Enhancing the Applicability of the Equivariant Diffusion Model to Diverse Molecule Datasets

Minju Na, Jumin Lee, Chaeyoung Moon

Keywords: Equivariant Diffusion Model (EDM), Molecular Generation, Deep Learning

TL;DR: Project applies bit diffusion to an equivariant diffusion model for molecule generation and evaluates its generalizability on larger 3D small molecule dataset and other biomolecules and inorganic materials.

Abstract: The success of deep learning-based molecular generation has been well-documented in various fields, but generating 3D molecular structures has been challenging due to their complex geometric symmetries. Recent developments in geometric deep learning have opened up new possibilities for 3D molecular representation and diffusion-based models for molecular generation. This thesis focuses on extending the Equivariant Diffusion for Molecule Generation in 3D (EDMs) model, which generates equivariant molecular structures that retain their properties irrespective of their position or orientation in 3D space. We propose using an equivariate graph neural network (GNN) for categorical data with bit diffusion and evaluate its performance on a larger three-dimensional small molecules dataset. Additionally, we explore the generalizability of our proposed models to other types of datasets such as dimers, dipeptides, solvated amino acids, and inorganic materials. Our results demonstrate the potential of deep learning-based molecular generation for various applications in materials science and drug discovery.