maltefranke/morph-data

Morph data

Ready-to-use processed QM9, GEOM-Drugs, ChEMBL3D, and CrossDocked data for Morph. Download the original processed files directly, including train/validation/test splits, ordered vocabularies, training statistics, SMILES sidecars, and saved scaffold cuts. No source-data download or molecular preprocessing is needed to use the published caches with Morph. CrossDocked uses the vocabulary/statistics of the pretrained ligand model, as described below.

This dataset repository is public and supports unauthenticated downloads. It belongs to the Morph collection.

Available datasets

Sizes are decimal GB. GEOM counts are conformer records; QM9 and ChEMBL3D counts are molecule records. CrossDocked counts are ligand/pocket pairs. Sizes include the published scaffold caches.

Directory	Representation / selection	Train	Validation	Test	Size
qm9	Explicit H; original Morph QM9 splits	100,000	20,000	12,947	0.467 GB
geom	GEOM-Drugs, explicit H; at most 72 total atoms	5,645,654	703,830	706,568	76.344 GB
chembl3d	Explicit H; processed cache without the 100-atom cap	1,479,308	184,947	184,960	14.795 GB
chembl3d_max100	Explicit H; at most 100 total atoms	1,474,008	184,295	184,242	14.726 GB
chembl3d_max100_implicit_h	Implicit H; same capped source molecules	1,474,008	184,295	184,242	9.925 GB
crossdocked	Implicit H; DiffDec ligand/pocket preprocessing	96,874	3,107	100	1.895 GB

The ChEMBL3D 100-atom cap counts hydrogens before removal. The implicit-H cache has its own vocabularies/statistics and includes the charge/bond normalization used for the released implicit-H models.

Total distributed data and scaffold manifest: 118,152,947,286 bytes (118.15 GB) in 90 files. The checksum list excludes the dataset card, manifest.json, and SHA256SUMS itself.

Download

From a Morph checkout, download a selected directory into data/:

# Choose one or more datasets:
hf download maltefranke/morph-data --type dataset --include 'qm9/**' --local-dir data
hf download maltefranke/morph-data --type dataset --include 'geom/**' --local-dir data
hf download maltefranke/morph-data --type dataset --include 'crossdocked/**' --local-dir data
hf download maltefranke/morph-data --type dataset --include 'chembl3d_max100/**' --local-dir data
hf download maltefranke/morph-data --type dataset --include 'chembl3d_max100_implicit_h/**' --local-dir data

# Uncapped ChEMBL3D cache:
hf download maltefranke/morph-data --type dataset --include 'chembl3d/**' --local-dir data

For example, QM9 will be available as data/qm9/processed/train.pt, and GEOM as data/geom/processed/train_data.lmdb/data.mdb.

Download everything, including the manifest and checksums:

hf download maltefranke/morph-data --type dataset --local-dir data
(cd data && sha256sum -c SHA256SUMS)

Or use Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="maltefranke/morph-data",
    repo_type="dataset",
    allow_patterns=["qm9/**"],
    local_dir="data",
)

For a reproducible download, pass a dataset commit hash with --revision in the CLI or revision= in Python. These are native Morph caches; load them with Morph's dataset classes rather than datasets.load_dataset.

Use a Morph checkout with support for loading processed QM9 caches without raw inputs. Older QM9 loaders, including those preserved in historical training bundles, may still fetch raw inputs during dataset initialization.

Matching the released checkpoints

Model subfolder in maltefranke/morph	Dataset directory here
qm9/explicit_h	qm9
geom/explicit_h	geom
chembl3d/explicit_h	chembl3d_max100
chembl3d/implicit_h	chembl3d_max100_implicit_h

When using a model's training bundle, point MORPH_DATASET_ROOT to the downloaded dataset directory that contains processed/. For example:

export MORPH_DATASET_ROOT="$PWD/data/chembl3d_max100_implicit_h"

Vocabulary/statistics SHA-256 hashes and split-file sizes were checked against the dataset references in the published training bundles. Those historical references contain split sizes, not split content hashes. This dataset snapshot adds full SHA-256 checksums for every distributed data file in SHA256SUMS and manifest.json.

File layout

Dataset directories contain a processed/ subdirectory with:

• QM9: train.pt, val.pt, test.pt, plus PyG preprocessing metadata.
• GEOM, ChEMBL3D, and CrossDocked: train_data.lmdb/data.mdb, val_data.lmdb/data.mdb, test_data.lmdb/data.mdb.
• QM9, GEOM, and ChEMBL3D: atom_tokens.txt, edge_tokens.txt, charge_tokens.txt, distributions.pt, and train_smiles.txt, val_smiles.txt, test_smiles.txt.
• ChEMBL3D: original preprocessing_summary.json; the capped variants also contain *_source_indices.npy mappings to their source-cache records.
• CrossDocked: crossdocked_summary.json, ordered *_smiles.txt, and ordered *_ligand_fns.txt sidecars. The LMDBs include pocket coordinates and source-pair identifiers alongside ligand tensors.
• Scaffold cuts: versioned LMDBs under qm9/processed/scaffold_groups/, geom/processed/scaffold_groups/, and crossdocked/scaffold_groups/.

Molecular LMDB records contain atomic numbers (z), coordinates (pos), charges, edge_index, edge_attr, and SMILES. Coordinates are stored in physical Angstroms; Morph applies the supplied training statistics when constructing model inputs. QM9, GEOM, and ChEMBL3D SMILES sidecars contain unique split SMILES and are not per-record index maps. CrossDocked sidecars follow record order and retain duplicates. Keep each cache's split files, vocabularies, and statistics together.

Only the LMDB data.mdb files are distributed; the Morph reader uses read-only LMDB access without lock files. Raw inputs, alternative statistics backups, and temporary files are omitted.

Saved scaffold cuts

All three splits are included for every listed kind: 18 scaffold caches in total.

Dataset	Published scaffold kinds
QM9	murcko, heavy
GEOM-Drugs	murcko, libinvent, linker
CrossDocked (implicit H)	libinvent

Scaffold caches store retained atom indices for each source record. Dataset loading constructs training examples from these selections and the matching molecular coordinates. Downloading qm9/**, geom/**, or crossdocked/** includes both the base molecular data and its scaffold caches.

The saved atom selections were preserved from the existing artifacts. QM9/GEOM LMDBs retain their original cuts and auxiliary metadata; the CrossDocked cut lists were converted from PyTorch to LMDB without recomputing cuts. Historical choices can differ from cuts recomputed with the current chemistry implementation. scaffold_manifest.json records each cache's kind, split, source-file SHA-256, total record count, and number of records with at least one cut. Artifact sizes and SHA-256 checksums are in manifest.json and SHA256SUMS.

Use a Morph checkout with portable scaffold cache version 2. Cache identities use source-file contents and dataset settings, so downloading or moving identical files preserves cache reuse. Morph verifies each source split once and shares that checksum across processes and later runs using a locked disk cache. File identity, size, modification time and change time invalidate the local checksum memo when the source changes. A new download/path requires one verification read, which can take minutes for the largest GEOM split. Keep source files immutable while readers are active. The dataset directory can be read-only.

Resources and verified source checksums default to ${XDG_CACHE_HOME:-~/.cache}/morph; set MORPH_CACHE_DIR to choose another writable location shared by your training processes/nodes. These local verification records do not change the portable scaffold-cache identity or the published cuts.

Select a published kind explicitly, for example:

python run.py defaults=qm9/scaffold_decoration \
  model.module.pretrained_ckpt=/path/to/unconditional.ckpt \
  data.scaffold_kind=murcko

The recipe's rgroup default is not among the published cuts and is computed on first use if selected. Evaluation panels are selected on first use; QM9/GEOM panels are also cached. Panels are not included in this snapshot. Other unpublished kinds are likewise built when requested; computing new LibINVENT cuts requires Morph's scaffold extra. The first LibINVENT use automatically downloads the original reaction rules and DiffDec filters from pinned revisions, checks their SHA-256 hashes, and caches them for offline reuse. The downloaded rules are byte-identical to those used for these artifacts. Loading the published cuts does not download rules or require the optional cutter dependency. To prepare the rules before an offline job:

python -c 'from morph.dataset.conditional.scaffold.resources import libinvent_resources; libinvent_resources()'

CrossDocked ligand/pocket data

The published CrossDocked cache uses DiffDec preprocessing with implicit ligand hydrogens, a validation fraction of 0.02 grouped by source directory, and seed 42. crossdocked/processed/crossdocked_summary.json records the upstream revision, original split checksum, processing options, validation groups, and rejected records. The 96,874 / 3,107 / 100 train/validation/test pairs have matching LibINVENT caches:

Split	Pairs with a saved cut	Enumerated cuts
Train	42,251	121,381
Validation	1,032	4,223
Test	31	82

For the released configuration, use the pretrained ligand vocabulary/statistics in chembl3d_max100_implicit_h/processed and an implicit-H ChEMBL3D checkpoint. The saved cuts were prepared with ligand elements B, Br, C, Cl, F, I, N, O, P, S, Se, Si. Pocket elements are C, N, O, S. CrossDocked does not ship a separate set of ligand vocabularies/statistics; the fine-tuning recipe loads these from the pretrained model's dataset. Its unconditional replay also needs the matching ChEMBL3D data.

The processed pocket coordinates are included. Original receptor PDB/SDF files are not included; optional external docking evaluation needs the original source files referenced by the pair identifiers.

Sources and citation

These are processed snapshots of existing molecular datasets. Upstream sources used by Morph's preprocessing include the corrected QM9 structures, the GEOM-Drugs Revisited split files, and LoQI's ChEMBL3D processed release (Figshare file IDs 62280814, 62280805, 62280802). CrossDocked provenance is recorded in its preprocessing summary, using the DiffDec preprocessing implementation in Morph. Consult the upstream releases for their dataset licenses and citation details.

For Morph, please cite Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport:

@misc{franke2026generativemolecularmorphingflexiblesize,
  title={Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport},
  author={Malte Franke and Stefan P. Schmid and Zarko Ivkovic and Kjell Jorner and Andreas Krause},
  year={2026},
  eprint={2606.07239},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2606.07239},
}