nitroprusside (15952d0a50480e89) is the only compound RDKit could not embed. Its 5 BioSNAP rows are now removed, so every pair left in the csv files has both a drug conformer and an AlphaFold 3 protein structure. It is still listed in drugs.csv with has_conformer = False.
AF3 structures 4517/4528: 78 more proteins (up to 3661 residues), csvs refiltered
Finer inference buckets plus GPU memory preallocation fixed the earlier out-of-memory failures, and unified memory covered the 4096 bucket, so 78 of the 89 missing proteins now have a structure. Only the 11 longest (4128-7073 residues) are still missing, so the csvs lose 20/50/32 rows instead of 709/338/104. README statistics updated.
12 days ago
Final AF3 structures (4439/4528); drop pairs of the 89 proteins without a structure; add protein_ca.npz and drug_conformers.sdf.gz
protein_ca.npz: Cα coordinates + pLDDT of every structure. drug_conformers.sdf.gz replaces sdf.tar.gz; sdf_status.csv is merged into drugs.csv. README updated.
Rename to "novel target benchmark"; drop superseded 202-row table