A training corpus is only as useful as its provenance—and molecular structure data can be especially difficult to audit once sources, annotations, and derived records are mixed. This collection’s notable contribution is to preserve the path from sampling metadata to structure shards, making reproducibility and source-level review part of the dataset rather than an afterthought.
What Makes It Useful
- More than protein monomers: Its 19 named configurations cover interfaces, protein–ligand systems, glycans, modified residues, RNA, cyclic peptides, and predicted complexes. That breadth matters for training models intended to work across varied molecular contexts.
- Metadata and structures are separated: Parquet tables describe sampling and indexing, while compressed mmCIF records live in binary shards. Researchers can inspect metadata before downloading large structure components.
- Training provenance is unusually explicit: The repository records source filters, recipe mixture weights, processing evidence, and component-level qualifications. This supports more informed replication and data audits.
- The scale calls for careful interpretation: The published store is about 2.13 TB, but rows across configurations are overlapping sampling views—not counts of unique structures or independent examples.
Who It’s For—and the Trade-Off
A good fit for teams reproducing RFD4 training, inspecting biomolecular training sources, or building compatible training pipelines. Look elsewhere if you need a ready-made evaluation split, uniformly experimental structures, or a small, simple download: predicted records are not experimental measurements, upstream source terms still apply, and target overlap must be audited separately. The added annotations and packaging are CC BY 4.0; that license does not replace the terms of the underlying structure sources.
