Datasetpublic

Dellboy/toppdblx-conditions

TopPDBLX v1.0.0 Every crystallisation condition in the Protein Data Bank, parsed, normalised and linked to the sequence that produced it. Citable archive: 10.5281/zenodo.21807134 · Code: bellcheddar/TopPDBLX The PDB holds about 200,000 crystallisation recipes, each typed free-hand in no agreed format. This dataset turns the free-text _exptl_crystal_grow.pdbx_details field into typed components (reagent, concentration, unit, role), cross-references them against published… See the full description on the dataset page: https://huggingface.co/datasets/Dellboy/toppdblx-conditions.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face

0likes27downloads

Dataset Card

TopPDBLX v1.0.0

Every crystallisation condition in the Protein Data Bank, parsed, normalised and linked to the sequence that produced it.

Citable archive: 10.5281/zenodo.21807134 · Code: bellcheddar/TopPDBLX

The PDB holds about 200,000 crystallisation recipes, each typed free-hand in no agreed format. This dataset turns the free-text _exptl_crystal_grow.pdbx_details field into typed components (reagent, concentration, unit, role), cross-references them against published commercial screen formulations, and attaches MMseqs2 cluster identifiers so redundancy can be controlled.


Records	199,185 from 198,691 PDB entries, keyed on `(pdb_id, crystal_id)`
Components	645,656, of which 87.3% resolve to a canonical reagent
Reagent lexicon	500 curated reagents, 1,587 spellings
Sequence linkage	195,985 records with UniProt accessions and clusters at 30/50/90% identity
Screen matches	81,802 records matched to a well in one of 9 commercial screens
Classification	152,006 conditions in a seven-class precipitant ontology

Provenance: every component says which parser produced it

`parser`	rows	what it is
`rules`	610,076	A deterministic rule parser. Filter to this for a fully reproducible subset with no model output
`slm`	35,580	Chemistry the rules left unnamed, recovered by a LoRA fine-tune of SmolLM2-360M reading the residual

Against 192 hand-labelled records the combination reaches 93.3% precision and 90.3% recall for reagent identity, against the rule parser alone at 95.1% and 69.5%.

Read this first

The PDB contains only successes. Every condition here produced a crystal, so this supports P(condition | crystallised) and says nothing about P(crystallised | sequence). Condition frequency reflects screen popularity, not intrinsic success rate, and reported conditions are often optimised rather than screen hits. DATASHEET.md carries the full set of limitations, and reading it before drawing conclusions is not optional.

Files

File	Contents
`toppdblx-conditions-v1.0.0.parquet`	Record level, sequence linkage flattened on
`toppdblx-components-v1.0.0.parquet`	One row per reagent
`toppdblx-components-v1.0.0.csv.gz`	The same, for spreadsheets
`schema-v1.0.0.json`	JSON Schema, generated from the pydantic model
`DATASHEET.md`	Datasheet, generated from the data at release time

Data is CC-BY-4.0. Attribution does not discharge the obligation to the sources this derives from: the Protein Data Bank (CC0), SIFTS and UniProt.