MarkitS: an image-to-SMILES parsing workflow for Markush structures

Two-dimensional image representations of chemical structure are common in patents and academic literature. Converting these images to a computer-readable format is crucial for downstream applications such as compound searching, comparison, and molecular modeling. Existing optical chemical structure recognition (OCSR) tools perform reasonably well on standard structures. However, for Markush structures they are generally ineffective. Here, we introduce MarkitS (Markush image-to-SMILES), a novel image-to-SMILES parsing workflow specifically designed to convert 2D Markush structures to SMILES representations. Built on a basal SMILES prediction model, MarkitS integrates three innovative correction modules to process Markush representations. Across three independently prepared datasets of 2D Markush structure images, MarkitS yields 85 \(~\) 90% SMILES prediction accuracy and delivers an approximately eight-fold improvement over widely used Markush-compatible tools, MolScribe and DECIMER. MarkitS also far outperforms the flagship vision-language model GPT-5, which apparently has not been trained on Markush structure data. In addition, MarkitS exhibits an extraction rate of SMILES from chemical images comparable to that of SureChEMBL, an industrial-grade baseline. For US patents in SureChEMBL, Markits also yields a consistently high accuracy of \(~\) 90%. MarkitS provides a reliable solution for converting Markush structure images to a computer-readable format for downstream applications. Scientific contribution : We introduce a novel workflow specifically for transforming Markush chemical structure images to the SMILES format. Compared with GPT-5 and existing methods capable of handling Markush representations, our workflow achieves a remarkably higher success rate in generating complete and topology-faithful SMILES. This approach provides a practical solution for transforming variable-group–rich Markush chemical structures commonly found in patents and scientific literature into machine-readable SMILES, thereby facilitating downstream applications such as chemical structure comparison and patent searching.