metabolites, 149 proteins, 38 clinical measures) resolved via text search, with hybrid text plus vector search as a fallback for synonyms; of the 61 unresolved, the one named label was the composite clinical lab measure OMEGA-3 INDEX. · Mechanistic interpretation of multiomic…; Analytical Tools

  3. SuggestedDegree-corrected permutation enrichment over the signature's proteins converged on inflammatory signaling pathways.

    Using the 88 more robustly selected signature proteins, 25 pathways were significant (false discovery rate < 0.05); named examples are leukocyte migration, chemokine-mediated signaling and immune-response regulation. Empirical p-values come from degree-matched random node sets with Benjamini-Hochberg correction. · Mechanistic interpretation of multiomic…; Enrichment analysis

  4. SuggestedSubgraph extraction around the resolved signature added connecting genes and proteins that grouped it into inflammatory, endocrine and matrix/vascular modules, and reached analytes the sparse regression model had dropped.

    Excluding disease, anatomical and information-content connectors gave 660 nodes and 18,774 edges (3 cell types, 6 clinical findings, 30 processes, 197 genes/proteins, 245 drug-annotated chemicals and 179 chemicals lacking drug annotation); added hubs include IL6, TNF, NFKB1, INS, IGF1, PPARG, TGFB1 and MMP9; twelve entities were measured analytes discarded in most training folds. · Mechanistic interpretation of multiomic…

  5. SuggestedTraversal from asthma polygenic-risk-correlated plasma proteins to compounds, with text-mined edges excluded, returned mepolizumab as a positive control and marimastat, an oncology matrix metalloproteinase inhibitor, as a repurposing candidate.

    Plasma proteins correlated with asthma polygenic risk in UK Biobank (text notes 2,932 proteins) were linked to compounds, then optionally to approval, use or trials in asthma-related conditions; a later literature search found a double-blind cross-over pilot in twelve atopic asthmatics where marimastat lowered bronchial hyper-responsiveness to allergen, leaving FEV1 and symptoms unchanged. · From polygenic risk to molecular intermediates…

Methods, data and reproducibility

A modular Python pipeline maps each source (RTX-KG2, ROBOKOP, Translator KG Open, RefMet, LIPID MAPS, PGS Catalog, NIH Common Data Elements, others) to Biolink-compliant newline-delimited JSON, merges equivalent identifiers into canonical nodes while keeping source predicates and provenance, and serves the graph via the KESTREL engine with search, multi-hop, subgraph and degree-corrected enrichment tools.

  • Directly ingested knowledge sources (Table 1): RTX-KG2, ROBOKOP, Translator KG Open, RefMet, LIPID MAPS Structure Database, PGS Catalog, among others · n = over 100 primary knowledge sources (including via aggregator graphs)
  • Analytes associated with a combined bBMI measure from a deep phenotyping source study (study not named) · n = 543 analytes: 356 metabolites, 149 proteins, 38 clinical measures
  • Plasma proteins whose levels correlate with polygenic risk for asthma · human (UK Biobank cohort) · n = 2,932 proteins (bracketed note in text; scope unclear)
Methods & tools
KESTREL, Biolink Model, BioMapper (biomapper2), SRI Node Normalizer, all-MiniLM-L6-v2, Model Context Protocol (Streamable HTTP), RTX-KG2, ROBOKOP, Translator KG Open, RefMet, LIPID MAPS Structure Database, PGS Catalog
Assay conditions & compute
Full build under 3 hours on a MacBook Pro with Apple M4 chip, with peak memory kept under 48 GB by file streaming; source download time excluded.
Reproducibility gaps
No bulk data download; content is reachable only through the hosted interfaces under each source's licence. Code for the KESTREL query engine is not linked, and its query and ranking methods are to be reported separately.
Limitations (6)
  • KRAKEN inherits the errors of its sources and does not itself validate any assertion. · stated by the authors
  • Coverage across the Biolink hierarchy is uneven, some categories come from a single source, and the wellness content covers only part of what deep phenotyping studies collect. · stated by the authors
  • Most canonical merges reuse Node Normalizer equivalences inherited from aggregator graphs; metabolite vocabularies split compound classes and members inconsistently, and the graph does not show which source a merge relied on. · stated by the authors
  • Semantic search embeddings come from names and synonyms, so they reflect nomenclature rather than function; adding a source is semi-manual, needing a parser and a mapping decision. · stated by the authors
  • Both use cases are single illustrative runs with no comparison to other knowledge graphs or enrichment tools, and the correctness of the 482 label mappings is not reported. · noted in this summary
  • The marimastat lead rests on one twelve-subject pilot found by a post hoc literature search, and only two compounds from the asthma traversal are shown, with no count of the full returned set. · noted in this summary

Read this if you need to map mixed metabolomic, proteomic and clinical feature lists onto canonical entities, or want a provenance-carrying biomedical knowledge graph that language-model agents can query through a Model Context Protocol server.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISZhang '26Mohamed '26Zhang '25Mathé '26Ahooyi '25Geleta '21
65% semantic matchBioinformatics
Heming Zhang, S. Liang, Tim Xu+16 more·Bioinformatics·Jun 1, 2026· 2 citations