Integrating the Expression and Discrimination via Bilateral Compensation for Molecular Property Prediction

Jiahao Chen, Yurou Liu, Bin Qin, Jiangmeng Li, Bing Su

Keywords: Molecular property prediction, Self-supervised learning

Abstract:

Predicting molecular properties plays an important role in both scientific research and industrial applications. Given that different molecular properties are influenced by specific atoms or functional groups, it is essential to incorporate both types of information. Previous approaches either leverage subgraph information in self-supervised learning to pre-train atom-based architectures or develop subgraph-based architectures tailored to specific downstream tasks. However, these methods often lack a thorough analysis or theoretical support concerning the expressive capabilities of these two types of representations. Moreover, they typically rely on fixed coupling representations, which cannot adaptively prioritize more discriminative information for various downstream tasks.
In this paper, we introduce a Route-guided Bilateral Compensation (RBC) architecture that explicitly extracts atom-wise and subgraph-wise information through two decoupled branches and integrates them via a route module. Theoretically, we demonstrate that our decomposition-polymerization subgraph-wise branch exhibits greater expressive power than the atom-wise branch, and that the integration process reduces the generalization error bound. Furthermore, we propose a coordinated self-supervised learning strategy that incorporates node-level masked graph reconstruction tasks for atomic and lexicalized subgraph tokens, alongside a graph-level contrastive learning task. For different downstream tasks, the route module facilitates dynamic integration, enhancing the discriminative power of the final representation. External experiments verify the effectiveness of our method.

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: 2139