UNICORNE-BENCH-1: benchmark data for sequence-to-ensemble protein structure predictors
Authors/Creators
- 1. Center for AI in Protein Dynamics, Vanderbilt University
Description
Benchmark data for The UNICORNE Protein Conformation Database Enables Large-Scale Benchmarking of Dynamic Structure Prediction (NeurIPS 2026, Evaluations & Datasets Track).
UNICORNE-BENCH-1, the first release of the UNICORNE benchmark, is a set of multi-conformation proteins, each a single UniProt entry with at least two experimentally resolved conformations (839 evaluated; 829 after fold-quality review). This record contains the benchmark definition, the superposed ground-truth structures, the OOD60 and DOMAINMOTION calibration sets, GT×GT TM-score matrices, and per-ensemble and per-conformation scores for eight predictors (BioEMU, ESMFlow, ESMDiff, AlphaFlow, CFrandom, AFsample2, AFsample3, SPEACH_AF). Together with the code repository it reproduces every figure and table in the paper's results section.
The archive's layout matches the code repository; extract it at the repository root. UNICORNE-BENCH-1.README.md lists the contents. Predicted structures and raw TM-align output are not included.
Version 1.1 recomputes the maximum-displacement matrices (max_disp_xx.npy) after superposing with the TM-align rotation, which version 1 omitted, and updates the paper title. All other files are unchanged.
Files
UNICORNE-BENCH-1.README.md
Additional details
Related works
- Is part of
- https://unicornedb.org (URL)
- Is supplemented by
- Software: https://github.com/OliverLaboratory/UNICORNE-BENCH (URL)
- References
- Dataset: 10.5281/zenodo.20069060 (DOI)