KG-TransomicNet — materialised database

ArangoDB dump of the semantic–quantitative property graph described in Modeling Omics with Semantics for Dynamic Construction of Knowledge-Based Trans-Omic Networks.

This is the materialised instance: restoring it reproduces the results and performance figures reported in the paper without rebuilding the corpus from UCSC Xena and the GDC (a ~4 hour pipeline).

• Software and build pipeline: https://github.com/johndef64/KG-TransomicNet
• Archived release: https://doi.org/10.5281/zenodo.21629419

Contents

One dump directory, KG-TransomicNet-v1.1.1_full/, holding the ontology backbone and all five quantitative layers: 12,010,793 documents, 5.74 GB compressed, 11.4 GB restored.

Collection	Documents		Collection	Documents
nodes	780,753		GENES	61,855
edges	11,082,103		SAMPLES	19,287
GENE_EXPRESSION	15,475		CASES	16,086
CNV	10,665		PROJECTS	42
MIRNA	13,441		PROTEIN	7,936
METHYLATION	3,150

The quantitative layers join to nodes through entity_id, so they are not meaningful without the backbone; the backbone alone is a usable ontology-grounded knowledge graph.

PKT/ holds that semantic backbone on its own, as standalone node and edge tables (nodes.zip, edges.zip, 289 MB) with simple_pandas_lookup.py to load and query them. Use it to work on the ontology graph in pandas without running ArangoDB; use the dump above whenever the quantitative layers are needed.

Requirements

ArangoDB ≥ 3.11 (this dump was produced with 3.11.8) and ~12 GB of free disk for a full restore.

Download and restore

pip install huggingface_hub python-arango

python scripts/db_dump.py download --out ./kg-dump
python scripts/db_dump.py restore --input ./kg-dump --db PKT_main --create

scripts/db_dump.py ships with the software repository. Both commands count documents and refuse to report success on a mismatch, checking against kg_transomicnet_manifest.json inside the dump.

This matters: a dump of this database can fail silently. Dumping the nine quantitative collections concurrently (arangodump's default of 8 threads) drops documents from GENE_EXPRESSION while exiting 0, leaving intact gzip streams and a valid last line. Restoring such a dump yields a database that looks healthy and quietly under-reports the transcriptomic layer. If you re-dump this database yourself, use --threads 1 and verify by counting.

Plain arangorestore also works if you prefer, but then nothing verifies the result:

arangorestore --server.database PKT_main \
              --input-directory KG-TransomicNet-v1.1.1_full

PKT_main is the database name the repository scripts expect by default; any name works if you pass it to them with --db.

What is in the graph

Semantic backbone — PheKnowLator v3.0.2, instance-based OWL-NETS build (Zenodo 10689968), derived from the OBO Foundry and the Relation Ontology: 780,753 entities and 11,082,103 typed relations. Each node carries its source identifier in entity_id, typed by class_code (EntrezID, UniProtKB, MONDO, HP, dbSNP); that pair is the join key against the quantitative layers.

Quantitative layers — five omic modalities over 16,938 distinct samples across 42 TCGA and TARGET projects, at native precision:

Layer	Platform	Projects	Samples	Features/cohort
Transcriptomics	STAR TPM	42	15,433	60,660
CNV (gene-level)	ASCAT3	33	10,632	60,623
miRNA	miRNA-Seq	38	13,403	1,881
Proteomics	RPPA (TCPA)	32	7,904	487
Methylation	Illumina HM27	13	3,137	27,578

Layer availability is uneven: 11 projects carry all five layers, 22 carry four, 8 carry two or fewer. Together the layers hold 1.70 × 10⁹ measurements in 50,509 per-sample vector documents plus 158 cohort index documents.

Each layer follows a vector + index layout: one *_index document per cohort maps array positions to feature identifiers, and one *_vector document per sample stores the dense array. Retrieval dereferences the index once, then slices vectors by position.

Provenance and terms

Quantitative data were downloaded from UCSC Xena and the NCI Genomic Data Commons in 2026, and are redistributed here in reorganised form; they are TCGA/TARGET open-access molecular data. Normalisation was inherited from the upstream providers and not re-applied. Users remain subject to the NIH Genomic Data Sharing policy and the GDC data-use terms, and should cite TCGA/TARGET and PheKnowLator alongside this resource.

Citation

De Filippis, G. M., Rinaldi, A. M. KG-TransomicNet: Semantic–Quantitative Property Graph of PheKnowLator and Multi-Omics, Zenodo, https://doi.org/10.5281/zenodo.21629419.

License

MIT (software and mappings). Underlying TCGA/TARGET data remain subject to their original terms.