Equivariant Diffusion Models

Overview

Equivariant Diffusion Models is the official code release for the paper Equivariant Diffusion for Molecule Generation in 3D. The repository implements an E(3)-equivariant diffusion model, referred to as EDM, and documents research workflows for training molecular generators and examining their outputs. It is a model implementation with training and evaluation scripts, rather than a hosted molecular-design service.

The documented workflows use QM9 or GEOM-Drugs data for training, then pass a saved model directory to separate scripts for sample-quality analysis and molecular visualization. For GEOM-Drugs, the README directs users to additional dataset-preparation instructions. Conditional QM9 generation accepts a selected molecular property and supports sampling across different property values. The listed conditioning choices are alpha, gap, homo, lumo, mu, and Cv. The README also identifies paths for a pretrained generator and property classifier for alpha.

A separate property-prediction workflow trains an EGNN classifier or a baseline using only the number of nodes, then evaluates a classifier on EDM-generated samples. These routines support investigating generated molecules, but the source excerpts do not report performance results or establish suitability for downstream experimental use. Resource planning matters: the authors warn that fully connected EGNN message passing can be memory-intensive, particularly in the GEOM-Drugs workflow. RDKit is described as optional, while the excerpts do not establish a complete, tested environment configuration.

Key Features

• Separate documented training workflows for EDM on QM9 and GEOM-Drugs.
• Post-training scripts for molecular sample-quality analysis and visualization using a saved model directory.
• Conditional QM9 generation using `alpha`, `gap`, `homo`, `lumo`, `mu`, or `Cv`, including sampling across property values.
• Property-classifier training with an EGNN model or a node-count-only baseline, followed by evaluation on EDM-generated samples.
• Documented paths for a pretrained `alpha`-conditioned generator and an associated pretrained property classifier.

Use Cases

• Suggested evaluation: train an EDM on QM9 and inspect generated 3D molecules using the documented analysis and visualization workflows.
• Suggested evaluation: investigate how generated samples change across conditioning values for a selected supported QM9 property.
• Suggested evaluation: assess property-conditioned samples with the documented classifier workflow and investigate the node-count-only baseline.
• Suggested evaluation: explore GEOM-Drugs generation while assessing the memory requirements of fully connected EGNN message passing.

How to Use

1. Start with the official README and choose the QM9, GEOM-Drugs, or conditional QM9 workflow. Use its supplied examples rather than assuming one configuration covers all three.
2. Inspect the package setup source for declared dependencies. The README offers optional RDKit environment guidance; these excerpts do not establish a complete installation procedure.
3. Prepare the selected dataset. For GEOM-Drugs, follow the README's reference to data/geom/README.md in the repository. That file's contents are not supplied here.
4. Follow the relevant training example and retain its model output directory. Account for the documented memory warning when planning GEOM-Drugs runs.
5. Use the documented analysis and sampling scripts with the saved model directory. For conditional generation, select a supported property and follow the property-sweep example.
6. If evaluating conditioning, follow the classifier-training and generated-sample evaluation examples. Treat resulting measurements as your own evaluation, not as performance established by this entry.