3DCS: Datasets and Benchmark for Evaluating Conformational Sensitivity in Molecular Representations

3DCS is a benchmark for 3D conformational sensitivity in molecular representations (MRs). It tests whether the representations of different conformers of the same molecule (i) preserve geometric variation, (ii) capture chirality, and (iii) reflect the energy landscape. This is the Geometry–Chirality–Energy (GCE) evaluation framework from the ICLR 2026 paper.

This repository holds the benchmark's conformer datasets as four named configs (see Configs).

• Paper (ICLR 2026): https://openreview.net/forum?id=JAb0y8lkqL
• Code (evaluation toolkit, Python module three_dbench): https://github.com/ComDec/3DCS
• Standalone copies of each config. The parquet files are byte-identical to the ones here: EscheWang/3dcs-chirality, EscheWang/3dcs-rotation, EscheWang/3dcs-traj-frames, EscheWang/3dcs-traj-energies

Configs

Each config has one split, train, that holds all rows of that config. It is not a training subset.

Config	GCE axis	Rows	Conformers	Parquet files (size)
chirality	Chirality	14,903 (one per stereoisomer of a parent molecule)	52,391	1 (96.0 MB)
rotation	Geometry	1,559,779 (one per molecule)	10,097,643	14 (7.53 GB)
traj_frames	Energy	999,988 (one per frame)	999,988	1 (345.8 MB)
traj_energies	Energy	10 (one per molecule)	999,988 energies	1 (1.7 MB)

All counts in this card were computed from the parquet files in this repository.

Structures are stored as MDL MolBlock strings (V2000) with 3D coordinates.

chirality

This config has 3,903 distinct parent molecules (ChEMBL IDs) and 14,903 stereoisomers. The number of stereoisomers per parent is 1 (61 parents), 2 (1,148), 3 (121), 4 (682) or 5 (1,891).

Field	Type	Description
key	string	Stereoisomer identifier <mol_id>::en<en_id>_<stereo>. <stereo> lists stereocenters as A<atom index>:<R or S> (0-based atom index in MolBlock atom order), separated by ;. Example: CHEMBL100259::en3_A5:S;A7:S;A10:R;A12:S. In 7 rows, <stereo> is achiral. The key is unique per row.
mol_id	string	ChEMBL ID of the parent molecule.
en_id	string	Index of the stereoisomer within its parent molecule, from "0" to "4". It matches en<k> in key.
n_conformers	int64	Number of conformers of this stereoisomer (1 to 20). It equals len(mol_blocks).
offset	int64	Position of this row's first conformer in the flattened conformer order (rows in file order, then mol_blocks order). It equals the sum of n_conformers over all preceding rows, so a flat embedding array of shape (52391, d) is sliced as [offset : offset + n_conformers].
mol_blocks	list<string>	One MolBlock per conformer, with explicit hydrogens.

rotation

Field	Type	Description
key	string	Molecule identifier, unique per row (e.g. 0-R0B0G12_16-R10B3G0_1_23).
shard	int32	Source shard (0–15) of the original generation output. Each of shards 0–14 has 97,487 rows and shard 15 has 97,474. Rows are grouped by shard. The 14 parquet file names (train-000NN-of-00014) do not correspond to shard values.
n_conformers	int32	Number of conformers kept for this molecule (1 to 20). It equals len(mol_blocks) and len(torsion_deg).
offset	int64	Position of this row's first conformer within its shard. It restarts at 0 for each shard value and equals the sum of n_conformers over the preceding rows of the same shard. For a single flat array covering the whole config, use the running sum of n_conformers over all rows instead.
mol_blocks	list<string>	One MolBlock per conformer. Hydrogens are implicit: only 11 of the 10,097,643 MolBlocks contain an explicit H atom.
torsion_deg	list<float32>	Torsion angle in degrees for each conformer, aligned with mol_blocks. Values lie in [-180, 180].

This release stores structures and torsion angles only. Per-conformer xTB energies are not included.

traj_frames

Field	Type	Description
mol_type	string	rMD17 molecule: rmd17_aspirin, rmd17_azobenzene, rmd17_benzene, rmd17_ethanol, rmd17_malonaldehyde, rmd17_naphthalene, rmd17_paracetamol, rmd17_salicylic, rmd17_toluene or rmd17_uracil.
frame_idx	int32	0-based index of the structure in the rMD17 file. It runs from 0 to n_frames - 1, and the rows for each molecule form one contiguous block in frame_idx order.
mol_block	string	MolBlock of the frame, with explicit hydrogens (coordinates in Å).

Each molecule has 100,000 frames, except rmd17_azobenzene, which has 99,988 (the same count as in rMD17).

traj_energies

Field	Type	Description
mol_type	string	Same values as in traj_frames.
n_frames	int64	Number of frames. It equals the number of traj_frames rows for this mol_type.
energies	list<float64>	rMD17 total energy per frame in kcal/mol, aligned with frame_idx. The values are the original rMD17 float64 energies (see Versions).

We checked this data against the official rMD17 .npz files for all 10 molecules and every frame: energies equals the rMD17 energies bitwise (float64), and the MolBlock coordinates match rMD17 coords to within 5e-5 Å (all frames, all molecules), with atoms in rMD17 nuclear_charges order.

Usage

from datasets import load_dataset

chirality = load_dataset("EscheWang/3dcs", name="chirality", split="train")
traj_frames = load_dataset("EscheWang/3dcs", name="traj_frames", split="train")
traj_energies = load_dataset("EscheWang/3dcs", name="traj_energies", split="train")
rotation = load_dataset("EscheWang/3dcs", name="rotation", split="train")  # 7.53 GB of parquet

Rebuild RDKit molecules from MolBlocks with the helper from the 3DCS toolkit:

from three_dbench.datasets.serialization import mol_from_block  # https://github.com/ComDec/3DCS

row = chirality[0]
mols = [mol_from_block(b) for b in row["mol_blocks"]]  # removeHs=False, sanitize=False

Without the toolkit, use the equivalent RDKit call: Chem.MolFromMolBlock(block, removeHs=False, sanitize=False).

The three_dbench CLI reads datasets saved with save_to_disk. The paths below follow the toolkit README and the CLI defaults:

chirality.save_to_disk("data/hf/chirality")
rotation.save_to_disk("data/hf/rotation")
traj_energies.save_to_disk("data/hf/traj/energies")
traj_frames.save_to_disk("data/hf/traj/frames")

python -m three_dbench evaluate chirality \
  --dataset-dir data/hf/chirality \
  --embeddings your_embeddings.npz \
  --embedding-key arr_0 \
  --model-name your_model

See the GitHub README and docs/EMBEDDINGS.md for the trajectory and rotation evaluators and the expected embedding formats.

Provenance

How each dataset was built:

• chirality is derived from ChEMBL (Gaulton et al., 2012), using drug-like molecules with annotated stereocenters. Their stereoisomers were enumerated with RDKit, embedded with ETKDG and geometry-optimized once with xtb --opt lax, giving 15,218 stereoisomers with one optimized geometry each. The 52,391 conformers released here were derived from those geometries: hydrogens were re-added at RDKit's idealized geometry, rotatable bonds were rotated by a small random torsion (median about 10-12 degrees per bond), and isotropic Gaussian noise of sigma about 0.09 A was added to every atom. A conformer was kept only if its heavy-atom RMSD to every already-accepted conformer of the same stereoisomer exceeded 0.350 A, up to 20 per stereoisomer, so stereoisomers with no rotatable bond carry exactly one conformer, as do 34% of the 14,903 stereoisomers overall. Every released conformer reproduces the CIP (R/S) assignment of its source geometry, as re-perceived from 3D with RDKit.
• rotation (the geometry dataset) was generated by the 3DCS authors. It is a combinatorial library of bi-aryl/heteroaryl scaffolds decorated with substituents. RDKit was used for fragment connection, sanitization and ETKDG 3D seeding. Each molecule has an xTB relaxed dihedral scan around the inter-ring single bond in 2.5° increments. Redundant conformers were then removed per molecule with DBSCAN (eps = 0.5).
• traj_frames / traj_energies are derived from the revised MD17 dataset (rMD17) (Christensen & von Lilienfeld, 2020; https://doi.org/10.6084/m9.figshare.12672038).

License

The data in this repository are released under CC BY-SA 4.0. The upstream sources have these licenses:

• ChEMBL data, used for chirality, are licensed under CC BY-SA 3.0.
• rMD17, used for traj_frames and traj_energies, is released under CC0.
• The rotation data were generated by the authors.

The evaluation code in the GitHub repository is licensed separately (MIT). If you use the chirality or trajectory configs, please also cite ChEMBL or rMD17 (see below).

Versions

• Current revision (since 2026-09-18): traj_energies.energies holds the original rMD17 values in float64.
• Revision 40c6cfa829ae5df84346d07d34c715c2924d9b78 and earlier: the same energies cast to float32 (1,096-3,334 distinct values per molecule, at most 0.0156 kcal/mol from the float64 values). Metrics computed on these energies can be sensitive to that precision, so use the current revision.

Baseline embeddings

The embedding files of the baseline models evaluated on these data are published at EscheWang/3dcs-embeddings, together with the per-molecule metric outputs of the original evaluation runs and a manifest.csv listing the key, shape, dtype and SHA-256 of every file. The evaluation toolkit is at https://github.com/ComDec/3DCS. The per-model scripts that extract the chirality baseline embeddings, each with its pinned environment and the SHA-256 of the weights it loads, are in the baselines/ directory of that repository.

Citation

@inproceedings{wang2026threedcs,
  title     = {3{DCS}: Datasets and Benchmark for Evaluating Conformational Sensitivity in Molecular Representations},
  author    = {Wang, Xi and Zhang, Yang and Zhang, Yingjia and Cai, Yejia and Wang, Shengjie},
  booktitle = {The Fourteenth International Conference on Learning Representations (ICLR)},
  year      = {2026},
  url       = {https://openreview.net/forum?id=JAb0y8lkqL}
}

Upstream data sources:

@article{gaulton2012chembl,
  title   = {ChEMBL: a large-scale bioactivity database for drug discovery},
  author  = {Gaulton, Anna and Bellis, Louisa J and Bento, A Patricia and Chambers, Jon and Davies, Mark and Hersey, Anne and Light, Yvonne and McGlinchey, Shaun and Michalovich, David and Al-Lazikani, Bissan and others},
  journal = {Nucleic Acids Research},
  volume  = {40},
  number  = {D1},
  pages   = {D1100--D1107},
  year    = {2012}
}

@article{christensen2020role,
  title   = {On the role of gradients for machine learning of molecular energies and forces},
  author  = {Christensen, Anders S and von Lilienfeld, O Anatole},
  journal = {Machine Learning: Science and Technology},
  volume  = {1},
  number  = {4},
  pages   = {045018},
  year    = {2020}
}