3D Autoencoding Diffusion Model for Molecule Interpolation and Manipulation
Primary Area: generative models
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.
Keywords: diffusion models, 3D molecule optimization, controllable generation, equivariant GNN
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Manipulating known molecules and interpolating between them is useful for many applications in drug design and protein engineering, where exploration around the molecular templates is involved. Recent studies using equivariant diffusion models have made significant progress in the de novo generation of high-quality molecules, but using these models to directly manipulate a specified template remains less explored. This is mainly due to an intrinsic property of diffusion models: the lack of a latent semantic space that is easy to operate on. To address this issue, we propose the first semantics-guided equivariant diffusion model that leverages the “semantic” embedding of a 3D molecule, learned from an auxiliary encoder, to control the generative denoising process. By modifying the embedding, we can steer the generation towards another specified molecule or a desired molecular property. We show that our model can effectively manipulate basic chemical properties, outperforming several baselines. We further verify that our approach can achieve smoother interpolation between 3D molecular pairs compared to standard diffusion models.
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: 6828
Loading