dargason/ADMET-CYP-cofolding

OpenADMET CYP PDB cofolding structures v1

This is a coordinate-complete, one-row-per-structure release for the current UniProt PDB cross-references of human CYP1A2, CYP2C9, CYP2D6, and CYP3A4. It contains 16,739 rows: 149 aligned experimental ground truths and 16,590 valid aligned cofolded structures.

Every coordinate file is gzip-compressed mmCIF and every row points to exactly one relative coordinate_path. All structures use the documented cyp3a4_1tqn_conserved_scaffold_ca_v1 frame: a proper Kabsch rotation over 187 conserved CYP scaffold C-alpha anchors, with at least 180 anchors and every block required. The same rigid transform is applied to every atom; ligand, heme, and pocket atoms are never fitted independently.

Method coverage

Coverage below is against the 141 ligand-bearing PDB IDs. The other eight experimental structures are apo/heme-only and remain in the dataset with ligand QC marked not_applicable_no_selected_ligand.

Method	Rows	Ligand PDB IDs	Coverage
boltz2	2,600	130/141	92.2%
chai1	2,780	139/141	98.6%
esmfold2	5,650	113/141	80.1%
openfold3	2,780	139/141	98.6%
protenix_v1	2,780	139/141	98.6%

The union of generated methods covers 139/141 ligand PDB IDs; 4WNU and 4WNV have no valid generated ligand complex. ESMFold2 was run only for CYP3A4. Protenix v2 seed outputs and raw apo cofold attempts are outside this v1 core because they have not passed the same alignment/QC protocol.

Row-level derived data

data/structures.parquet contains identifiers and hashes, ligand chemistry and role fields, alignment metrics, native method confidence, OpenStructure lDDT-PLI where assignment succeeded, training-window labels, ProLIF interaction counts and contact JSON, and the complete PoseBusters dock report (all raw report fields are prefixed posebusters_). Null lDDT-PLI values with ost_status are retained; they are not silently filtered.

ProLIF and PoseBusters use the selected CCD SMILES as authoritative ligand topology, element/graph-isomorphism atom mapping, and the coordinate pose without numerical movement. ProLIF uses a sequence-template RDKit protein graph and canonical UniProt residue numbering. PoseBusters uses the CYP protein chain only as mol_cond; heme, waters, additives, and other cofactors are deliberately excluded from the receptor condition. Ir and Ru use explicit 2.00 and 2.05 A ProLIF van der Waals radii. Legacy PDB CONECT records are used when distance inference cannot recover a deposited organometallic graph.

The automated 6CSB nearest-heme choice was role-curated: the experimental ground truth uses complete RTZ chain A residue 602 rather than a partial HEGA-10 (2CV) detergent fragment. Existing generated 2CV structures are retained because they are real campaign outputs, but they carry ligand_role=crystallization_additive_automated_selection and ost_target_role_valid=false.

Quality states

ProLIF and PoseBusters executed successfully for all 16,731 selected-ligand rows; the eight apo/heme-only rows are explicitly not applicable. PoseBusters reports an all-check scientific pass for 4,212 rows and at least one failed check for 12,519 rows. These are scientific results, not execution statuses.

• qc_status=complete: ProLIF and PoseBusters both executed. A complete row can still fail one or more PoseBusters scientific checks; inspect posebusters_overall_pass and posebusters_failed_checks_json.
• qc_status=not_applicable_no_selected_ligand: experimental apo/heme-only structure.
• ost_status records lDDT-PLI assignment or the reason no value is available.
• No invalid/silent ligand-omission coordinate is included. The 40 excluded Chai-1 attempts are listed in provenance/excluded_attempts.parquet.

Files and verification

• data/structures.parquet: canonical one-row-per-structure table.
• structures/experimental/: 149 rigidly aligned PDB ground truths.
• structures/predicted/: 16,590 verified aligned predictions.
• provenance/: coverage, exact transforms, anchor map, protocols, excluded attempts, field groups, and checksums.
• scripts/verify_release.py: full hash, schema, row/path, and coordinate-parse verification.

Run python scripts/verify_release.py . from the dataset root. provenance/SHA256SUMS covers every published file except itself and the post-verification provenance/verification.json report.

Scope and limitations

Training-window fields classify PDB release dates relative to documented method cutoffs; they do not prove training-set membership. Prediction method names are campaign labels; exact predictor commit and container digests were not consistently captured in the source manifests. Ligand role selection is mostly an automated nearest-heme heuristic, with the explicit 6CSB correction above. PoseBusters internal-energy checks can be unavailable/fail for organometallic ligands because UFF lacks parameters; this is preserved as a failed scientific check, not treated as a pipeline failure. OpenADMET-generated structures, derived tables, documentation, and release scripts are provided under Apache-2.0. Experimental PDB coordinate data retain their upstream CC0 status; see LICENSES.md.

Citation

Use CITATION.cff for this dataset snapshot and cite the original PDB entries and structure-prediction methods appropriate to your analysis.