Kaipengm2/CrystalX-dataset

CrystalX Dataset

This repository provides the preprocessed CrystalX dataset for training and dataset-level evaluation.

Files

• equiv_pt.zip: compressed archive containing the equiv_pt/ folder
• splits/sorted_by_journal_year.txt: publication-year metadata used for splitting
• LICENSE

The year metadata covers 1994-2024.

Data Format

After extracting equiv_pt.zip, each equiv_*.pt file is a Python dictionary saved with torch.save.

The input z and pos describe the coarse electron-density peaks obtained from SHELXT initial phasing. pos is read from the initial phasing result and converted directly into Cartesian coordinates of the electron-density peaks; entries are ordered from strongest to weakest peak. z is an input peak descriptor used to encode the relative strength or initial chemical interpretation of those coarse density peaks, not the final supervised atom label. In practice, z can be assigned from the known or estimated elemental composition ordered from heavier/stronger-scattering elements to lighter/weaker-scattering elements, or it can use the initial element assignment produced by SHELXT.

Main fields:

• z: input peak descriptors or initial element guesses aligned with pos
• equiv_gt: heavy atoms with symmetry-expanded context
• gt: heavy-atom labels
• hydro_gt: hydrogen-count labels
• pos: Cartesian coordinates of coarse electron-density peaks with shape [N, 3] in angstrom, ordered from stronger to weaker peaks

Other auxiliary fields may also be present.

Split File

Each line in splits/sorted_by_journal_year.txt has this format:

year timestamp cif_name

The current CrystalX code uses:

• column 1 as the publication year
• column 3 as the CIF stem

Usage

Extract equiv_pt.zip first. This creates the equiv_pt/ folder.

python -m crystalx_train.trainers.trainer_heavy \
  --pt_dir equiv_pt \
  --txt_path splits/sorted_by_journal_year.txt

Source

The dataset is derived from the open-access Crystallography Open Database (COD) and CrystalX preprocessing.

Citation

If you use this dataset, please cite the CrystalX paper and the COD.

@article{doi:10.1021/jacs.5c21832,
  author  = {Zheng, Kaipeng and Huang, Weiran and Ouyang, Wanli and Zhong, Han-Sen and Li, Yuqiang},
  title   = {CrystalX: High-Accuracy Crystal Structure Analysis Using Deep Learning},
  journal = {Journal of the American Chemical Society},
  doi     = {10.1021/jacs.5c21832}
}