ogbg-ppa

Given a protein-association neighborhood, predict the broad taxonomic group of its source species. The output is one of 37 classes and is scored by accuracy.

OGB samples 100 center proteins from each of 1,581 STRING species, extracts their two-hop association neighborhoods, removes the center, and subsamples graphs to fewer than 300 protein nodes. The resulting 158,100 graphs cover diverse branches of the tree of life.

Nodes carry no intrinsic biological feature in the standard artifact. Each association edge has seven confidence values for evidence channels such as co-expression, gene neighborhood, co-occurrence, and gene fusion.

Validation and test graphs come from species absent from training while remaining within the same 37 taxonomic groups. Accuracy therefore measures cross-species generalization rather than memorization of neighborhoods from the same organism.

The target is the graph’s taxonomic origin, not a protein function. STRING edges mix experimental and computational evidence, and the seven confidence channels can encode species-specific collection patterns. Random graph splits would leak species identity and materially change the task.