Morph: 3D molecular generation and scaffold decoration

Checkpoints for unconditional 3D molecular generation on QM9, GEOM-Drugs and ChEMBL3D, plus GEOM-Drugs scaffold decoration and CrossDocked scaffold decoration in protein pockets. The releases include the configurations, vocabularies and training distributions needed for their inference workflows. Unconditional generation requires no training dataset; scaffold decoration uses a supplied scaffold and the dedicated scaffold bundle.

See our paper, Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport.

Unconditional checkpoints

Dataset	Model name	Hydrogen representation	Completed epochs	Files
QM9	qm9-explicit-h	Explicit H atoms	1,000, 1,500	Bundle
GEOM-Drugs	geom-explicit-h	Explicit H atoms	50, 100	Bundle
ChEMBL3D	chembl3d-explicit-h	Explicit H atoms	200, 300, 400	Bundle
ChEMBL3D	chembl3d-implicit-h	Heavy atoms; implicit H	100, 200, 300	Bundle

Epoch numbers count completed training epochs: epoch_050.ckpt means 50 completed epochs and stores the zero-based Lightning counter epoch=49. Checkpoints are original training files containing current weights, EMA weights and optimizer state.

The catalog is the source of truth for unconditional model names and epochs. Each entry pins an immutable release revision so its weights and supporting files stay together. New catalog entries are discovered without updating the Python package.

Scaffold decoration

GEOM-Drugs scaffold decoration is also available: the geom-scaffold-murcko bundle contains the checkpoint after 25 fine-tuning epochs, initialized from the GEOM explicit-H epoch-100 EMA weights. It uses rank-16 LoRA, trainable output heads and substituent-constrained insertion supervision.

Model	Hydrogen representation	Completed fine-tuning epochs	Checkpoint
geom-scaffold-murcko	Explicit H atoms	25	epoch_025.ckpt

The model generates decorations conditioned on a Murcko scaffold. Scaffold atom identities, internal bonds and charges are preserved; scaffold positions can move. This conditional checkpoint requires the scaffold bundle's saved configuration and runtime. It is loaded through that workflow, outside the unconditional morph generate interface and catalog.

The bundle includes the original checkpoint, training configuration and source, portable configuration, vocabularies, distributions, evaluation runtime and fixed validation/test scaffold panels. See its usage and reproduction instructions.

Scaffold decoration in a protein pocket

crossdocked-scaffold is the evaluated CrossDocked checkpoint after 100 fine-tuning epochs. It uses implicit ligand hydrogens, rank-8 LoRA and 10% in-vacuum ChEMBL3D replay.

Model	Hydrogen representation	Completed fine-tuning epochs	Checkpoint
crossdocked-scaffold	Heavy atoms; implicit H	100	epoch_100.ckpt

The bundle contains the evaluated checkpoint, portable configuration, vocabulary/statistics assets and complete metrics table. Use the pocket evaluation workflow with a scaffold and protein pocket. Like the GEOM scaffold model, it is separate from the unconditional CLI/catalog.

Evaluation results

Values are mean ± sample standard deviation across seeds 42, 43 and 44 (ddof=1). Rates are percentages; diversity and similarity are on a 0–1 scale. Vina scores are in kcal/mol. All evaluations use EMA weights and 100 integration steps.

Unconditional generation — all 10 checkpoints

10,000 generation attempts per seed and checkpoint (30,000 per checkpoint). PoseBusters uses the full offline mol checks, including flatness and internal energy, on the generated structures.

Model	Completed epochs	Novelty (%)	PoseBusters validity (%)	Diversity (0–1)	Uniqueness (%)	Connected-sanitization validity (%)
QM9 · explicit H	1,000	24.99 ± 0.50	93.19 ± 0.30	0.8969 ± 0.0003	94.77 ± 0.19	98.99 ± 0.14
QM9 · explicit H	1,500	22.96 ± 0.23	93.53 ± 0.19	0.8966 ± 0.0001	94.91 ± 0.11	99.14 ± 0.08
GEOM-Drugs · explicit H	50	99.77 ± 0.02	88.21 ± 0.39	0.8855 ± 0.0004	100.00 ± 0.00	95.22 ± 0.35
GEOM-Drugs · explicit H	100	99.75 ± 0.07	88.30 ± 0.24	0.8848 ± 0.0001	100.00 ± 0.00	95.46 ± 0.31
ChEMBL3D · explicit H	200	99.69 ± 0.05	91.46 ± 0.24	0.8953 ± 0.0002	99.99 ± 0.02	95.80 ± 0.17
ChEMBL3D · explicit H	300	99.75 ± 0.05	91.38 ± 0.16	0.8956 ± 0.0005	100.00 ± 0.00	95.72 ± 0.15
ChEMBL3D · explicit H	400	99.73 ± 0.02	91.21 ± 0.37	0.8953 ± 0.0001	100.00 ± 0.00	95.41 ± 0.16
ChEMBL3D · implicit H	100	99.72 ± 0.04	94.61 ± 0.43	0.9141 ± 0.0002	100.00 ± 0.00	98.60 ± 0.08
ChEMBL3D · implicit H	200	99.64 ± 0.10	95.27 ± 0.16	0.9140 ± 0.0002	100.00 ± 0.01	98.87 ± 0.18
ChEMBL3D · implicit H	300	99.61 ± 0.04	95.32 ± 0.09	0.9142 ± 0.0001	100.00 ± 0.00	98.80 ± 0.07

Novelty is measured against each checkpoint's training split. Novelty and uniqueness use canonical, H-stripped molecular identities with stereochemistry ignored.

Scaffold decoration — GEOM-Drugs, 25 fine-tuning epochs

512 held-out test scaffolds × 8 generation attempts per scaffold per seed (4,096 per seed; 12,288 total), using the fixed released Murcko scaffold panel. The generated coordinates are not relaxed.

Model	Completed epochs	Novelty (%)	PoseBusters validity (%)	Diversity (0–1)	Uniqueness (%)	Connected-sanitization validity (%)
geom-scaffold-murcko	25 (fine-tuning)	97.52 ± 0.10	94.10 ± 0.19	0.9076 ± 0.0004	63.54 ± 0.59	97.99 ± 0.20

PoseBusters protocol: the native scaffold evaluation checks chemistry and distance geometry, excluding the full offline energy and flatness checks. The value above is posebusters/validity, the strict all-check pass rate for that configuration; it does not apply the scaffold-aware inherited-clash exemption. It is therefore not directly comparable to the full offline PoseBusters scores in the unconditional table. Novelty, diversity and uniqueness describe the pooled generated molecules within each seed, across all 512 scaffolds; uniqueness is not averaged per scaffold.

See the scaffold bundle documentation for the saved evaluation runtime and test panel.

Scaffold decoration in a pocket — CrossDocked, 100 fine-tuning epochs

crossdocked-scaffold, evaluated on the published DiffDec single- and multi-decoration test panels, with 100 attempts per condition per seed (43,500 total).

Panel	Conditions	Attempts per seed	Validity (%)	Uniqueness (%)	Similarity (0–1)	Recovery (%)	Vina score (kcal/mol)	High affinity (%)
Single	43	4300	88.72 ± 0.71	53.32 ± 1.02	0.8157 ± 0.0042	39.00 ± 1.15	-8.041 ± 0.007	44.18 ± 3.67
Multi	102	10200	89.80 ± 0.16	66.43 ± 0.38	0.7671 ± 0.0032	37.83 ± 5.33	-8.088 ± 0.010	40.33 ± 3.14

Mean ± sample standard deviation across seeds 42/43/44. The six columns use DiffDec's grouping and formulas; validity and uniqueness have different definitions from the unconditional and GEOM scaffold tables. Docking re-embeds and UFF-minimizes prediction SMILES, so it does not score the original generated coordinates. Invalid predictions are replaced with the scaffold for docking. Reproduction differences include reconstructed reference grouping strings, predicted bond topology, newer RDKit and deterministic embedding seeds.

See the full CSV, including individual seeds and docking failures and protocol and reproduction notes.

Metric definitions for the unconditional and GEOM scaffold tables:

• Novelty: fraction of generated, SMILES-convertible molecules absent from the corresponding training reference, retaining duplicates in the denominator.
• PoseBusters validity: fraction of all generation attempts passing every enabled PoseBusters check; missing checks and failed molecule conversions count as failures.
• Diversity: mean pairwise Tanimoto distance between ECFP6 fingerprints (radius 3, 2,048 bits), retaining duplicates among fingerprintable molecules.
• Uniqueness: distinct canonical molecular identities divided by all SMILES-convertible molecules within a seed.
• Connected-sanitization validity: fraction of all attempts producing connected, RDKit-sanitizable molecules, with no additional charge or radical restriction (connected-sanitization-validity).

Install and authenticate

The runtime bundled with GEOM includes the shared inference interface and catalog discovery for all models in the unconditional checkpoint table. Install it in a Python environment:

python -m pip install 'huggingface_hub>=1.0,<2'
hf auth login
hf download maltefranke/morph \
  --local-dir morph-runtime \
  --include 'geom/explicit_h/runtime/*'
python -m pip install ./morph-runtime/geom/explicit_h/runtime
morph models list

Authentication requires an account with access to this model repository. Python >=3.11 and a compatible PyTorch 2.9 installation are required; CUDA generation needs a CUDA-capable PyTorch build. Dependency constraints and tested versions are included in the runtime. An existing Morph environment with catalog discovery can use the examples directly. On a cluster, run model loading and generation inside a compute allocation.

Load and generate from Python

from morph.inference import load_pretrained

model = load_pretrained("geom-explicit-h", epoch=100, device="cuda")

trajectories, validity, vocab, coordinate_std = model.generate(
    n_mols=64,
    batch_size=16,
    seed=42,
    return_traj=False,
)

Use device="cpu" for CPU inference. The same interface loads QM9 and ChEMBL3D:

model = load_pretrained("qm9-explicit-h", epoch=1500, device="cuda")
# Use epoch=1000 for the earlier QM9 checkpoint.
model = load_pretrained("chembl3d-explicit-h", epoch=400, device="cuda")
# Or choose the heavy-atom model:
model = load_pretrained("chembl3d-implicit-h", epoch=300, device="cuda")

Loading downloads only the selected checkpoint and its six required asset files, then reuses the Hugging Face cache. Each checkpoint is approximately 510 MB. Omitting epoch selects the latest published epoch for the chosen model.

EMA weights are used by default. The inference loader constructs one network and loads only the selected weights; the unselected network and optimizer tensors are not materialized from the memory-mapped checkpoint. To select current training weights, pass use_ema=False to load_pretrained.

Generation defaults to 100 integration steps. Each attempt remains in the returned lists, including invalid molecules. trajectories[i][-1] is its final graph and validity[i] is its RDKit validity result. Pass return_traj=True to retain all integration frames. Returned coordinates are already in Angstrom; do not multiply them by coordinate_std again.

CLI and downloads

Generate GEOM-Drugs molecules and save valid structures to SDF:

morph generate --model geom-explicit-h --epoch 100 \
  --device cuda --batch 64 --batch-size 16 --output geom-e100.sdf

--batch is the total number of generation attempts. Add --save-trajs to save trajectories and validity flags beside the SDF. --no-ema selects current weights, and --steps and --seed control sampling.

Download a portable bundle without loading the model:

morph models download geom-explicit-h --epoch 100 --local-dir morph-models

Load that local bundle without network access or catalog lookup:

model = load_pretrained(
    weights_dir="morph-models/geom/explicit_h",
    epoch=100,
    device="cuda",
)

Use another model name from the table for QM9 or ChEMBL3D. The interface also accepts direct Hub repository IDs with subfolder and revision. Pass cache_dir to choose a cache location, or local_files_only=True to load a previously cached release without network access. The corresponding CLI options are --cache-dir and --offline.

The root sample.py remains available as the archived ChEMBL3D sampling entry point with its corresponding bundled runtimes. The shared interface above supports all four published unconditional model variants.

Reproducibility

Each unconditional model has a training bundle with its original training configuration, launch settings, frozen source, dependency versions and data references. The bundles live alongside the corresponding model files on Hugging Face:

Model	Configuration and source
QM9, explicit H	Training bundle
GEOM-Drugs, explicit H	Training bundle
ChEMBL3D, explicit H	Training bundle
ChEMBL3D, implicit H	Training bundle

For example, download and unpack the GEOM training bundle:

hf download maltefranke/morph geom/explicit_h/training_bundle.tar.gz --local-dir .
tar -xzf geom/explicit_h/training_bundle.tar.gz

Each bundle also includes a portable training recipe; its changes from the original run are recorded alongside it. See the training bundle guide for setup, data requirements and resuming. Version identifiers and checksums are recorded in the release metadata and bundle manifests.

Hydrogen handling and training data

Dataset	Training filter	Representation
QM9	Original QM9 split	Explicit-H variant
GEOM-Drugs	At most 72 total atoms, including H	Explicit-H variant
ChEMBL3D	At most 100 total atoms, including H before hydrogen removal	Explicit-H and implicit-H variants

Both ChEMBL3D variants use the same split membership: 1,474,008 training, 184,295 validation and 184,242 test molecules. The implicit-H model predicts heavy atoms, bonds and formal charges; RDKit infers hydrogens during conversion and validity checking. Its geometry statistics and coordinate normalization were recomputed for the heavy-atom representation.

QM9 uses 100,000 training, 20,000 validation and 12,947 test molecules. Its run started from scratch, validates every epoch and retains checkpoints every 25 epochs.

The GEOM-Drugs epoch 50 and 100 checkpoints come from the run resumed from the best checkpoint after 48 completed epochs. The resumed run validates every epoch and retains checkpoints every fifth epoch. The published unconditional epochs are listed in the catalog. The scaffold-decoration checkpoint adds 25 fine-tuning epochs on top of the epoch-100 model.

Explicit-H models generate H atoms as graph vertices. Every model uses its own ordered vocabularies, distributions and saved configuration; keep those files together. The training atom limits describe dataset filtering and do not impose a hard atom-count limit during generation.

Bundle contents and verification

Each unconditional model directory contains:

• epoch_NNN.ckpt: the published original training checkpoints.
• config.yaml: portable inference configuration with dataset loading disabled.
• atom_tokens.txt, edge_tokens.txt, charge_tokens.txt: ordered vocabularies.
• distributions.pt: original training statistics, including atom counts, coordinate scale, type frequencies and geometry histograms.
• distributions.json: a readable export of those statistics.
• metadata.json: dataset, hydrogen mode, epochs, optimizer steps, original filenames and checkpoint SHA-256 hashes.
• training_config.yaml: the original training configuration for provenance.
• runtime/: archived source, configs and dependency specifications.
• training_bundle.tar.gz: original run configuration, source, environment and a portable training recipe; training_bundle.json records its checksum and provenance.
• validation.json: results of checkpoint loading and sampling checks.

The published unconditional bundles were checked for strict current/EMA weight loading and end-to-end sampling. These functionality checks are separate from the generation benchmarks above. The GEOM and QM9 releases also passed remote file-size and hash verification.

The ChEMBL3D release manifest describes the original ChEMBL3D release, including its historical model card. QM9 and GEOM-Drugs have their own bundle manifests. Original training configs contain historical cluster paths; use config.yaml through the inference loader for portable generation.

Citation

If you use Morph, please cite our paper:

@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},
}

Machine-readable citation metadata is available in CITATION.cff.