# 3DReact data

## Cite
- [J. Chem. Inf. Model. 2024, 64, 5771−5785](https://doi.org/10.1021/acs.jcim.4c00104)
- <https://github.com/lcmd-epfl/EquiReact>
- <https://wandb.ai/equireact>
- materials cloud

## Data

Each dataset (`gdb/` for GDB7-22-TS [1],
`cyclo/` for Cyclo-23-TS [2],
`proparg/` for Proparg-21-TS [3,4]) directory contains:

* `xyz/` — the original (DFT) geometries.
* `xyz-xtb/` — GFN2-xTB geometries.
* `{dataset}.csv` — the CSV file that contains:
  * `idx` / `rxn_id` / (`mol`,`enan`): reaction indices used to find the corresponding xyz files.
  * `dE0` / `G_act` / `Eafw`: target property.
  * `rxn_smiles`: unmapped reaction SMILES
  * `rxn_smiles_mapped`: the original ("true") atom-mapped SMILES
  * `rxn_smiles_rxnmapper`: SMILES mapped by RXNMapper [5]
  * `rxn_smiles_rxnmapper_full`: SMILES mapped by RXNMapper including hydrogens
  * `bad_xtb`: is the reaction is excluded from the geometry quality tests (xTB optimization failed)

Additionally,
  * `proparg/proparg-weird-smiles.csv`: "bad" SMILES for Proparg-21-TS automatically obtained from xyz
    taken from [6].
    They are also mapped by RXNMapper but were not used to produce the results of the paper.

## References
- [1]: [Sci. Data 2022, 9, 417](https://doi.org/10.1038/s41597-022-01529-6)
- [2]: [Sci.  Data 2023, 10, 66](https://doi.org/10.1038/s41597-023-01977-8)
- [3]: [ACS Catal. 2016, 6, 7948−7955](https://doi.org/10.1021/acscatal.6b02366)
- [4]: [Chem. Sci. 2021, 12, 6879−6889](https://doi.org/10.1039/D1SC00482D)
- [5]: [Sci. Adv. 2021, 7, eabe4166](https://doi.org/10.1126/sciadv.abe4166)
- [6]: [Digital Discovery, 2024, 3, 932–943](https://doi.org/10.1039/d3dd00175j)
      ([github repo](https://github.com/lcmd-epfl/benchmark-barrier-learning/))