Skip to yearly menu bar Skip to main content


Poster
in
Workshop: AI for Science: Scaling in AI for Scientific Discovery

Energy-Free Guidance of Geometric Diffusion Models for 3D Molecule Inverse Design

Aksh Garg · Jiaqi Han · Sanjay Nagaraj · Minkai Xu · Stefano Ermon

Keywords: [ SE(3) Equivariance ] [ Inverse Molecular Design ] [ Classifier-Free Guidance ] [ Geometric Diffusion Models ] [ Molecular Generation ]


Abstract:

Molecule inverse design is of critical significance in drug discovery which requires molecules to be generated based on certain chemical properties or structural compositions. Generative models, most popularly diffusion models, have shown great promise in performing inverse design through conditioning techniques and/or explicit energy guidance during sampling. In this work, we propose a novel guidance framework, Energy-Free Guidance for Geometric Diffusion Models (EFG-GDM), that effectively boosts the utility of molecule inverse design without any auxiliary energy head for guidance. The key innovation lies in the joint training strategy for the conditional and unconditional score models via random masking, which are then composed during sampling in an SE(3)-equivariant fashion, ensuring the critical physical symmetry of the geometric distribution. This feature alleviates practitioners from needing additional efforts in training energy prediction heads and avoids the adversarial gradient coming from them. We conduct experiments on a diverse range of inverse design tasks on QM9, showing that our approach achieves state-of-the-art on 4 out of 6 design targets without leveraging any external energy gradients.

Chat is not available.