TopPDBLX: a parsed, normalised and sequence-linked database of crystallisation conditions from the Protein Data Bank

Every crystallisation condition in the Protein Data Bank, parsed into typed components and linked to the sequence that produced it. The PDB holds about 200,000 crystallisation recipes, each typed free-hand by a different scientist in no agreed format. TopPDBLX turns the free-text _exptl_crystal_grow.pdbx_details field into structured components (reagent, concentration, unit, role), cross-references them against published commercial screen formulations, and attaches MMseqs2 cluster identifiers so redundancy can be controlled. Contents 199,185 records from 198,691 PDB entries, keyed on (pdb_id, crystal_id) 645,656 typed components, 87.3% resolved to a canonical reagent from a curated 500-reagent lexicon 195,985 records with a linked protein sequence, carrying UniProt accessions and MMseqs2 cluster ids at 30%, 50% and 90% identity 81,802 records matched to a well in one of 9 commercial screens, extracted verbatim from vendor support materials Seven-class precipitant ontology; 152,006 conditions classified, the remainder Unclassified with the reason recorded Provenance Every component carries a parser field. 610,076 come from a deterministic rule parser; 35,580 were recovered by a LoRA fine-tune of SmolLM2-360M reading the residual the rules could not parse. Filter to parser = 'rules' for a fully reproducible subset containing no model output. Quality Measured against 192 hand-labelled records: 93.3% precision and 90.3% recall for reagent identity, against the rule parser alone at 95.1% and 69.5%. Archive fidelity is 100%: every condition string was byte-compared against the source mmCIF. Inferred units, inferred cryoprotectant roles and pH attribution are flagged rather than presented as fact. Limitations Condition frequency reflects screen popularity, not intrinsic success rate. Reported conditions are often optimised rather than screen hits. Only successful crystallisations are represented: there are no negative examples. See DATASHEET.md for the full set. Code: MIT. Data: CC-BY-4.0. Attribution does not discharge the obligation to the sources this derives from: the Protein Data Bank (CC0), SIFTS and UniProt.