TopoGuide: A Finetuning Framework for Topologically-Consistent 3D Molecule Generation

ICLR 2026 Conference Submission25260 Authors

Keywords: Molecule Generation, Diffusion Models, Equivariant Neural Networks, Drug Discovery

Abstract: Equivariant diffusion models can generate high-quality 3D molecular geometries but often struggle with chemical validity due to a lack of explicit guidance from the 2D molecular graph. While prior works have addressed this by adding graph-based information to the model's input, this often increases architectural complexity and slows inference. We propose a new finetuning framework that instills 2D topological awareness into pre-trained 3D generative models without altering their core architecture. Our method enforces consistency between the representations of a target 2D graph and a generated 3D structure within a shared embedding space, guided by a consistency loss. By applying our framework to state-of-the-art models, we demonstrate a significant improvement in topological accuracy and chemical validity while preserving the original model's high-quality geometry and inference efficiency.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 25260