Similarity-Driven Regularization for Aligning Chemical and Latent Spaces in Molecular Design
Keywords: Molecular design，Consistency Regularization，Generative model with latent space，Molecular Similarity
Abstract: Generative models play a pivotal role in molecular design by effectively generating target molecules. Among these, generative models with latent space stand out due to their robust latent space representation capabilities, powerful dimensionality reduction ability and controllability of generation. In molecular design applications, generative models with latent space convert input molecules into latent variables, capturing essential molecular features including both structural and property-related characteristics. Ideally, similar molecules should map to proximate latent variables. However, previous studies have shown an inconsistency between molecular similarity in the chemical space and that in the latent space. This inconsistency will impede the accurate representation and complicate subsequent design process,such as leading to higher optimization budget. To address this, we propose Molecular Similarity-Aware Consistency Regularization (MSCR), a straightforward regularization approach aimed at preserving the molecule similarity consistency. Our method proposes a brief but effective regularization technique to align chemical space and latent space,clearly reflect similarity relationships in latent space. We leverage Matched Molecules Pairs (MMPs) to introduce more robust similarity information than other conventional augmentation methods. Extensive experiments demonstrate that MSCR not only maintains molecules pairs similarity but also enhance optimization performance in molecular latent space tasks, without additional costs. Furthermore, our visualizations highlight molecular inconsistencies, thus underscoring the significance of our approach and improving the interpretability and relevance of our work.
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: 13533
Loading