GEOM-QM9 & GEOM-DRUGS preprocessed conformer pickles (Align Your Structures, ICLR 2026)

Preprocessed GEOM-QM9 and GEOM-DRUGS conformer pickles used for conformer-level diffusion pretraining (Stage 1) in Align Your Structures: Generating Trajectories with Structure Pretraining for Molecular Dynamics (ICLR 2026). Contents (6 pickles, ~3 GB uncompressed): GEOM-QM9_Train.pkl — QM9 train split GEOM-QM9_Val.pkl — QM9 validation split GEOM-QM9_Test_Actual_compat.pkl — QM9 test split (ConfGF 200-molecule official benchmark, repackaged) GEOM-DRUGS_Train.pkl — DRUGS train split GEOM-DRUGS_Val.pkl — DRUGS validation split GEOM-DRUGS_Test_Actual_compat.pkl — DRUGS test split (ConfGF 200-molecule official benchmark, repackaged) Format: each pickle is a Python list of dicts with keys atom_type, boltzmannweight, edge_index, edge_type, idx, nx, pos, rdmol, smiles, totalenergy. Same fields as GeoDiff's PyTorch-Geometric Data format, repackaged as dicts to decouple loading from PyG version drift. Provenance: train / val derived from the GeoDiff preprocessed GEOM archive (MinkaiXu/GeoDiff), itself built on top of the ConfGF preprocessing pipeline (DeepGraphLearning/ConfGF). Test split is ConfGF's official 200-molecule benchmark. Upstream raw GEOM: Axelrod & Gómez-Bombarelli, Scientific Data 2022, doi:10.7910/DVN/JNGTDF. Extraction: tar xf align-your-structures-conformer-pkls-v1.tar.gz -C ${MD_DATA_ROOT}/ The archive expands into processed_input_data/GEOM-{QM9,DRUGS}/, matching the paths referenced in the configs_official/ YAMLs. Reference code: https://github.com/ani11452/Align_Your_Structures If you use this data please cite the paper above and the upstream GEOM dataset.