kfold-affinity-labels

Upstream binding-affinity databases, redistributed as downloaded, as the label sources for the protein-ligand side of hsjang0/k-fold-multimer-tuning.

Kept in its own repo rather than folded into hyosoon0/kfold-multimer-lmdb because these are third-party databases under their own licenses and citation requirements, while that repo holds data we derived. One license statement should not have to cover both.

Contents

	size
chembl/chembl_37_sqlite.tar.gz	5.4 G	ChEMBL 37, full SQLite release
bindingdb/BindingDB_All_202608_tsv.zip	566 M	BindingDB, August 2026 full TSV
mf_pcba/	1.2 G	MF-PCBA, 50 retrieved assays, AID*/ with SD.csv, DR.csv and metadata
pubchem/*.zip	6.7 G	2013 PubChem BioAssay archives, named by AID range

pubchem/ is stored in six commits rather than one. A single commit at that file count returns 504 from the hub, and although the commit does land, the client cannot tell.

The .git directories inside the cloned upstream repos (mf_pcba/mf-pcba) are excluded; they carry nothing the data needs and one of them held a 686 MB pack.

Licenses and citation

These are not ours. Each carries its own terms and each expects its own citation.

• ChEMBL is CC BY-SA 3.0. Cite the current ChEMBL database paper.
• BindingDB is CC BY 3.0 US. Cite Gilson et al., Nucleic Acids Research.
• PubChem BioAssay is public domain as US government work, with the caveat that individual depositor records may carry their own terms. Cite Kim et al.
• MF-PCBA follows its own repository's license. Cite the MF-PCBA paper.

Check each upstream source's current terms before redistributing further. The copies here are a snapshot at the dates in the filenames, not a maintained mirror.

What this is not

No labels were computed here. This is the raw upstream material; the derived affinity tables that training actually reads are built by the scripts listed in the GitHub repo's CODEMAP.md and described in AFFINITY_LABELS.md.