Knowledge graph information bottleneck enhanced molecular representation learning

Effective molecular representation learning (MRL) is essential for advancing molecular property prediction. In recent years, graph-based MRL methods have made significant progress by effectively utilizing the topology structure of molecules. Researchers have started to explore the integration of fundamental domain knowledge into semantic graphs using knowledge graphs (KGs), which provide valuable priori chemical information to enhance MRL performance. However, external knowledge from the real world may contain considerable redundant and noisy information for the tasks. Previous efforts have primarily focused on integrating KG and molecular structures into a unified framework, with few assessing the minimal sufficiency of knowledge. In this work, we propose a novel knowledge graph information bottleneck (KGIB) enhanced molecular representation learning framework consisting of a knowledge compression module and a knowledge alignment module. We construct a molecular knowledge graph (MKG) to describe the "molecule-functional group-element" hierarchy relationship. The knowledge compression module recognizes the predictably compressed minimal sufficient knowledge subgraph for each molecule in MKG, ensuring the preservation of essential information while effectively minimizing irrelevant knowledge noise for downstream tasks. The knowledge alignment module captures complex patterns and dependencies between topology structures and semantic information, thereby enriching the semantic relevance of molecular representation. We compare KGIB with various state-of-the-art baselines across ten real-world tasks and demonstrate that it consistently outperforms them on six tasks while achieving highly competitive results on the remaining four tasks, showcasing the superiority of our method.