Rethinking Diffusion Models with Symmetries through Canonicalization with Applications to Molecular Graph Generation
Image generated by Gemini AI
Researchers propose a new approach to generative tasks in chemistry, moving away from traditional invariant and equivariant models. They introduce a canonicalization method that simplifies training and enhances performance by mapping samples to a standardized form before applying non-equivariant models. This framework, tested on molecular graph generation with $S_n \times SE(3)$ symmetries, outperforms existing models, particularly in 3D molecule generation, demonstrating state-of-the-art results on the GEOM-DRUG dataset.
New Approach to Diffusion Models Enhances Molecular Graph Generation
A recent study introduces a novel method for molecular graph generation by employing a canonicalization approach that leverages group symmetries. This method demonstrates improved efficiency and performance over traditional strategies.
Traditionally, generative models have relied on equivariant denoisers to handle distributions invariant to group symmetries. The latest research proposes a three-step process: mapping samples to an orbit representative, training a non-equivariant diffusion model on this canonical slice, and recovering the invariant distribution through random symmetry transforms.
Key Findings
- The correctness and universality of canonical generative models, which outperform traditional invariant targets.
- Increased expressivity of these models, leading to enhanced training efficiencies.
- Training acceleration through canonicalization, reducing complexity associated with group mixtures.
Applications in Molecular Graph Generation
The authors implemented this framework in molecular graph generation under the symmetries of \(S_n \times SE(3)\). Their method, Canon, significantly outperformed existing equivariant baselines in 3D molecule generation tasks with comparable or reduced computational demands.
CanonFlow achieved state-of-the-art performance on the GEOM-DRUG dataset, showing advantages even in few-step generation scenarios.
Related Topics:
📰 Original Source: https://arxiv.org/abs/2602.15022v1
All rights and credit belong to the original publisher.