ogbn-proteins

The Open Graph Benchmark node property prediction dataset ogbn-proteins, repacked as numpy arrays.

The graph is an undirected protein association network from STRING: 132,534 proteins across eight species, 39,561,252 associations, each association carrying eight weights (homology, neighborhood, fusion, cooccurrence, coexpression, experimental, database, textmining). Every node carries 112 binary labels, one per gene ontology function annotation.

Files

graph/graph.npz edges (39561252, 2) int32, each association once with both endpoints, the graph is undirected edge_feat (39561252, 8) float16, the eight association weights in [0, 1] node_species (132534,) int32, the NCBI taxonomy id of each protein

graph/targets.csv node_id, the 24,679 nodes of the species split's target group, shuffled with seed 20260909

labels/train_labels.npz node_ids (86619,) int32 and labels (86619, 112) int8, the training group of the species split

labels/valid_labels.npz node_ids (21236,) int32 and labels (21236, 112) int8, the validation group of the species split

Provenance

Built from the official archive http://snap.stanford.edu/ogb/data/nodeproppred/proteins.zip (ogbn-proteins RELEASE v1, May 2020) with no change to node ids, edges or labels. The only transformations are the container format, the edge weights cast from float32 to float16 (maximum absolute error 2.42e-4 on values in [0.001, 1.0]), and the species split written out as three files.

OGB distributes ogbn-proteins under CC0 1.0. Underlying data: STRING v11 (Szklarczyk et al., 2018) and the Gene Ontology (Ashburner et al., 2000).