Protein Interface Dataset · Dror Lab, Stanford
DIPS
Tens of thousands of binary protein complexes, for learning which residues form an interface.
What DIPS is
Interface prediction was being trained on the few hundred complexes in the standard docking benchmark, which is far too small for a structure-aware network. DIPS is the alternative: a pipeline that mines the Protein Data Bank's biological assemblies for binary complexes, applies the filtering that makes them usable as training data, and emits interacting residue pairs — 42,826 complexes in the authors' run. What the repository distributes is the pipeline rather than the assembled data, so the collection is reproduced from a PDB mirror on the machine that will train on it; the intermediate parsed and paired representations are kept, which is what makes it useful beyond its original paper. The cleaned Docking Benchmark 5 set used for testing is published ready-made.
Get the data
Its authors publish DIPS outside the Hugging Face Hub, so the dataset explorer has no table to read and there are no rows on this page. The files are linked below, and the research agent can fetch them onto its machine and work through them there.
Try asking
“Set up the DIPS pipeline on the machine against a small sample of PDB biological assemblies and show me the interacting-pair output for one complex.”
Licence
Commercial use permitted · CC0 1.0 · code MIT
Read 2026-10-24: the processing pipeline declares MIT, and the companion Dataverse deposit holding the prepared test set and trained-model data declares CC0 1.0. The complexes themselves are built from Protein Data Bank biological assemblies, which place no restriction on reuse — but they are fetched from the PDB by the pipeline, not redistributed here.
Questions
Can I browse DIPS in the browser?
No. Its authors publish it outside the Hugging Face Hub, so the dataset explorer has no table to read and there are no rows to show here (format: Pickled pandas frames, paired structures and TFRecords, produced by the pipeline). This page describes what it holds and links to the files; the research agent can fetch and read them on its machine.
What licence is DIPS under?
CC0 1.0, with the accompanying code under MIT. Read 2026-10-24: the processing pipeline declares MIT, and the companion Dataverse deposit holding the prepared test set and trained-model data declares CC0 1.0. The complexes themselves are built from Protein Data Bank biological assemblies, which place no restriction on reuse — but they are fetched from the PDB by the pipeline, not redistributed here.
How to cite it
Townshend, R. J. L., Bedi, R., Suriana, P. A. & Dror, R. O. End-to-end learning on 3D protein structure for interface prediction. NeurIPS (2019).