MolCraftDiffusion

One platform for diverse 3D molecular generation workflows.

MolCraftDiffusion is an open-source framework for building, training, applying, and comparing 3D molecular generative models in computational chemistry. It supports de novo, property-directed, structure-guided, shape-conditioned, pocket-conditioned, fragment-based, and pharmacophore-driven molecular design.

The platform unifies data preparation, training and fine-tuning, guided generation, checkpoint handling, and evaluation behind a shared CLI and configuration system. Different generative paradigms can therefore use the same surrounding infrastructure without requiring separate end-to-end codebases. A no-code browser UI, AutomaticMolCraft, is also available for users who prefer not to work from the CLI.

---

Key Features

MolCraftDiffusion is built around a common task interface, allowing generators with different architectures and conditioning inputs to share the platform infrastructure.

• Data Module — Preprocess, compile, and manage raw .xyz files into unified .db (ASE Database) pipelines, and annotate properties.

• Training & Fine-Tuning Module — Train or adapt generative models, property regressors, and time-aware guidance models.

• Broad Generator Coverage — Apply multiple 3D generation paradigms across de novo and conditioned molecular-design tasks.

• Generation & Guidance Module — Generate 3D molecules using a variety of mechanisms:

  • Unconditional Generation: Generate 3D molecules without any specific constraints or guidance.

  • Property-Targeted Guidance: Steer generation towards desired properties using Classifier-Free Guidance (CFG), Gradient Guidance (GG), or a hybrid approach.

  • Structure-Guided Generation: Perform inpainting (scaffold decoration), outpainting (fragment extension) and SILVR (soft reference steering / fragment merging) with precise 3D geometric constraints.

• Analysis & Evaluation Module — Assess generated molecules with structural validity metrics, xTB geometry optimisation, RMSD comparisons, and quantum-chemical property calculation or prediction.

---

Web Interface

Prefer a browser to the CLI? AutomaticMolCraft is a no-code web UI built on top of MolCraftDiffusion — property-guided generation, structure-guided inpainting/outpainting, training configuration, dataset curation, and linked 2D/3D visualization, served locally via dev.sh.

---

Quick Start

# Train a diffusion model
MolCraftDiff train my_config

# Generate molecules
MolCraftDiff generate my_gen_config

# Analyse outputs
MolCraftDiff analyze metrics generated_molecules/

No model of your own yet? The Model Zoo ships pretrained weights, the datasets behind them, and a runnable config for every model on the platform — so you can generate molecules before training anything:

MolCraftDiff zoo list                        # see what is available
MolCraftDiff zoo fetch --model kgdiff
MolCraftDiff generate examples/kgdiff_generate.yaml

Ready-to-use template configuration files for common workflows are listed in Configuration Templates, with the full packaged Hydra defaults under src/MolecularDiffusion/configs/. Copy the relevant file, fill in your paths, and run:

# Example: unconditional generation with the template
cp docs/cfg_examples/gen_unconditional.yaml my_gen.yaml
# edit my_gen.yaml → set chkpt_directory
MolCraftDiff generate my_gen

---

Contents

Getting Started

• Installation
• Model Architectures

Model Zoo

• Model Zoo
• Using a Pretrained Model
• Using Zoo Datasets
• Registering Your Own Model

Tutorials

• Tutorials

Applications

• Applications

Workflows

• Workflows

Configuration Templates

• Configuration Templates

API Reference

• API Reference

---

Citation

If you use MolCraftDiffusion in your research, please cite:

Modular Framework for 3D Molecular Generation in Computational Chemistry Applications, Journal of the American Chemical Society, 2026.

@article{worakul_modular_2026,
	title = {Modular {Framework} for {3D} {Molecular} {Generation} in {Computational} {Chemistry} {Applications}},
	url = {https://pubs.acs.org/doi/10.1021/jacs.5c19960},
	doi = {10.1021/jacs.5c19960},
	journal = {Journal of the American Chemical Society},
	author = {Worakul, Thanapat and Azzouzi, Mohammed and Wodrich, Matthew D. and Corminboeuf, Clémence},
	month = jun,
	year = {2026},
	pages = {jacs.5c19960},
}

Related paper:

A Diffusion Framework for Geometrically Valid and Practically Viable 3D Molecular Generation.

@article{worakul_diffusion_2026,
	title = {A {Diffusion} {Framework} for {Geometrically} {Valid} and {Practically} {Viable} {3D} {Molecular} {Generation}},
	url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15005231/v1},
	doi = {10.26434/chemrxiv.15005231/v1},
	publisher = {American Chemical Society (ACS)},
	author = {Worakul, Thanapat and Corminboeuf, Clémence},
	month = jun,
	year = {2026},
}