PlantMetWiki Full Dataset: 421-species enzyme-reaction graph with node embeddings and link-prediction splits

The complete PlantMetWiki plant-metabolism knowledge graph (424 plant species), prepared for heterogeneous-graph link prediction. Task: predict (Protein, catalyzes, Interaction) edges ("this enzyme catalyses this reaction"), where the destination is always a Conversion-subtype node (the reaction itself). Three complementary train/val/test split strategies are included: a random 80/10/10 split, an organism-held-out split (tests cross-species generalisation), and a pathway-held-out split (tests generalisation to unseen biological contexts). Includes pre-computed node embeddings, already attached to the PyTorch Geometric HeteroData object as node features: MAP4 binary fingerprints for metabolites, ESM-C (esmc_300m) embeddings for proteins, and PlantCaduceus (l32) DNA embeddings for gene products. See README.md inside the archive for the full file list, coverage numbers, and loading code, and LICENSE.md for provenance of each embedding's source sequences. v4 changes (2026-07-03 — learnathon baseline updated to v6; data files unchanged): The graph (heterodata.pt) and splits (splits_taxa.pt) are identical to v3. Changes are in learnathon training code and documentation: (1) dataset.py adds disjoint_train_ratio=0.2 default — 80% of training positives enter the message-passing graph, 20% are supervision-only, preventing the 1-hop memorisation shortcut; (2) dataset.py removes (Interaction, catalyzed_by, Protein) from the MP graph by default (keep_catalyzed_by=False), closing the 2-hop P→C→P shortcut; (3) train.py default negative sampling changed from 5+5 (random+corrupt-protein) to 5+0 — corrupt-protein negatives hurt P-H@50 when used aggressively (see README §Baseline); (4) learnathon metrics clarified: P-H@50 is the primary leaderboard metric, CP-AUC/CP-AP are secondary diagnostics, standard AUC/AP (random negatives, always ≥0.95) are omitted from reporting; (5) is_about edges confirmed load-bearing — ablation (v7) that removed them caused test P-H@50 to drop below random (0.003) and AUC to invert to 0.188. Baseline results: val P-H@50=0.124 (21× random), test P-H@50=0.071 (12× random), val CP-AUC=0.817 (emergent, no CP training). v3 changes (graph redesign): (1) generic DataNode/External nodes eliminated; (2) Inhibition/Stimulation/ComplexBinding-subtype nodes dropped; (3) TranscriptionTranslation collapsed into a direct (GeneProduct, encodes, Protein) edge. v2 changes: target edge collapses through Catalysis to Conversion; embedding-coverage-aware taxa split. See README.md inside the archive for full changelog. This is the full-graph counterpart of the 2-species Learnathon subset (10.5281/zenodo.20736061) — same task and file format, complete species coverage instead of an A. thaliana / G. max slice.