SkyeAv/multiomics-kp

Multiomics Knowledge Provider (MOKG)

A KGX knowledge graph built from supplementary tables of open-access PubMed Central articles in the rare-disease cohort. Tablassert maps configured table records to Biolink assertions and resolves entities against BABEL fullmap. Primary knowledge source: infores:multiomics-kp.

Latest build: 3.1.3 - 222,994 nodes, 18,206,965 edges. Built with Tablassert 22.1.0, validated against Biolink Model 4.4.6. This repository remains private.

Releases

main contains only the current build, matching the version-directory layout used by DAKP. The Dataset Viewer configs are 3.1.3_edges (default) and 3.1.3_nodes, each with one train split. The release directory contains exactly the NDJSON pair and Resource Ingest Guide:

3.1.3/
  MULTIOMICS_KP_3.1.3.edges.ndjson
  MULTIOMICS_KP_3.1.3.nodes.ndjson
  MULTIOMICS_KP_3.1.3.RIG.yaml

The previous repository snapshot remains available in repository history at revision 232ff2be936ea7323a70d4d16424854591f8c3f4. No release tag is asserted for this build.

Checksums and sizes

File	Bytes	SHA-256	Records
MULTIOMICS_KP_3.1.3.nodes.ndjson	31,304,275	67c11a2de63c8d4374946f3af64770bb502ceda56866911b0227f356846905f2	222994
MULTIOMICS_KP_3.1.3.edges.ndjson	21,031,119,505	148a73057d8f9ffe72d2bc68d441eb514394e6a5dceb923cb8f36215f829f45c	18206965
MULTIOMICS_KP_3.1.3.RIG.yaml	48,032	91831f8614c098f8000515eca00881042d127f6f47cb1247f5c90fa0acce61d4	Resource Ingest Guide

Validation and limitations

The complete NDJSON pair passed tablassert validate-kgx: 222,994/222,994 nodes and 18,206,965/18,206,965 edges valid, with 0 validation failures. All emitted infores CURIEs were registered. This verifies the validator's structural and Biolink checks, not the scientific correctness of every source mapping.

This build carries forward the 3.1.2 quality-control round: 190 flagged mis-resolved entities removed through source-config fixes (literal blanking, category constraints, exclusion backstops), mis-modeled relations (membership, derivation, geography, drug-mirror, and wrong-predicate tables) disabled in their owning configs, and publication prunes orphan nodes. Node identifiers are unique and no flagged entity or forbidden predicate remains in the published files.

Changes specific to 3.1.3: the biolink:in_taxon and biolink:related_to predicates are excluded entirely - 17 table configs were disabled in the graph configuration before the rebuild (16 whose sections emitted the default related_to predicate and one declaring in_taxon), so neither predicate appears in the published edges. The build also moved to Tablassert 22.1.0, which folds in upstream build fixes (failed dedup runs no longer destroy the previously published graph, and generic header labels are blocked from entity resolution in their normalized forms).

Scientific audit remains incomplete. The checks above are enumerated rules, not an exhaustive semantic review; no fresh domain-and-range audit was re-run for this build. Known remaining conventions: statistical significance qualifiers use bare enum tokens (significant, suggestive, ...) per Translator practice, not biolink: prefixed forms; treats-style assertions are modeled as biolink:applied_to_treat without approval evidence. Treat this as a private working build, not a scientifically finalized release or a clinical decision resource.

Schema

Nodes: id, name, category, provided_by, and optional in_taxon. Category counts below count category memberships.

Category	Count
biolink:Gene	139,368
biolink:Protein	37,125
biolink:SmallMolecule	19,038
biolink:Disease	11,640
biolink:BiologicalProcess	5,789
biolink:ChemicalEntity	4,618
biolink:PhenotypicFeature	1,500
biolink:MolecularActivity	919
biolink:OrganismTaxon	704
biolink:MolecularMixture	466
biolink:CellularComponent	342
biolink:GeneFamily	280
biolink:Pathway	219
biolink:Phenomenon	199
biolink:Drug	147
biolink:Cell	116
biolink:GrossAnatomicalStructure	100
biolink:Procedure	96
biolink:NamedThing	95
biolink:AnatomicalEntity	87
biolink:InformationContentEntity	75
biolink:PopulationOfIndividualOrganisms	19
biolink:Food	13
biolink:MacromolecularComplex	10
biolink:Behavior	9
biolink:ClinicalFinding	5
biolink:BiologicalEntity	4
biolink:NucleicAcidEntity	3
biolink:Cohort	3
biolink:Activity	2
biolink:ClinicalIntervention	1
biolink:PhysiologicalProcess	1
biolink:Device	1

Edges: subject, original_subject, object, original_object, predicate, category, knowledge_level, agent_type, sources, publications, has_supporting_studies, supporting_text, id, effect_size, effect_type, p_value, statistical_significance_qualifier, adjusted_p_value, subject_aspect_qualifier, subject_direction_qualifier, has_evidence, object_direction_qualifier, anatomical_context_qualifier, population_context_qualifier. Optional evidence and qualifier fields are populated only where emitted by source configs. p_value and adjusted_p_value are strings in this build; inequality notation is preserved. has_supporting_studies is a JSON object keyed by publication CURIE, not a fixed struct. sources is a list of provenance records with resource_id, resource_role, and optional upstream_resource_ids and source_record_urls lists.

Predicate	Count
biolink:associated_with	15,111,427
biolink:affects	1,476,867
biolink:binds	401,948
biolink:interacts_with	380,051
biolink:actively_involved_in	235,970
biolink:participates_in	180,970
biolink:genetically_interacts_with	134,736
biolink:regulates	78,975
biolink:directly_physically_interacts_with	73,402
biolink:contributes_to	57,913
biolink:correlated_with	18,816
biolink:causes	17,019
biolink:expressed_in	12,096
biolink:has_gene_product	10,762
biolink:genetically_associated_with	3,907
biolink:has_phenotype	3,218
biolink:gene_associated_with_condition	2,830
biolink:associated_with_likelihood_of	2,364
biolink:expresses	1,415
biolink:affected_by	1,055
biolink:physically_interacts_with	390
biolink:disease_has_location	334
biolink:associated_with_response_to	190
biolink:applied_to_treat	176
biolink:response_associated_with	54
biolink:affects_likelihood_of	51
biolink:associated_with_increased_likelihood_of	13
biolink:gene_fusion_with	10
biolink:coexpressed_with	6

Edge IDs are whole-record UUIDs under uuid_domain: infores:multiomics-kp. Exact duplicate records are dropped; distinct records are not merged or folded. The RIG summarizes observed endpoint categories, predicates, qualifiers, evidence, source scope, exclusions, and pipeline attribution. Statistical association and expression changes do not by themselves establish causation. Gene-set significance must not be interpreted as individual gene-term significance.

Loading

Access requires a Hugging Face account authorized for this private dataset. Authenticate locally; do not put access tokens into scripts, dataset cards, or shared logs.

hf auth login
hf download SkyeAv/multiomics-kp --repo-type dataset --include "3.1.3/*" --local-dir multiomics-kp
tablassert validate-kgx \
  multiomics-kp/3.1.3/MULTIOMICS_KP_3.1.3.nodes.ndjson \
  multiomics-kp/3.1.3/MULTIOMICS_KP_3.1.3.edges.ndjson

Use a recent datasets version supporting the Json feature (tested with 5.1.0). The card declares every observed field explicitly. Publication-keyed supporting-study objects use dtype: json to avoid thousands of sparse publication-specific columns. Numeric JSON fields such as effect_size are rounded to single precision by the datasets NDJSON parser regardless of the declared float64; read the NDJSON directly (pyarrow, polars, jq) when exact double-precision values matter. The KGX NDJSON files themselves are unchanged. Streaming avoids materializing the approximately 21 GB edge file in memory:

from datasets import load_dataset

edges = load_dataset(
    "SkyeAv/multiomics-kp", "3.1.3_edges", split="train", token=True, streaming=True
)
nodes = load_dataset(
    "SkyeAv/multiomics-kp", "3.1.3_nodes", split="train", token=True, streaming=True
)
first_edge = next(iter(edges))

For reproducible access, set revision to the uploaded commit SHA shown in repository history. Dataset Viewer processing is asynchronous. Private viewing requires an eligible PRO account or Team/Enterprise organization. For this non-Parquet edge file, Hugging Face limits automatic conversion to the first 5 GB, so the viewer can be partial. Download access is independent of viewer processing.

Sources and provenance

• PubMed Central Open Access subset supplementary tables, with per-publication citations in publications and study metadata in has_supporting_studies.
• Deterministic supplementary-table extraction and upstream staging by PMCProcessing.
• Entity resolution through the BABEL fullmap database.
• KGX generation and validation by Tablassert 22.1.0.

The upstream ingest gates articles to CC0, CC BY, or CC BY-SA according to extraction ledgers. See the RIG for access locations and scope. Not all upstream ledgers or table configs are included in this artifact-only repository.

License

Copyright 2026 Skye Lane Goetz. The MULTIOMICS_KP build artifacts in this repository are licensed under Apache-2.0: https://www.apache.org/licenses/LICENSE-2.0

Upstream content terms still apply to the data they describe. Source publications retain their individual CC0, CC BY, or CC BY-SA terms, including attribution and share-alike requirements where applicable; consult each publication and its extraction ledger before reuse of the underlying content.

• PMC copyright and reuse
• PMC Open Access subset

Links

• Tablassert
• MultiomicsKG publication
• Resource Ingest Guide