Pharmacogenomic Knowledge Graph (PGx-KG) Dataset

Pharmacogenomic Knowledge Graph (PGx-KG) Dataset

This repository contains the processed pharmacogenomic knowledge graph (PGx-KG) prepared for benchmarking link prediction models. The graph integrates curated relationships from PharmGKB, ClinVar, SIDER, and Reactome to standard biomedical vocabularies so that entities and relations align across resources.

Contents

• nodes.csv: Consolidated catalogue of biomedical entities with harmonized identifiers, semantic types, and source provenance.
• edges.csv: Full edge list covering all relation types with integer IDs referencing the node maps.
• triples/: Relation-specific subject–predicate–object triples in CSV format, named as <REL>__HEADTYPE__TAILTYPE.
• relations/: Per-relation tables with descriptive metadata, confidence scores, and source annotations.
• id_maps/: Mapping tables that link integer IDs to canonical biomedical identifiers (HGNC, RxNorm, MeSH, ChEBI, etc.).
• degree/: Per-relation in/out degree distributions for downstream analysis or sampling.
• splits/: Stratified train/validation/test partitions for each relation type, with only CSV versions (including reproducible random/ test/train/val).
• stats.json: Summary statistics computed after harmonization (entity counts, relation counts,).

Data Statistics

Total nodes: 3,744,727 Total edges: 9,645,367

Node types:

• Variants,
• Genes,
• Diseases,
• Drugs,
• ADRs, 

Pathways Relation types:

• VAR_ASSOC_DIS,
• VAR_IN_GENE,
• DRUG_CAUSES_ADR,
• GENE_IN_PATH,
• DRUG_IN_PATH,
• GENE_AFFECTS_DRUG

How to Use

1. Load the relation-specific triples in triples/ (or the global edge list in edges.*) into your preferred graph ML framework (e.g., PyTorch Geometric, DGL).
2. Use the mapping tables in id_maps/ (e.g., GENE_ids.csv, DRUG_ids.csv) to translate integer IDs back to canonical biomedical identifiers.
3. Evaluate models on the curated train/validation/test splits under splits/<REL>/ and report metrics such as MRR, Hits@K, or AUROC.

Source Databases

This dataset was derived from the following resources (not redistributed here):

• PharmGKB (https://www.pharmgkb.org/)
• ClinVar (https://www.ncbi.nlm.nih.gov/clinvar/)
• SIDER (http://sideeffects.embl.de/)
• Reactome (https://reactome.org/)

License

This processed dataset is released under the CC BY 4.0 License. You are free to share and adapt, provided appropriate credit is given.

Citation

Citation

If you use this dataset, please cite both the dataset and the preprint:

• Faruk, M.O. (2025). Pharmacogenomic Knowledge Graph (PGx-KG): Processed Dataset for Link Prediction. Zenodo. https://doi.org/10.5281/zenodo.17189995

• Faruk, M.O. (2025). A large-scale pharmacogenomic knowledge graph for drug-gene-variant-disease discovery. medRxiv. https://doi.org/10.1101/2025.09.24.25336269v1