OpenBind-0 model release benchmarking materials

Protein-ligand benchmarking set of 462 systems, each with a ground-truth structure and a model input query, together with the multiple sequence alignments (MSAs). Developed for the OpenBind-0 model release. At the time of release, the official updated PLINDER has not been released. We urge potential users of this benchmark to check that the updated release has occured and to use their official benchmark instead of this preview version.

• queries.json — 462 queries (483 protein and 890 ligand chain entries; 599 individual protein chains)
• ground_truth.tar.gz — 462 mmCIF structures
• protein_deduplicated_chain_map.csv — maps each protein chain to one of 302 deduplicated MSA representatives
• rep_msas_processed.tar.gz — combined MSAs. Some models only take one MSA as input, so this combines the the raw MSAs from different databases.
• rep_msas_raw_01..12.tar.gz — raw search hits; extract all 12 into one directory. This contains the raw MSAs from each database. OB-0 can use multiple MSAs as input

9.3 GB compressed, ~75 GB uncompressed. 

We would like to thank Peter Skrinjar, Janani Durairaj, and Vladas Oleinikovas and the Schwede Lab at Biozentrum for providing access to an updated PLINDER database, curating this updated Runs N’ Poses style benchmark set, and aiding with training set similarity calculations.