Controllable Molecule Generation by Sampling in Continuous Parameter Space
Keywords: Molecular generation; bayesian flow networks
Abstract: Deep generative models have made significant strides for continuous data generation, such as producing realistic images and 3D protein conformations. However, due to the sensitivity of topological graphs to noise and the constraints of long-range discrete relationships, the generation of purely discrete data—such as topological graphs—remains a long-standing challenge, with property control proving even more elusive. In this paper, we propose a novel molecular graph generative framework, called CtrlMol, to learn the topological graphs of molecules in a differentiable parameter space. Unlike diffusion models that iteratively refine samples, CtrlMol optimizes distribution parameters at different noise levels through a pre-defined Bayesian flow. At each of the sampling step, we leverage a property guided output distribution to have a fine-grained control of the topological structures toward the given property. Experimental results demonstrate CtrlMol outperforms all the competing baselines in generating natural molecule graphs. In addition, CtrlMol advances the state of the art in producing the molecules with the desired properties.
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13653
Loading