Knowledge Graph Catalog

Purpose: Scholarly + comparative context for the knowledge graphs we use — paper citations, published baselines, domain-fit rationale, known issues, and pretraining background. This file is not the source of truth for live state.

Live state — current load status, node/edge counts, license, version, and composition membership — lives in the mods_kg_source and mods_kg_composition_* tables. Query via the mods_kg_list_sources and mods_kg_list_compositions MCP tools. Always defer to the DB for exact counts and current status; the snapshots in this file are illustrative.

Updated: 2026-05-29

Live state quick reference (snapshot 2026-05-29)

Run mods_kg_list_sources for live values.

Source ID	Status	Approx scale (ent / rel)	Section
`gp_kg`	loaded	61K / 1.24M	GP-KG
`prime_kg`	loaded	129K / 8.0M	PrimeKG
`civic`	loaded	6.0K / 11.5K	CIViC
`clinvar`	loaded	933 / 1.6K	ClinVar
`oncokb`	loaded	176 / 483 (1,221 / 7,217 on 2026-10-05, after #744)	OncoKB
`pharmgkb_variants`	loaded	5.0K / 9.4K	PharmGKB CPIC
`maple_patient`	loaded	5 / 339	Maple Patient Overlay
`rtx_kg2`	partial	3.54M / 10.2M	RTX-KG2
`drkg`	running	(ingest in progress)	DRKG
`dbsnp`	cross_ref_only	2.7K / 0	Variant cross-ref services
`gnomad`	cross_ref_only	30 / 0	Variant cross-ref services
`myvariant_aggregator`	cross_ref_only	116 / 0	Variant cross-ref services
`translator_ncats`	api_only	—	NCATS Translator
`txgnn`	api_only	—	TxGNN

Status vocabulary (from mods_kg_source.status):

loaded — Fully ingested, usable in compositions
running — Ingest in progress
partial — Some data loaded, raw data on disk, full load not done. RTX-KG2 is partial; see GH #85.
cross_ref_only — Identifiers loaded for joining; no edges contributed
api_only — Live API; no local copy
planned — Identified, not started

Loaded KGs

GP-KG (Genotypic-Phenotypic Knowledge Graph)

Attribute	Value
Paper	Gao Z, Ding P, Xu R. “KG-Predict: A knowledge graph computational framework for drug repurposing.” J Biomed Inform 2022
DOI	10.1016/j.jbi.2022.104133
Data URL	http://nlp.case.edu/public/data/GPKG-Predict/
Entity types (7)	drug, gene, disease, mammalian_phenotype, human_phenotype, go_term, tissue
Relation types (9)	drug-target-gene, disease-associate-gene, drug-treat-disease, gene-relate-MP, gene-relate-GO, gene-relate-UT, drug-relate-HP, drug-relate-MP, disease-relate-HP
Upstream sources	DrugBank, MGI, HPO, GTEx, GOA, Phenomebrowser, TreatKB (NLP from FAERS, FDA labels, MEDLINE, clinical trials)
Algorithm in original paper	CompGCN (embeddings) + ConvE (link prediction)
DB row	`mods_kg_source.id='gp_kg'` (status: `loaded`)
Connector	`backend/connectors/gpkg_connector.py`
Domain fit	Primary — this is the KG used by KG-Predict and the PDR papers we replicate. Drug-disease focus with strong phenotype coverage.

Published metrics

Source	Metric	Value	Task
KG-Predict (Gao 2022) CV	AUROC	0.981	10-fold CV on GP-KG, full head+tail link prediction
KG-Predict (Gao 2022) CV	AUPR	0.409	Same
KG-Predict (Gao 2022) CV	MRR	0.261	Same
KG-Predict (Gao 2022) AD case study	AUROC	0.868	Drug ranking vs FDA-approved + active-trial AD drugs as positives
KG-Predict (Gao 2022) AD case study	AUPR	0.364	Same
PDR1 (Oğuztüzün 2025) Table 3	MR / MRR / Hits@1 / Hits@3 / Hits@5 / Hits@10 / AUC	1390.6 / 0.19231 / 0.13126 / 0.20721 / 0.24959 / 0.31331 / 0.866	5-fold CV foundational model; per-patient models comparable (mean AUC 0.856)

Paper taxonomy (avoids the KG-Predict / PDR1 / PDR2 mix-up)

Three distinct papers reuse the same GP-KG + CompGCN+ConvE stack. Different numbers, different tasks — don’t cross-compare:

Key	Paper	Year	Authors	Published metrics’ task
KG-Predict	”KG-Predict: A knowledge graph computational framework for drug repurposing”	2022	Gao / Srivastava	Introduced GP-KG. Reports CV numbers (0.981/0.409/0.261) and an AD case study (0.868/0.364).
PDR1	”Precision Drug Repurposing (PDR): Patient-level modeling and prediction…“	2025	Oğuztüzün et al.	Reuses KG-Predict’s foundational model, adds per-patient biobank data. Table 3 foundational AUC 0.866, MRR 0.192.
PDR2 / PDR-II	”Learning to take it personally…” (DOI 10.1016/j.jbi.2026.105039)	2026	Oğuztüzün, Gao, Li, Qi, Koyutürk, Xu	Code + PDF in `docs/References/PDR2/`. Same GP-KG, new backbone (ULTRA + NBFNet). Personalization in the loss function (learnable disease/protein weights × PRS/biomarker). Reports 1.3×–5.4× AUPRC improvement per patient vs PDR1.

Which comparator applies where:

Our measurement	Right comparator
`compute_metrics.py` AUROC/AUPR on AD drug ranking vs 60 NCT positives	KG-Predict AD-CS AUROC 0.868, AUPR 0.364
Val MRR during training (with `-skip_head`)	Not directly comparable — tail-only eval roughly doubles MRR vs head+tail. For fair comparison to KG-Predict CV MRR 0.261, must re-run without `-skip_head` (Run 4).
Test-set MRR/Hits@k/AUROC without `-skip_head` (Run 4)	KG-Predict CV AUROC 0.981, MRR 0.261
Per-patient drug ranking vs prescription history	PDR1 Table 3 per-patient AUC 0.856, MRR 0.192 — not attempted; we have no UK Biobank data.

PrimeKG (Precision Medicine Knowledge Graph)

Attribute	Value
Paper	Chandak P, Huang K, Zitnik M. “Building a knowledge graph to enable precision medicine.” Sci Data 2023
DOI	10.1038/s41597-023-01960-3
Data URL	Harvard Dataverse DOI:10.7910/DVN/IXA7BM
Entity types (10)	drug, disease, gene/protein, effect/phenotype, biological_process, molecular_function, cellular_component, anatomy, pathway, exposure
Relation types (30)	indication, contraindication, off-label use, target, enzyme, carrier, transporter, side_effect, phenotype_protein, disease_protein, protein_protein, etc.
Upstream sources	20+ resources including DrugBank, OMIM, DisGeNET, HPO, GO, Uberon, Reactome, SIDER, PharmGKB
Supplementary files	`disease_features.tab` (MONDO, Orphanet, Mayo descriptions), `drug_features.tab` (MoA, ATC codes, pharmacokinetics)
DB row	`mods_kg_source.id='prime_kg'` (status: `loaded`)
Connector	`backend/connectors/prime_kg_connector.py`
Known issues	64 node collisions from temporal name ambiguity — see `docs/primekg_anomaly_report.md`. Report sent to authors.
Domain fit	Strong — broadest entity coverage, rich edge types (indication/contraindication/off-label), disease+drug feature tables. More granular than GP-KG but different relation schema.
Wishlist	Orient PrimeKG’s seven hierarchies from their source ontologies. PrimeKG labels them all `parent-child` and stores each edge both ways. Before the doubling it built MONDO, HPO and GO parent→child but UBERON child→parent, so only the ontologies can orient them. Training exports keep them in both directions until then (D-Q4-14, #610; the anatomy and cellular-component mapping question is #866). Not scheduled.

CIViC

Attribute	Value
Resource	Clinical Interpretations of Variants in Cancer
Project	civicdb.org (WashU)
DB row	`mods_kg_source.id='civic'` (status: `loaded`)
Connector	`backend/connectors/civic_connector.py`
What it adds	Curated variant→drug response / prognostic / predictive / diagnostic evidence with evidence-tier grading. Critical for variant-aware reasoning (Phase 8.5).
Domain fit	Strong — clinically graded somatic-variant evidence is the closest thing to oncology gold-standard for KG-based reasoning. Used in `maureen_focus_v1`, `primekg_civic_v1`, `tier1_mvp_v1`.

ClinVar

Attribute	Value
Resource	NCBI ClinVar — variant ↔ condition assertions
Connector path	Via `eutils-esummary` (see DB version string)
DB row	`mods_kg_source.id='clinvar'` (status: `loaded`)
What it adds	Variant identity + germline disease assertions. Smaller than CIViC because we only ingested the variation IDs relevant to our cohort.
Domain fit	Germline complement to CIViC’s somatic-oncology focus. Used in Maureen compositions.

OncoKB

Attribute	Value
Resource	Memorial Sloan Kettering OncoKB precision-oncology knowledgebase
Project	oncokb.org
DB row	`mods_kg_source.id='oncokb'` (status: `loaded`)
License gotcha	Commercial, with a free academic tier. We’re on academic-annotate. Surface to user before any redistribution / publication use; see `memory/feedback_user_handles_ethical_calls.md`.
What it adds	FDA + NCCN level-of-evidence tagging for variant→therapy claims in cancer. Smaller graph (176 ent / 483 rel; 1,221 / 7,217 after #744’s re-annotation, 2026-10-05) but very high-precision.
Domain fit	High-precision oncology layer for Maureen compositions.
Wishlist	Fusion annotation through OncoKB’s structural-variant endpoint (`annotate/structuralVariants`). `byProteinChange` cannot express a fusion: all 480 `::` queries came back `geneExist=False`, the annotator now refuses them, and #744’s compound-profile retirement took the graph’s only OncoKB fusion edges. Not scheduled.

PharmGKB CPIC

Attribute	Value
Resource	PharmGKB — Clinical Pharmacogenetics Implementation Consortium variant guidelines
Project	pharmgkb.org
DB row	`mods_kg_source.id='pharmgkb_variants'` (status: `loaded`)
What it adds	CPIC-graded pharmacogenetics: variant→drug toxicity / response / metabolism evidence. Layered on top of `primekg_civic_v1` to make `primekg_civic_pharmgkb_v1`.
Domain fit	PGx layer (germline drug metabolism) — complements CIViC (somatic oncology) and ClinVar (germline disease).

Maple Patient Overlay

Attribute	Value
What it is	Internal patient-as-node overlay. Currently just Maureen.
DB row	`mods_kg_source.id='maple_patient'` (status: `loaded`)
Pattern	Unioned into compositions at export via `--include-source maple_patient`. NOT a permanent member of any composition; it’s overlaid on top of whatever base composition the export targets.
Reference	`memory/project_pdr2_mechanism.md`, `memory/project_paan_thinking_doc.md`, `docs/paan_planning.md`
Domain fit	Required for any per-patient training run. The whole PDR2 personalization story passes through this.

Partial / Running KGs

RTX-KG2

Attribute	Value
Resource	NCATS Translator canonicalized KG (~70 upstream biomedical sources merged + normalized to Biolink)
Project	github.com/RTXteam/RTX-KG2
DB row	`mods_kg_source.id='rtx_kg2'` (status: `partial`)
Status detail	~7% loaded; full raw data on disk; full load tracked in `GH #85`. See `memory/project_kg_source_completeness.md`.
What it adds (when complete)	Biolink-normalized predicates across most of the public biomedical-KG ecosystem — absorbs most of what PrimeKG carries, plus much more.
Compositions	`rtx_kg2_full`, `rtx_kg2_tier1` (empty until D0 lands); `tier1_mvp_v1` headline (empty until D0/D1/D2/D3 land)
Domain fit	Highest potential — the single largest source we can ingest, and Biolink-canonical, so it composes cleanly. Bottleneck right now is finishing the load.

DRKG (Drug Repurposing Knowledge Graph)

Attribute	Value
Paper	Ioannidis et al. / SNAP group
Repository	github.com/gnn4dr/DRKG
Entity types	Drug, disease, gene, anatomy, atc, pharmacologic class, side effect, symptom, pathway
Upstream sources	DrugBank, Hetionet, GNBR, String, IntAct, DGIdb
Format	TSV triples
DB row	`mods_kg_source.id='drkg'` (status: `running` — ingest in progress as of 2026-05-29)
Domain fit	Strong — drug repurposing focus, active community, well-documented. Similar scale to GP-KG (~97K ent / ~5.9M rel). Good third-KG comparator once loaded.

Variant cross-reference services

Loaded with identifiers only — no edges contributed to the canonical KG. Used to validate / normalize variant references that come in via CIViC / ClinVar / OncoKB / PharmGKB.

Source ID	Resource	Purpose
`dbsnp`	NCBI dbSNP rsIDs	Standardize variant identifiers
`gnomad`	gnomAD allele frequency	Population-frequency context for variants
`myvariant_aggregator`	MyVariant.info	Convenience aggregator across variant sources

API-only KGs

NCATS Biomedical Data Translator

Attribute	Value
Type	Federated query API (TRAPI 1.6.0), not a static downloadable KG
Endpoints	ARAGORN, ARAX
Coverage	200+ biomedical data sources
DB row	`mods_kg_source.id='translator_ncats'` (status: `api_only`)
Connector	`backend/connectors/translator_connector.py`
Queries	`query_treatments_for_disease()`, `query_drug_targets()`, `query_drug_pathways()`
Domain fit	Complementary — on-demand enrichment, not suitable for batch GNN training

TxGNN (Therapeutic Genome-wide Neural Network)

Attribute	Value
Type	Pre-trained GNN model (uses PrimeKG internally)
Paper	Huang et al., Harvard
DB row	`mods_kg_source.id='txgnn'` (status: `api_only`)
Connector	`backend/connectors/txgnn_connector.py`
Capabilities	`predict_for_disease()` → (indications, contraindications) with scores
Disease coverage	Colorectal, breast, lung, pancreatic, ovarian cancer
Domain fit	Comparator — independent predictions from the same underlying KG (PrimeKG)

Reference / not-yet-evaluated

HGTDR (Heterogeneous Graph Transformer for Drug Repurposing)

Attribute	Value
Paper	2024, Bioinformatics
Type	Heterogeneous KG + HGT embedding + link scoring
Mentioned in	`docs/_archive/2026-06-02/deep-research-report.md` (archived)
Domain fit	Architectural reference — 3-step KG construction approach may inform our pipeline. Not in DB.

ULTRA Pretraining KGs (non-biomedical)

These three datasets aren’t biomedical and don’t enter the MODS pipeline directly. They’re the standard pretraining mix for ULTRA (Galkin et al., ICLR 2024) — the relation-invariant foundation model that PDR2 builds on. ULTRA’s whole pitch is that it doesn’t memorize entities or relations; it learns the structure of the relation graph. To learn that, it needs exposure to varied relation patterns — dense vs sparse, broad vs taxonomic, factual vs modern. These three together cover that space well enough that the trained model transfers zero-shot to 50+ unseen KGs (including biomedical ones like Hetionet and PrimeKG) without fine-tuning.

ULTRA paper: Galkin M, Yuan X, Mostafa H, Tang J, Zhu Z. “Towards Foundation Models for Knowledge Graph Reasoning.” ICLR 2024. arxiv:2310.04562. Mirror at docs/References/ULTRA/.

These KGs are not tracked in mods_kg_source; they live in the PDR2 vendor tree (lib/pdr2/) and are only used for pretraining the ULTRA backbone.

FB15k-237

Attribute	Value
Paper	Toutanova K, Chen D. “Observed versus latent features for knowledge base and text inference.” Proc. 3rd Workshop on Continuous Vector Space Models and their Compositionality, 2015
Origin	Subset of FB15k (Bordes et al. 2013), with inverse-relation leakage removed (e.g., `(A, born_in, B)` / `(B, contains, A)` redundancies stripped)
Source domain	Freebase — general world knowledge (people, places, films, organizations)
Entities	14,541
Relations	237
Triples	272,115 train / 17,535 val / 20,466 test
Character	Dense, many-to-many, broad-domain factual
Standard download	Bundled with most KGE libraries (PyKEEN, OpenKE, DGL-KE); also via the ULTRA repo
Role in mix	Teaches the model dense, many-relation graph patterns

WN18RR

Attribute	Value
Paper	Dettmers T, Minervini P, Stenetorp P, Riedel S. “Convolutional 2D Knowledge Graph Embeddings.” AAAI 2018
Origin	Subset of WN18 (Bordes et al. 2013), with inverse-relation leakage removed (same fix as FB15k-237)
Source domain	WordNet — English lexical database; relationships between word senses (synsets)
Entities	40,943 synsets
Relations	11 (`hypernym`, `hyponym`, `member_meronym`, `derivationally_related_form`, etc.)
Triples	86,835 train / 3,034 val / 3,134 test
Character	Sparse, deep-hierarchy, taxonomic — tree-like chains, mostly chain-following reasoning
Standard download	Bundled with most KGE libraries; also via the ULTRA repo
Role in mix	Teaches the model sparse, deep-hierarchy graph patterns

CoDEx (CoDEx-M used by ULTRA)

Attribute	Value
Paper	Safavi T, Koutra D. “CoDEx: A Comprehensive Knowledge Graph Completion Benchmark.” EMNLP 2020
Source domain	Wikidata — modern, type-aware, multi-domain
Variants	CoDEx-S (small), CoDEx-M (medium — used by ULTRA), CoDEx-L (large)
CoDEx-S size	~2,000 entities, 42 relations, ~36K triples
CoDEx-M size	~17,050 entities, 51 relations, ~206K triples ← the one ULTRA pretrains on
CoDEx-L size	~77,950 entities, 69 relations, ~612K triples
Special features	Includes entity-type information and curated hard negatives (designed to be harder than FB15k-237’s near-trivial negatives)
Standard download	github.com/tsafavi/codex
Role in mix	Teaches the model mid-density, modern Wikidata-style patterns with type signal and harder negatives

Planned experiment — 64d vs 128d ULTRA pretraining

Current ULTRA pretraining (lib/pdr2/) uses 64-dimensional hidden representations. Planned sweep: re-pretrain on the same three-KG mix (FB15k-237 + WN18RR + CoDEx-M) at 128d and compare:

(a) Standard link-prediction metrics (MRR, Hits@k) on the three pretraining datasets
(b) Zero-shot transfer to GP-KG and PrimeKG (the targets we actually care about)
(c) Downstream impact on per-patient PDR2 fine-tune AUPRC (Maureen + RGCC validation)

Hypothesis: larger capacity may help on harder transfer targets, but the pretraining set is small (combined ~565K triples) so over-parameterization is a real risk. Worth measuring rather than assuming. Track in a follow-up writeup under docs/pdr2_* when run.

Compositions

Compositions are named, versioned graph mixes that training runs reference by ID. Live list via mods_kg_list_compositions. Snapshot 2026-05-29:

ID	Members	Purpose
`gp_kg_baseline_v1`	gp_kg	Reproduces Gao 2022 / PDR1 baseline
`primekg_baseline_v1`	prime_kg	Full PrimeKG, PDR2 baseline
`primekg_civic_v1`	prime_kg + civic	First base composition for canonical → PDR2 pipeline
`primekg_civic_pharmgkb_v1`	prime_kg + civic + pharmgkb_variants	Adds CPIC PGx evidence on top of `primekg_civic_v1`
`maureen_focus_v1`	PrimeKG (anatomy/CC/MF/exposure excluded) + civic + pharmgkb_variants + oncokb + clinvar	Maureen proof-out, scoped PrimeKG
`maureen_full_v1`	Same members, no PrimeKG domain exclusions	Maureen full-vs-focus ablation
`rtx_kg2_full`	rtx_kg2	All of RTX-KG2 (empty until D0 lands)
`rtx_kg2_tier1`	rtx_kg2 (Tier 1 domains only)	Faster iteration variant (empty until D0 lands)
`tier1_mvp_v1`	rtx_kg2_tier1 + civic + oncokb + pharmgkb_variants	Headline composition — skips PrimeKG because RTX-KG2 absorbs it (empty until D0/D1/D2/D3 land)
`tier1_mvp_with_primekg_v1`	tier1_mvp_v1 + prime_kg	Cross-KG corroboration variant

Patient overlay pattern: maple_patient is unioned at export via --include-source maple_patient, not as a permanent composition member. This keeps Maureen out of the canonical composition graph while letting any base composition be exported with the patient overlay on top.

Composition lock state, exact filters (include_domains / include_relations / weight), and member counts are in mods_kg_composition + mods_kg_composition_member. Query via mods_kg_list_compositions for live state.

Comparison Matrix

High-level fitness comparison for the biomedical KGs. For live counts / status / license, query mods_kg_list_sources.

Feature	GP-KG	PrimeKG	RTX-KG2	DRKG	Translator
Static downloadable	Yes	Yes	Yes	Yes	No (API)
Drug-disease edges	Yes (treat)	Yes (indication, contraindication, off-label)	Yes	Yes	Yes
Drug-gene edges	Yes (target)	Yes (target, enzyme, carrier, transporter)	Yes	Yes	Yes
Disease-gene edges	Yes (associate)	Yes (disease_protein)	Yes	Yes	Yes
Phenotype coverage	MP + HP	effect/phenotype	Yes	symptom	Limited
GO term coverage	Yes	biological_process, molecular_function, cellular_component	Yes	No	No
Tissue/anatomy	Yes (tissue)	Yes (anatomy)	Yes	Yes (anatomy)	Limited
Drug feature tables	No	Yes (MoA, ATC, pharmacokinetics)	Via upstreams	No	No
Disease feature tables	No	Yes (MONDO, Orphanet, Mayo)	Via upstreams	No	No
Biolink-canonical predicates	No	Partial	Yes	No	Yes
Published repurposing baselines	Yes — KG-Predict CV AUROC 0.981; AD-CS AUROC 0.868	Via TxGNN	—	Some	—

Adding a New KG

To add a new knowledge graph to the pipeline:

Connector: Create backend/connectors/<kg_name>_connector.py implementing:
- is_downloaded() → bool
- download(progress_callback) → Dict
- stream_triples(batch_size) → Generator[List[Tuple], None, None]
- get_stats() → Dict
- data_summary() → Dict
- count_triples() → int
Register in DB: Insert a row in mods_kg_source (or use the bootstrap path that wires it during worker/task_handlers.handle_import_kg). Set initial status='planned', fill citation / source_url / license / version. Status auto-advances to running then loaded during ingest.
Load: Trigger via mods_kg_import MCP tool or directly through ModsKGCanonicalService.import_from_connector(kg_source, connector).
Validate: Run mods_kg_canonical_stats and mods_kg_list_sources to confirm live counts line up with what the connector said it would emit.
Composition (optional): If the new KG should participate in a training run, add a member row in mods_kg_composition_member for the relevant composition(s), with any include_domains / include_relations filters.
Document here: Add a section under “Loaded KGs” with paper / DOI / domain fit narrative and known issues. Don’t duplicate node/edge counts or license text — those live in the DB.
Feature tables (optional): If the KG ships supplementary node features, extend mods_kg_disease_feature / mods_kg_drug_feature or create new tables via migration.

Critical: Preserve source entity types and relation types exactly. Never collapse finer-grained types into coarser ones during import. See “Data and Mapping Integrity Rules” in CLAUDE.md — this rule was paid for in real data loss.

MODS Notebook

Explorer