Model Architectures¶
MolCraftDiffusion ships several backbone families as first-class options. You
select one by choosing a task config (MolCraftDiff train <config>); the
config’s _target_ factory builds the matching model, so adding or swapping an
architecture never touches the core engine.
Available architectures¶
Task config ( |
|
Model |
Notes |
|---|---|---|---|
|
|
EDM (E(n)-equivariant diffusion, EGCL backbone) |
Default. Cartesian-space DDPM; the checkpoints on Hugging Face use this. |
|
|
EDM + extra atom features |
Same backbone, feature-aware conditioning. |
|
|
EDM |
Fine-tuning from a pretrained EDM checkpoint. |
|
|
EGT (equivariant graph transformer) |
Transformer backbone via |
|
|
GFMDiff |
Geometric full-molecule diffusion ( |
|
|
TABASCO |
Flow-matching architecture. |
|
|
FlowMol (SE(3)-equivariant GVP) |
Flow matching for coordinates, atom types, and formal charges. This integration considers the bond-free variant only: bonds are not modeled or generated; graph edges are used only for geometric message passing. |
|
|
ADiT / DiT-based LDM |
Latent diffusion with a DiT denoiser. |
|
|
DiffLinker |
Fragment linking / linker design. |
|
|
GeoLDM VAE (transformer enc/dec) |
Trains the autoencoder for latent diffusion. |
|
|
GeoLDM VAE (Equiformer enc/dec) |
Equivariant autoencoder variant. |
|
|
Pharmacophore-conditioned dynamics |
Requires |
|
|
Self-supervised 3D pretraining |
Prefix-matched to the |
|
|
Property predictor / guidance head (EGCL backbone) |
Default. Used for property prediction and gradient guidance. |
|
|
Property predictor / guidance head (eSEN backbone) |
Same task classes as above, eSEN backbone. |
|
|
Property predictor (EquiformerV2 backbone) |
No |
ShEPhERD is integrated as a scoring/architecture module (modules/models/shepherd_arch/,
utils/shepherd_score/) used by the guidance and analyze metrics --metrics shepherd
paths rather than as a standalone training config.
References¶
Backbones and objectives integrated here are based on the following work.
EDM — Hoogeboom, Satorras, Vignac & Welling. Equivariant Diffusion for Molecule Generation in 3D. ICML 2022. arXiv:2203.17003
EGT — Vignac et al. MiDi: Mixed Graph and 3D Denoising Diffusion for Molecule Generation. ECML PKDD 2023. arXiv:2302.09048
GFMDiff — Xu et al. Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation. AAAI 2024. arXiv:2401.02683
TABASCO — Vonessen, Harris, Cretu & Liò. TABASCO: A Fast, Simplified Model for Molecular Generation with Improved Physical Quality. 2025. arXiv:2507.00899
FlowMol — Dunn & Koes. Mixed Continuous and Categorical Flow Matching for 3D De Novo Molecule Generation. 2024. arXiv:2404.19739
ADiT — Joshi et al. All-atom Diffusion Transformers: Unified generative modelling of molecules and materials. 2025. arXiv:2503.03965 — built on the DiT backbone of Peebles & Xie, Scalable Diffusion Models with Transformers, ICCV 2023 (arXiv:2212.09748).
DiffLinker — Igashov et al. Equivariant 3D-Conditional Diffusion Model for Molecular Linker Design. Nature Machine Intelligence 2024. arXiv:2210.05274
GeoLDM — Xu, Powers, Dror, Ermon & Leskovec. Geometric Latent Diffusion Models for 3D Molecule Generation. ICML 2023. arXiv:2305.01140
eSEN — Fu et al. Learning Smooth and Expressive Interatomic Potentials for Physical Property Prediction. 2025. arXiv:2502.12147
EquiformerV2 — Liao, Wood, Das & Smidt. EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations. ICLR 2024. arXiv:2306.12059
ShEPhERD — Adams, Abeywardane, Fromer & Coley. ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design. ICLR 2025. arXiv:2411.04130