Equivariant Diffusion for Molecule Generation in 3D

TL;DR

This paper presents an E(3) equivariant diffusion model for 3D molecule generation that improves sample quality and training efficiency by jointly modeling atom coordinates and types with an equivariant network.

Contribution

It introduces a novel equivariant diffusion model that handles both continuous and categorical features for 3D molecule generation, with likelihood computation capabilities.

Findings

Outperforms previous methods in sample quality

Achieves higher training efficiency

Provides probabilistic likelihood analysis

Abstract

This work introduces a diffusion model for molecule generation in 3D that is equivariant to Euclidean transformations. Our E(3) Equivariant Diffusion Model (EDM) learns to denoise a diffusion process with an equivariant network that jointly operates on both continuous (atom coordinates) and categorical features (atom types). In addition, we provide a probabilistic analysis which admits likelihood computation of molecules using our model. Experimentally, the proposed method significantly outperforms previous 3D molecular generative methods regarding the quality of generated samples and efficiency at training time.