GEOM-QM9: Conformers (Mirrored)

Note: This is a mirrored and lightly sanitized version of the GEOM-QM9 dataset. It contains a single 3D conformation for the molecules originally defined in the QM9 dataset. Credit for the original molecule compilation belongs to Ramakrishnan et al., while credit for the conformational sampling belongs to Axelrod and Gómez-Bombarelli. This repository exists to guarantee availability and exact reproducibility for downstream machine learning projects.

Dataset Description

This dataset applies the Geometric Ensemble Of Molecules (GEOM) methodology to the famous QM9 dataset. While the standard GEOM-QM9 provides an ensemble of multiple energetically annotated conformers for these molecules, this specific repository has been filtered for simplicity.

Preprocessing Note (Single Conformer): Unlike the standard GEOM-QM9 which contains conformational ensembles, this version has been filtered down to exactly one conformer per molecule (e.g., the lowest energy conformer). Therefore, while it structurally resembles the original QM9 format (a 1:1 molecule-to-conformer ratio), the 3D geometries and calculated energies are derived from the GEOM computational pipeline, not the original QM9 level of theory.

• Original GEOM Paper: GEOM, energy-annotated molecular conformations for property prediction and molecular generation
• Original QM9 Paper: Quantum chemistry structures and properties of 134 kilo molecules
• Canonical Data Source (GEOM): learningmatter-mit/geom

---

Data Structure & Organization

To maximize flexibility and reproducibility, this repository is organized with pre-computed splits alongside the full dataset. Because this dataset only requires a single version, all files and folders are located directly at the root level.

Directory Layout

At the root of the repository, you will find the unsplit (qm9_*) dataset files, alongside pre-computed train/ and test/ subdirectories:

geom_qm9/
├── train/
│   ├── train_confs.sdf
│   ├── train_counts.csv
│   └── train_smiles.smi
├── test/
│   ├── test_confs.sdf
│   ├── test_counts.csv
│   └── test_smiles.smi
├── qm9_confs.sdf      <-- Unsplit dataset
├── qm9_counts.csv     <-- Unsplit counts
└── qm9_smiles.smi     <-- Unsplit SMILES

File Types

Each split (and the unsplit root) contains three complementary files representing the exact same pool of molecules:

1. *_confs.sdf: The main Structure-Data File containing the 3D molecular conformers.
2. *_counts.csv: A summary file mapping each unique SMILES string to the number of conformers it has in the corresponding .sdf file (which will always be 1 in this specific dataset).
3. *_smiles.smi: A strict 1:1 text file containing the SMILES string for every single entry in the .sdf file (useful for fast, sequential parsing or redundancy checks).

---

Train / Test Splitting Strategy

The train/ and test/ subfolders contain a standard 80/20 split of the data.

Because this version contains exactly one conformer per unique SMILES string, the split inherently guarantees that there is absolutely zero molecular overlap between the training and testing sets, ensuring robust validation of model generalizability.

If you prefer to use alternative splitting strategies (e.g., scaffold splits or specific property cutoffs), you can ignore the subdirectories and build custom splits using the unsplit files located at the root of the repository.

---

Citation Information

If you use this dataset, you must cite both the GEOM creators (for the conformers) and the original QM9 creators (for the base molecules):

@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},
  publisher={Nature Publishing Group UK London}
}

@article{ramakrishnan2014quantum,
  title={Quantum chemistry structures and properties of 134 kilo molecules},
  author={Ramakrishnan, Raghunathan and Dral, Pavlo O and Rupp, Matthias and Von Lilienfeld, O Anatole},
  journal={Scientific Data},
  volume={1},
  number={1},
  pages={1--7},
  year={2014},
  publisher={Nature Publishing Group}
}