Topological Representations Enhance Molecular Learning

ICLR 2025 Workshop LMRL Submission109 Authors

Track: Full Paper Track

Keywords: topological deep learning, molecular representation, oversmoothing, molecular property prediction

TL;DR: We implement and benchmark topological deep learning methods for molecular representations showing enhance results over graph learning counterparts

Abstract:

Molecular representation learning underpins computational chemistry and drug discovery, yet standard graph-based approaches struggle with oversmoothing and limited long-range interaction modeling. We explore topological deep learning (TDL) as an alternative, leveraging hypergraphs and cell complexes to incorporate higher-order molecular structures. By systematically comparing these representations against graph-based models, we evaluate their capacity to mitigate oversmoothing and capture richer molecular features. Our empirical analysis across QM9 and ZINC benchmarks demonstrates that topological representations enhance predictive performance, particularly in complex molecular graphs. These findings highlight the potential of TDL for more expressive and structurally aware molecular learning frameworks.

Submission Number: 109