Fragment-Augmented Diffusion for Molecular Conformation Generation

Xiaozhuang Song, YUZHAO TU, Tianshu Yu

Keywords: Molecular Conformation Generation, Data Augmentation

Abstract:

Molecular conformer generation is a fundamental challenge in computational chemistry, particularly for large and complex molecules.
In this work, we propose a novel approach called Fragment-Augmented Diffusion (FADiff), which integrates molecular fragmentations into diffusion models as a data augmentation strategy to enhance molecular conformation generation. By decomposing molecules into smaller, manageable fragments for the purpose of data augmentation, FADiff enhances the diffusion generation process, effectively capturing local structural variations while preserving the integrity of the entire molecule. Extensive experiments across multiple datasets demonstrate that FADiff consistently outperforms state-of-the-art methods, particularly in data-scarce scenarios, where the fragment-based augmentation approach significantly enhances model performance. We also provide a comprehensive analysis of different fragmentation rules and their impact on model performance, and theoretically validate FADiff's effectiveness in improving generalization. Overall, FADiff advances molecular conformation generation by enhancing the exploration of conformational space, offering a powerful tool for computational chemistry. The code is available at https://anonymous.4open.science/r/fragaug-5960/.

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.

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: 2082