THBKG — Temporal Heterogeneous Biomedical Knowledge Graph

A dated biomedical knowledge graph built from Open Targets 26.03 (with Reactome, ChEMBL and ClinicalTrials.gov), plus a clinical-advancement benchmark: rank target–disease pairs by their likelihood of advancing to Phase II, scored only from evidence datable strictly before each pair's decision year.

Every temporal edge carries the year its evidence first appeared, so the graph can be queried as of any historical decision point without leakage.

Files

Path	What it is
graph.safetensors	All graph tensors (node features, edge_index, edge_attr, edge_time), pickle-free.
metadata.json	Schema to reassemble the graph: node/edge types, tensor keys, shapes, dtypes.
mappings/<ntype>.parquet	node_id ↔ index for each node type (target, disease, molecule, go, reactome).
labels/train.parquet, labels/eval.parquet	Advancement labels: target_id, disease_id, transition_year, outcome.
load_thbkg.py	Standalone loader — rebuilds a PyG HeteroData, reads the label splits.
croissant.json	MLCommons Croissant metadata.

Graph at a glance

• Nodes: 110,396 across five types — target, disease, molecule, Gene Ontology term (go), Reactome pathway (reactome). Feature dims: molecule 1024, disease 256, target 56, GO 64, Reactome 64.
• Edges: ~11.1M over nineteen relations (sixteen temporal, three static); each temporal edge carries edge_time (year first observed) and edge_attr = [weight, novelty].
• Supervision is external. The graph holds only nodes and evidence edges; the clinical-advancement outcomes live in the label parquets and are joined to the graph by node ID at train/eval time. (Baking the label in as an edge made supervision indistinguishable from evidence and leaked it into message passing.)

Advancement benchmark

28,795 target–disease pairs, split temporally: 21,602 train (Phase II entry 1995–2015) / 7,193 eval (2016–2021, ~9.3% positive). outcome is True iff a Phase III trial for the same pair was registered within three years of Phase II entry. Each pair is scored using only graph evidence dated strictly before its transition_year.

Primary metric: Relative Success @ K (an importance-weighted hit rate), reported per therapeutic area and Wilcoxon-tested against a randomized-decisions baseline.

Quick start

from huggingface_hub import snapshot_download
from load_thbkg import load_graph, load_labels, node_index

path  = snapshot_download("<user>/THBKG", repo_type="dataset")
data  = load_graph(path)              # PyG HeteroData (reverse edges added, training view)
train = load_labels(path, "train")    # DataFrame: target_id, disease_id, transition_year, outcome
eval_ = load_labels(path, "eval")

# resolve a label endpoint to its graph node index
t2i, d2i = node_index(path, "target"), node_index(path, "disease")

The label parquets also load directly as a 🤗 datasets config:

from datasets import load_dataset
ds = load_dataset("<user>/THBKG", "advancement")   # splits: train, eval

License & citation

Dataset artifacts: CC-BY-4.0. Source code (construction pipeline, benchmark harness): MIT, at https://github.com/jackysiupuichung/THBKG. Archival DOI on Zenodo: concept 10.5281/zenodo.20795231.

Siu, Cabrera, Mudaliar, and Zubiaga.
THBKG: A Temporal Biomedical Knowledge Graph for Leak-Free Clinical Advancement Prediction.