ibm-research/S23D.GEOM.processed

GEOM-DRUGS for S23D

GEOM-DRUGS molecular conformers in the format read by S23D, the conformer generation model from the preprint A standard transformer and attention with linear biases for molecular conformer generation. It holds the training data and the test_1000 benchmark of the released models S23D-S (ibm-research/S23D.8m.v1) and S23D-B (ibm-research/S23D.25m.v1).

Download

The evaluation needs only the benchmark files (212 files, 0.62 GB), fetched here with checksums.sha256:

hf download ibm-research/S23D.GEOM.processed --repo-type dataset \
    --include "zarr/test_1000/*" --include "index/eval_index/test_1000/*" \
    --include checksums.sha256 --local-dir /abs/path/drugs_20
export SMILES23D_DATA=/abs/path/drugs_20

Then run s23dtest as described in the S23D README. The hf command comes with huggingface_hub, which smiles23d installs.

To train as well, download everything (589 data files, 54.7 GB):

hf download ibm-research/S23D.GEOM.processed --repo-type dataset --local-dir /abs/path/drugs_20

The same eval-only download from Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ibm-research/S23D.GEOM.processed",
    repo_type="dataset",
    allow_patterns=[
        "zarr/test_1000/*",
        "index/eval_index/test_1000/*",
        "checksums.sha256",
    ],
    local_dir="/abs/path/drugs_20",
)

Download to a fixed absolute path with --local-dir (local_dir), not into the Hugging Face cache: the cache path changes with every commit to this repo, and a restarted training run expects its data at the same path.

Checksums

checksums.sha256 lists the SHA-256 of every data file. After an eval-only download, check the benchmark files:

cd /abs/path/drugs_20
grep -E '  (zarr|index/eval_index)/test_1000/' checksums.sha256 | sha256sum --check --quiet

After a full download, check every file:

cd /abs/path/drugs_20
sha256sum --check --quiet checksums.sha256

Both checks print nothing and exit with status 0 when every file matches. On macOS use shasum -a 256 --check in place of sha256sum --check. The record decoder can unpickle object arrays, so only read a copy of this data that passes the check.

Layout

The repository root is the directory that SMILES23D_DATA points to:

drugs_20/
  train/                       data*.bin, index.json      4,103,418 conformers, one item each
  val/                         data*.bin, index.json      30,414 molecules, one item each
  test_1000/                   data*.bin, index.json      the 1,000 benchmark molecules
  zarr/test_1000/              .zarray, 0                 the 1,000 benchmark molecules, Zarr v2
  index/eval_index/test_1000/  index.json, chunk_index*   208,304 evaluation entries
  checksums.sha256

Folder	Files	Size	Read by
train/	343	46.8 GB	training
val/	32	5.8 GB	training, when validation is turned on
test_1000/	2	1.44 GB	nothing in the release configs
zarr/test_1000/	2	0.62 GB	evaluation
index/eval_index/test_1000/	210	0.12 MB	evaluation

• train/, val/ and test_1000/ are litdata folders; each index.json lists the chunks and the record schema. val/ and test_1000/ are ordered by atom count.
• test_1000/ holds the same 1,000 molecules (104,152 reference conformers) as zarr/test_1000/, in the training format.
• Each entry of index/eval_index/test_1000 is the position of a molecule in zarr/test_1000, and a molecule with K reference conformers appears 2K times. The two folders only work as a pair.
• The evaluation groups conformers by each record's smiles string, so keep the strings exactly as stored: do not re-key or canonicalise them.

Read the data with the smiles23d package. It was tested with smiles23d v1.0.0, litdata 0.2.30, zarr 3.0.10, posebusters 0.6.5 and RDKit 2026.3.6.

Evaluation

The S23D benchmark generates 2K conformers for a molecule with K reference conformers, with 300 diffusion steps, and scores heavy-atom RMSD after alignment with a coverage threshold of 0.75 Å and no chirality correction. A complete run on this data prints n rmsd calcs: 77763496 and writes 208,304 rows to output_0.jsonl. A lower count means molecules were skipped; because test_1000 is ordered by atom count, a short run skips the largest ones and overstates the metrics.

Results from the preprint for S23D-S-1/9 (S) and S23D-B-1/13 (S). Mean / median:

Model	Recall COV (%) ↑	Recall AMR (Å) ↓	Precision COV (%) ↑	Precision AMR (Å) ↓
S23D-S (ibm-research/S23D.8m.v1)	80.7 / 89.9	0.483 / 0.458	57.5 / 57.5	0.757 / 0.700
S23D-B (ibm-research/S23D.25m.v1)	84.5 / 91.7	0.420 / 0.387	62.5 / 63.4	0.690 / 0.626

Sampling is random, so a rerun will not repeat these numbers exactly.

Training

Training reads only train/, plus val/ when validation is turned on. The released models were trained on this data. Training is not seeded, so a new run gives a different model. The training configs read train/ in stored order; with litdata 0.2.30, 2 GPUs (+task.trainer.devices=2) and the batch_size and num_workers of the released configs, every run sees the records in the same order.

Provenance

1. GEOM (Axelrod and Gómez-Bombarelli, 2022): the GEOM-DRUGS conformer ensembles, from Harvard Dataverse, doi:10.7910/DVN/JNGTDF, under CC0 1.0.
2. GeoMol (Ganea et al., 2021): the train, validation and test split and the 1,000-molecule test set.
3. MCF (Wang et al., 2024): the processed GEOM-DRUGS archives, with at most 20 conformers per molecule in train and validation, all conformers in test, the smi keys, and precomputed eigenvectors (eig_vals, eig_vecs) that S23D does not read.
4. Conversion with the scripts in smiles23d/tools/preprocessing: conformers are grouped into molecules by the smi key, and conformers with a disconnected bond graph or at most one atom are dropped (in test_1000, 106,775 conformers become 104,152). A rebuild from the MCF archives reproduces this content, but not the record order.

License

This dataset is released under the Community Data License Agreement – Permissive, Version 2.0. The GEOM data it is built from is under CC0 1.0. If you use it, please cite GEOM, GeoMol, MCF and S23D.

Citation

@article{axelrod2022geom,
  title   = {{GEOM}, energy-annotated molecular conformations for property prediction and molecular generation},
  author  = {Axelrod, Simon and G{\'o}mez-Bombarelli, Rafael},
  journal = {Scientific Data},
  volume  = {9},
  number  = {1},
  pages   = {185},
  year    = {2022},
  doi     = {10.1038/s41597-022-01288-4}
}

@inproceedings{ganea2021geomol,
  title     = {{GeoMol}: Torsional geometric generation of molecular {3D} conformer ensembles},
  author    = {Ganea, Octavian-Eugen and Pattanaik, Lagnajit and Coley, Connor W. and Barzilay, Regina and Jensen, Klavs F. and Green, William H. and Jaakkola, Tommi S.},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {34},
  pages     = {13757--13769},
  year      = {2021}
}

@inproceedings{wang2024swallowing,
  title     = {Swallowing the Bitter Pill: Simplified Scalable Conformer Generation},
  author    = {Wang, Yuyang and Elhag, Ahmed A. A. and Jaitly, Navdeep and Susskind, Joshua M. and Bautista, Miguel {\'A}ngel},
  booktitle = {Proceedings of the 41st International Conference on Machine Learning},
  series    = {Proceedings of Machine Learning Research},
  volume    = {235},
  pages     = {50400--50418},
  year      = {2024}
}

@misc{gurev2025standard,
  title         = {A standard transformer and attention with linear biases for molecular conformer generation},
  author        = {Gurev, Viatcheslav and Rumbell, Timothy},
  year          = {2025},
  eprint        = {2506.19834},
  archivePrefix = {arXiv},
  primaryClass  = {q-bio.BM},
  url           = {https://arxiv.org/abs/2506.19834}
}