Felixb7/geom-drugs-top1-rem3di

GEOM-Drugs top-1 conformers — the Rem3Di pretraining corpus

The exact molecule set used to pretrain the Rem3Di descriptor models (arXiv:2607.19977): one conformer per molecule, chosen by highest Boltzmann weight.

281,071 conformers · 278,679 unique molecules · 12,519,313 atoms

Contents

geom_drugs_top1.tar unpacks to a zarr store consumed by remedi's MoleculeDataset:

geom_drugs_top1/
├── positions/         3D coordinates
├── atomic_numbers/    elements
├── molecule_ptr/      per-molecule atom offsets
├── ids/               molecule identifiers
├── total_charge/      zero-filled (corpus is neutral)
├── total_spin/        zero-filled (closed shell)
└── dataset_config.yaml

Geometry only — no precomputed MLIP features. Rem3Di computes MACE features online at every training step, so the backbone is a runtime choice rather than baked into the store.

from remedi.data_handling.dataset.molecule_dataset import MoleculeDataset
ds = MoleculeDataset.open_existing_dataset_from_dir("geom_drugs_top1")

How it was built

1. Raw GEOM-Drugs pickles (304,340)
2. Keep conformers with Boltzmann weight > 0.01, max 128 atoms → 2,710,989 conformers
3. Top-1 by Boltzmann weight per (molecule_id, isomer_id, atomic_numbers) group → 281,071

Filters, all at defaults: elements restricted to {H, C, N, O, F, P, S, Cl, Br, I} (verified against the store); 3–128 atoms, actual min 7, max 128, mean 44.5; neutral only; single fragment, so salts rejected; no radicals; no isotopes. No canonical-SMILES dedup in this lineage.

⚠️ Two things that will not reproduce if you rebuild from GEOM-Drugs

Single-conformer molecules were dropped. The subset builder contains if len(members) < 2: continue, so any molecule with only one conformer above the Boltzmann threshold was excluded entirely. That is why the count is exactly 281,071. Rebuilding "top-1 GEOM-Drugs" the obvious way gives a different, larger set.

The train/val split seed is hardcoded to 1, and ignores any --seed argument. The split is molecule-level random 0.95/0.05 → 267,007 train / 14,064 val. Not scaffold.

Both are recorded here because neither is documented in the code that produced them.

Pretraining recipe this was used with

24 epochs · AdamW · LR 1e-4 · weight decay 1e-3 · denoising noise σ 0.3 · batch 40 · grad-norm clip 1.0. About 86 min/epoch on an RTX 4090.

Reference: a healthy MACE-POLAR run reaches epoch-24 denoising val loss ≈ 5.0×10⁻⁴.

Provenance and licence

Derived from GEOM (Axelrod & Gómez-Bombarelli, Sci Data 2022), released CC0. This is a filtered subset with no new experimental data, so it carries the same terms. Please cite GEOM as well as Rem3Di.

@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}
}