Mining Large-Scale Knowledge Graphs for Chemical Reaction Fingerprints

• DOI: 10.1109/bigdata59044.2023.10386300
• Publication Date: 2024-01-22T18:28:47Z
• Authors: Blake B. Gaines; Minghu Song; Jinbo Bi

Knowledge graphs have become a popular method for representing large, relational data. Similar to citation networks and social networks, relationships in chemical reaction data can also be uniquely captured using knowledge graph. However, relatively few studies exist concerning the application of graph mining techniques numerical representation reactions. In this study, we develop pipeline transforming large-scale databases reactions into heterogeneous graphs, which their reactants products are all characterized as nodes with connecting edges. We create templates, each links multiple enhance connectivity graph, then employ learning methods (Node2Vec RotatE) generate an embedding (or fingerprint) node. To evaluate efficacy method, construct classifiers label mechanisms based on these fingerprints. Experimental results show that our approach outperforms state-of-the-art fingerprints, specifically when class labels not available during process. When representations fine-tuned subsequent classification task, achieves comparable accuracy recent Transformer-based algorithm, but significantly lower computational cost.