PPI prediction from sequence, gold standard dataset

Gold Standard Dataset for sequence-based PPI prediction:

• Big dataset: 163,192 training points (Intra-1), 59,260 validation points (Intra-0), 52,048 test points (Intra-2)) + corresponding protein sequences from Swissprot
• No direct data leakage: proteins from training are not contained in validation or test, proteins from validation are not in training or test, proteins from test are not in validation or training
• Minimized sequence similarity between training, validation, test because whole human proteome was split with KaHIP such that sequence similarities are minimized w.r.t. length-normalized bitscores
• Redundancy-reduction with CD-HIT: inside of the datasets, no proteins with >40% pairwise sequence similarity