Text-guided Diffusion Model for 3D Molecule Generation

Yanchen Luo, Sihang Li, Zhiyuan Liu, Jiancan Wu, Zhengyi Yang, Xiangnan He, Xiang Wang, Qi Tian

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: Drug design, Generative Models, Diffusion Models

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: This paper introduces TEDMol, a novel approach to 3D molecular generation that leverages text-guided diffusion.

Abstract:

The *de novo* generation of molecules with desired properties is a critical task in fields like biology, chemistry, and drug discovery.
Recent advancements in diffusion models, particularly equivariant diffusion models, have shown promise in generating 3D molecular structures.
However, these models largely work under value guidance, typically conditioning on a single property value, which might limit their ability to address complex real-world requirements.
To address this, we propose the text guidance instead, and introduce TEDMol, a new *Text-guided Diffusion Model for 3D Molecule Generation*.
It aims to integrate the capabilities of language models with diffusion models, thereby providing a deeper level of language understanding in 3D molecule generation.
Specifically, TEDMol utilizes textual conditions to guide the reverse process, enabling the adept and flexible generation of 3D molecules.
Our experimental results on various tasks demonstrate that TEDMol not only enhances the stability and diversity of the generated molecules, but also excels in capturing and utilizing information derived from textual descriptions.
Our approach forms a flexible and efficient text-guided molecular diffusion framework, providing a powerful tool for generating 3D molecular structures in response to complex, textual conditions.

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: 5078