| --- |
| pretty_name: MET-dateset |
| tags: |
| - chemistry |
| - molecular-property-prediction |
| - 3d-molecules |
| - lmdb |
| --- |
| |
| # MET-dateset |
|
|
| LMDB dataset release for the MET project: |
|
|
| **Integrating equivariant architectures and charge supervision for data-efficient molecular property prediction** |
|
|
| ## Links |
|
|
| - Paper DOI: https://doi.org/10.1039/D5ME00173K |
| - GitHub repository: https://github.com/mint258/MET |
| - Intended Hugging Face dataset page: https://huggingface.co/datasets/Mint258/MET-dateset |
|
|
| ## Contents |
|
|
| This upload contains single-file LMDB datasets prepared for the MET codebase. |
|
|
| ### QM9 |
|
|
| - `QM9/full_database.lmdb` |
| - `QM9/train_valid_database.lmdb` |
| - `QM9/test_database.lmdb` |
|
|
| ### QM7 |
|
|
| - `QM7/full_database.lmdb` |
| - `QM7/train_database_300.lmdb` |
| - `QM7/test_database.lmdb` |
|
|
| ## Record Format |
|
|
| Each LMDB record stores the molecular graph and associated labels used by MET. The per-record payload includes: |
|
|
| - `z`: atomic numbers |
| - `pos`: 3D atomic coordinates |
| - `y`: main supervised target used by the corresponding pipeline |
| - `scalar_props`: graph-level scalar properties when available |
| - `filename`: source molecule identifier |
| - `chiral_inchi`: stored molecule identifier string when available |
|
|
| Metadata entries are also included, such as `__meta__/property_names`. |
|
|
| ### QM9 Graph-Level Properties |
|
|
| `rot_A`, `rot_B`, `rot_C`, `dipole`, `polarizability`, `HOMO_energy`, `LUMO_energy`, `gap`, `R2`, `zpve`, `U0`, `U298`, `H298`, `G298`, `Cv` |
|
|
| ### QM7 Graph-Level Properties |
|
|
| `atomization_energy` |
|
|
| ## Intended Local Layout |
|
|
| After download, place the files into the MET repository like this: |
|
|
| ```text |
| data/ |
| QM7/ |
| full_database.lmdb |
| train_database_300.lmdb |
| test_database.lmdb |
| QM9/ |
| full_database.lmdb |
| train_valid_database.lmdb |
| test_database.lmdb |
| ``` |
|
|
| ## Usage with MET |
|
|
| Example pretraining command: |
|
|
| ```bash |
| python pretrain/training_charge_model.py \ |
| --data_root data/QM9/train_valid_database.lmdb \ |
| --save_path pretrained_ckpt/best_model_dim128_reproduced.pth |
| ``` |
|
|
| Example evaluation command: |
|
|
| ```bash |
| python pretrain/charge_predict.py \ |
| --checkpoint_path pretrained_ckpt/best_model_dim128.pth \ |
| --test_data_root data/QM9/test_database.lmdb |
| ``` |
|
|
| ## Notes |
|
|
| - These LMDB files are derived from the local QM7 and QM9 inputs used by the MET repository. |
| - Downstream fine-tuning in MET still supports plain-file inputs such as `xyz` directories, manifests, and CSV files with SMILES. |
| - This dataset card is prepared for manual upload to Hugging Face. |
|
|
