Multi-Task Benchmark · IBIVU, VU Amsterdam
ProteinGLUE
Seven per-residue and per-protein tasks for evaluating protein representation models on one footing.
What ProteinGLUE is
A protein language model is usually reported on whichever downstream task its authors chose, which makes two models hard to compare and easy to oversell. ProteinGLUE fixes the tasks and the splits: secondary structure in three and eight classes, buried solvent accessibility, hydrophobic-patch ranking, epitope prediction and protein–protein interface prediction, each with its own training, validation and test files, plus the pretraining corpus the authors used. Everything is distributed as plain CSV alongside the TFRecord form, which means the benchmark can be read and scored with no framework commitment — unusual for a suite of this kind, and the reason it is still straightforward to run.
Get the data
Its authors publish ProteinGLUE 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
“Download the ProteinGLUE secondary-structure task CSVs on the machine and report the class balance in the training split.”
Licence
Commercial use permitted · CC0 1.0
Read 2026-10-24: the download directory states CC-0 and serves the CC0 1.0 Universal text; the accompanying code repository declares CC0-1.0 as well. CC0 is a public-domain dedication — no attribution is required, though citing the paper is the norm.
Questions
Can I browse ProteinGLUE 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: CSV per task and split, with the same data as TFRecords). 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 ProteinGLUE under?
CC0 1.0. Read 2026-10-24: the download directory states CC-0 and serves the CC0 1.0 Universal text; the accompanying code repository declares CC0-1.0 as well. CC0 is a public-domain dedication — no attribution is required, though citing the paper is the norm.
How to cite it
Capel, H. et al. ProteinGLUE multi-task benchmark suite for self-supervised protein modeling. Scientific Reports 12, 16047 (2022).