Understanding Markush Structures in Chemistry Documents With Deep Learning

The retrieval of chemistry documents depends on the recognition of chemical structural formula (CSF) except texts, tables, and ordinary figures. A Markush structure represents a series of chemicals with similar structures. This research proposes a complete approach to analyze the Markush structures, including a deep-learning-based detection of CSFs, recognition of Markush structures and their corresponding tables, and reconstruction of the synthetic CSFs represented by one Markush structure. The results of experiments demonstrate that the proposed method obtains CSFs accurately from the original chemistry documents and provides a more efficient way with the automatically reconstructed molecules than the existing manually-drawing-based approach for further retrieval of Markush structures.