wasarou/plinder-cistrans-209

PLINDER cis/trans conformer benchmark (209 systems)

A curated 209-system subset of PLINDER for measuring whether a structure predictor preserves the E/Z (cis/trans) configuration of a ligand double bond. Each system ships everything needed to fold it and score the result: the protein MSA, the ligand SMILES, and the crystal ligand as an SDF.

It is the evaluation set for the cistrans term of RGI (Restraint-Guided Inference), but nothing here is RGI-specific — any method that outputs ligand coordinates can be scored.

What makes it curated

Of 484 raw PLINDER systems with a cis/trans-capable ligand, 275 cannot be scored at all and are excluded. A double bond is benchmarkable only when it is restrained (acyclic and non-aromatic) and hydrogen-decidable (its E/Z is fixed by heavy atoms alone).

bucket	n	why
no restrained + decidable E/Z bond	240	mostly guanidine/amidine (N=C(N)N, ~168): which bond is double depends on an N–H position a structure predictor never outputs. Plus ring-only E/Z (macrolides/macrolactams, ~70) — see the limitation below.
≥1 restrained but undecidable bond	35	tautomerizable (enamines, free oximes/amidoximes): an exchangeable H shifts the double bond, so the tautomer is equally valid. Scoring is conjunctive, so one such bond would mis-score the whole ligand.
invalid a3m/sdf or unparseable SMILES	0	the 484 → 209 cut is purely the E/Z filter
kept	209	≥1 acyclic, H-decidable E/Z bond and no restrained-undecidable one

Oxime ethers, N-substituted Schiff bases and azo bonds are kept: no mobile H, so heavy-atom coordinates fix E/Z outright.

The two criteria address separate problems—bond topology and hydrogen placement—and were applied independently.

Known limitation: macrocycles

Ring double bonds are excluded because the reference implementation treats them as unable to isomerise — which is false for macrocycles. A 14–16-membered macrolide's C=C has a real E/Z that this set does not test; 70 of the 484 are exactly that.

This is a limitation of the restraint engine, not a curation preference. If you are evaluating a method that can handle macrocyclic E/Z, this set under-tests it, and those 70 systems are a reasonable extension.

Layout

path	what
index.csv	209 rows: sys_id, pdbid, ligid, smiles, release_date, + payload sha256/size
msa/<sys_id>.a3m.xz	ColabFold/MMseqs2 MSA, xz-compressed (~171 MB total; ~850 MB raw)
sdf/<sys_id>.sdf	crystal ligand from PLINDER, re-keyed to sys_id

Row order is meaningful and matches the published index.csv — it is frozen from the original draw, and downstream batch packing is positional, so re-sorting changes batch assignment.

The *_sha256 / *_bytes columns are checksums of the decompressed a3m and of the sdf, so a consumer can verify integrity after materializing.

Usage

The dataset is self-describing: read index.csv, xz-decompress the a3m, and read the SDF with RDKit. Any method that outputs ligand coordinates can be scored against the crystal ligand's E/Z configuration.

Provenance and licensing

Released under CC-BY-4.0. Attribution is required for:

• PLINDER (CC-BY-4.0) — the systems and the crystal ligand SDFs. Release 2024-06, iteration v2.
• RCSB PDB — the underlying experimental structures.
• UniProt / UniRef100 (CC-BY-4.0) and the MGnify protein database (CC0-1.0) — the MSA sequences.
• ColabFold / MMseqs2 — the MSA search.

The MSAs were generated with ColabFold/MMseqs2 against UniRef100 + MGnify. No ChEMBL-derived annotation is included, so PLINDER's CC-BY-SA'd ChEMBL subset does not apply here.

The filter used RDKit 2026.03.4 for stereo and ring perception. If the selection rule changes—for example, if macrocyclic E/Z becomes supported—the current set is superseded rather than silently redefined.