D2PPI

D2PPI benchmarks: sequence-based protein–protein interaction prediction Three benchmarks for predicting PPI from sequence, in one format. They are the benchmarks of the D2PPI paper. Each benchmark has a train/val/test split, and each split comes as two files: a pair table (`.tsv`): one protein pair per row, with both sequences and the label an embedding file (`.npz`): ESM-2-650M domain and whole-protein embeddings of every protein in the pair table The pair tables are all you need to evaluate any method. The embedding files are optional: they hold the exact inputs of the D2PPI models in the paper. directory benchmark train pairs val pairs test pairs test sets 1 bernett_human Bernett leakage-free human gold standard 163,192 59,260 52,048 human 2 dscript_xspecies D-SCRIPT cross-species (STRING v11) 421,792 52,725 242,000 mouse, fly, worm, yeast, E. coli 3 d2ppi_xspecies LF-XS: leakage-free cross-species (STRING v12.5) 2,244,473 40,997 1,912,383 fly, worm, yeast, E. coli, Arabidopsis Positives to negatives: 1 : 1 in Bernett, 1 : 10 in every split of D-SCRIPT and LF-XS. Train and validation are human in all three benchmarks.