MolecularDiffusion.modules.models.ditmc

DiTMC – graph-conditioned Diffusion Transformer conformer generator.

Frank, Ripken, Lied, Mueller, Unke, Chmiela. Sampling 3D Molecular Conformers with Diffusion Transformers. NeurIPS 2025. arXiv:2506.15378. Data + checkpoints: doi:10.5281/zenodo.15489212.

Unlike every other generator in this repo, DiTMC does not invent molecules: you hand it a molecule you already have (atoms and bonds) and it returns 3D conformers of exactly that molecule. No atom types, bonds or charges come out.

A faithful PyTorch port of all three published variants (dit_ape, dit_rpe, dit_so3), including classifier-free guidance, self-conditioning and sampler trajectories. The e3x surface dit_so3 depends on is reimplemented directly in modules/layers/e3x – not substituted with e3nn, whose irrep ordering and normalization would make the published checkpoints unconvertible.

Submodules