Download
Data

Both files are UTF-8, tab-separated, one binary interaction per line, with a header row. Empty or unavailable fields are written as -.

• TXT format
• MITAB format

Ready-made train / validation / test partitions of the current release, for training protein–protein interaction predictors without configuring anything. Each archive is a .zip holding train.csv, val.csv and two test sets — test_balanced.csv (1:1 positives to negatives) and test_realistic.csv (1:10) — plus sequences.fasta (the sequence of every protein in the splits, headed by its UniProt accession). Negatives are either non-observed pairs (Complement) or highly replicated non-interactions (HRNIs): pairs that were tested in at least 3 (HRNI ≥ 3) or 5 (HRNI ≥ 5) independent experiments and never observed to interact. Read the Machine Learning guide first — it covers the data-leakage pitfalls these splits are designed around. For a split filtered your own way (by tissue, experiment type, source database, node degree, …), build one on the ML Splits page.

All releases — click any link to download.

If you need HIPPIE inside a pipeline rather than on disk, the same data is reachable as JSON over a small REST API. It covers the queries this site runs — interactions of a set of proteins, pairwise lookups, network expansion, the paginated proteome browser and a bulk TSV export — and takes the same filters as the web interface (confidence score, source database, experiment type, interaction type, tissue expression, isoforms). The ML split jobs above can be built and fetched the same way. No key or registration is needed. See the REST API reference for the endpoint list and parameters.

HIPPIE also speaks the Model Context Protocol, so an AI assistant can query it directly instead of you pasting results into one. Add the endpoint below to any MCP-capable client — there is nothing to install and no key to request.

https://cbdm-01.zdv.uni-mainz.de/hippienew/mcp

Five tools are available: resolve_protein (which HIPPIE protein an identifier refers to), get_interactions (a protein's partners), check_pairs (whether specific pairs interact), get_interaction_detail (the PMIDs, source databases and experimental methods behind one interaction), and list_filter_options (the filter vocabulary). Filters take the same values as the web interface, and also accept a category label such as Two-hybrid & complementation in place of listing every individual method. Results are capped per call and always report the full match count, so a truncated answer is visibly truncated.

Scores assigned to experimental techniques (used to score the interactions in HIPPIE).

• Technique scores v3.0 (TSV)
• Technique scores v2.3 (TSV)

The current HIPPIE release annotated with AlphaFold DB model identifiers and the structure-based confidence scores ipSAE and pDockQ2, as described in this BioRxiv preprint from NVIDIA.

• Only AFDB ID, ipSAE_max, and pDockQ2_max (TSV)
• All scores (TSV)