MolecularDiffusion.modules.models.loqi

LoQI: low-energy molecular conformer generation on a fixed 2D graph.

Ported from https://github.com/…/LoQI (NVIDIA / Megalodon lineage, Apache-2.0). Unlike every other generative model in this package, LoQI does not invent molecules: it takes a molecule you already have – atom types, formal charges, bond orders and stereochemistry – and denoises only its 3D coordinates. Two variants share the same backbone:

  • diffusion_loqi – VDM diffusion, 25 discrete steps, self-conditioned

  • diffusion_loqi_flow – continuous flow matching, velocity prediction, rigid (Kabsch) optimal transport, no self-conditioning

References

Nikitin, Anstine, Zubatyuk, Paliwal, Isayev, “Scalable Low-Energy Molecular Conformer Generation with Quantum Mechanical Accuracy” (2025), DOI 10.26434/chemrxiv-2025-k4h7v.

Reidenbach, Nikitin, Isayev, Paliwal, “Applications of Modular Co-Design for De Novo 3D Molecule Generation” (Megalodon), arXiv:2505.18392 (2025).

Submodules