bernett_gold_ppi_2

The human protein-protein interaction benchmark of Bernett, Spannagl, Ås, List and Blumenthal (2026), "Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets" (arXiv:2609.10193). It is the HIPPIE v3.0 dataset in the paper's ILP split + ILP sampling setting, which the authors report "yields the least-biased datasets". It is redistributed unchanged from their figshare release (doi:10.6084/m9.figshare.33407437, folder results_esm2/Hippie-ILP-split-and-sampling). Every row, pair, label and row order is the authors'; the only addition is each protein's sequence, taken from the release's own data/sequences.fasta.

It succeeds Synthyra/bernett_gold_ppi, the 2024 gold standard of Bernett, Blumenthal and List (Briefings in Bioinformatics 25(2):bbae076), and keeps its columns, so code that reads one reads the other.

Splits

Split	Rows	Positives	Negatives	Proteins	Authors' file
train	957,610	478,805	478,805	15,763	train.csv
valid	112,678	56,339	56,339	4,614	val.csv
test	112,402	56,201	56,201	5,018	test_balanced.csv
test_realistic	618,211	56,201	562,010	5,018	test_realistic.csv

test and test_realistic hold the same positives and the same proteins. test is balanced 1:1 with negatives drawn by the same ILP sampler as train and valid; test_realistic has ten negatives per positive, sampled uniformly. No protein appears in two of train, valid and test, and no sequence is shared between train and a held-out split under another accession.

Columns

Column	What it holds
A	UniProt accession of the first protein (the authors' protein1); isoforms carry a suffix such as -2
B	UniProt accession of the second protein (the authors' protein2)
labels	1 for an interaction in HIPPIE, 0 for a sampled pair (int8)
SeqA	Sequence of A, from the release's sequences.fasta
SeqB	Sequence of B, from the release's sequences.fasta

How the authors built it

1. Positives. Human, experimentally detected interactions from HIPPIE v3.0: 1,179,347 unique pairs over 29,080 UniProt entries, with sequences, taxonomy and GO annotations fetched from UniProt.
2. Similarity. An all-against-all blastp search (BLAST 2.16.0), with bitscores normalized by the shorter sequence, and a KaHIP (3.25) partition of that similarity graph into 100 blocks.
3. ILP split. An integer linear program assigns blocks to train, validation and test near target fractions (deviation 0.05), minimizing the interactions lost across splits. 587,512 positives crossed splits and were discarded.
4. Similarity reduction. CD-HIT-2D removes every validation or test interaction with a protein at 40% identity or more to a training sequence (490 positives removed).
5. ILP sampling. Negatives come from a candidate pool four times the positives, chosen by a second integer linear program that matches the positives on self-interactions, per-protein degree (capped at six times the positive degree), taxon pairs and mean GO biological process similarity, so none of these separates the classes.

Proteins

The 15,763 + 4,614 + 5,018 proteins in the splits are 25,395 accessions with 24,052 distinct sequences. 5,903 are isoform accessions, 1,721 are unreviewed (TrEMBL) entries, and 1,312 sequences appear under more than one accession. 2,660 reviewed human entries (UniProt 2026_03, 20,431 in all) appear in no split.

Scores reported in the paper

The paper is an audit, not a benchmark: its only classifier is a Random Forest on protein language model embeddings. On this setting it reports (Tables A1 and B1):

Embeddings	Test set	AUROC	AUPRC	MCC
ESM-2	test	0.57	0.56	0.10
ESM-2	test_realistic	0.48	0.10	-0.01
ProtT5	test	0.58	0.58	0.09
ProtT5	test_realistic	0.52	0.11	0.01

The same classifier scores 0.85 AUROC in the paper's maximally biased HIPPIE setting, which is the gap this setting is built to remove. On test_realistic the positive rate is 1 in 11, so an AUPRC of 0.09 is chance.

Licence and citation

The data are CC BY 4.0, as released on figshare. HIPPIE, UniProt and the code (bionetslab/ppi-splitting-pipeline) keep their own terms.

@article{bernett2026shortcuts,
  title   = {Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets},
  author  = {Bernett, Judith and Spannagl, Anton and {\AA}s, Joel and List, Markus and Blumenthal, David B.},
  journal = {arXiv preprint arXiv:2609.10193},
  year    = {2026}
}