ogbl-ppa

Predict held-out biologically meaningful protein-protein association edges. Each positive is ranked against three million sampled negative edges and evaluated by Hits@100.

The undirected graph contains 576,289 proteins from 58 species and STRING associations arising from physical interaction, co-expression, homology, genomic neighborhood, and other evidence.

Each node has only a 58-dimensional one-hot species indicator in the standard artifact. The released training graph contains training edges but excludes validation and test associations.

The biological-throughput split uses high-throughput experimental or computationally derived associations for training and reserves resource-intensive low-throughput experimental associations for validation and test.

STRING associations are broader than direct physical binding. Randomly sampled absent pairs may include unknown biology rather than confirmed negatives, while Hits@100 against millions of candidates is much stricter than ROC-AUC. A random edge split would discard the intended evidence-quality shift.