# Leakage-Free Gold Standard PPI Dataset (Bernett et al.)

## Facts

- Provider: Bernett, Blumenthal and List (TUM daisybio); packaged by Synthyra
- License: CC-BY-4.0 (https://api.figshare.com/v2/articles/21591618)
- Size: 5 files
- Format: Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/Synthyra/bernett_gold_ppi
- Access: public, readable without an account
- DOI: https://doi.org/10.6084/m9.figshare.21591618.v4
- Category: Proteomics
- Source: https://doi.org/10.1093/bib/bbae076
- Page: https://scigantic.com/archive/3bdb3b5a-48ef-5138-ac52-91168829d59e
- Keywords: Proteomics, protein language model, benchmark, Parquet, Hugging Face

## About

Parquet packaging of the gold standard dataset for sequence-based protein-protein interaction (PPI) prediction from Bernett, Blumenthal and List, Cracking the black box of deep sequence-based protein-protein interaction prediction (Briefings in Bioinformatics 2024). The benchmark was built to remove the data leakage that inflates published PPI accuracies: the human proteome was partitioned with KaHIP so that sequence similarity between partitions is minimised, proteins in training never appear in validation or test, and CD-HIT removed redundancy above 40 percent pairwise identity inside each set. Positives come from the HIPPIE interaction database and sequences from Swiss-Prot; the data are balanced binary labels. Splits: 163,192 training pairs (Intra-1), 59,260 validation pairs (Intra-0) and 52,048 test pairs (Intra-2). Each row has the two UniProt accessions (A, B), a label and both amino acid sequences (SeqA, SeqB). Measured 2026-11-14: three parquet files under data/ totalling about 0.29 GB. Load with datasets.load_dataset('Synthyra/bernett_gold_ppi') or pandas. The authors' follow-up study reports that deep models on this benchmark plateau near 0.65 accuracy, which makes it a hard test for protein language model embeddings and PPI classifiers. The original figshare record is licensed CC BY 4.0; cite Bernett et al. 2024.

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("Synthyra/bernett_gold_ppi")
print(ds)
```

## Data structure

File types in sampled files: Parquet

Example file: `data/test-00000-of-00001.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/3bdb3b5a-48ef-5138-ac52-91168829d59e
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/3bdb3b5a-48ef-5138-ac52-91168829d59e
- Over MCP: `get_schema_card` with id `3bdb3b5a-48ef-5138-ac52-91168829d59e` at https://api.scigantic.com/mcp

## Related topics

- [Parquet datasets](https://scigantic.com/topics/parquet)
- [Benchmark datasets](https://scigantic.com/topics/benchmark)
- [Proteomics datasets](https://scigantic.com/topics/proteomics)
- [Protein language model datasets](https://scigantic.com/topics/protein-language-model)
