Abstract
The automated extraction of chemical reaction data from scientific literature remains very challenging mainly due to the visual complexity and heterogeneity of published graphical schemes. In particular, substrate scopes containing positional Markush structures pose a significant challenge for current Optical Chemical Structure Recognition (OCSR) and reaction-extraction systems, as their resolution requires combining molecular recognition with precise spatial reasoning. Herein, we introduce a deterministic, coordinate-based extraction pipeline designed to resolve positional Markush structures in reaction schemes. Building upon a fine-tuned version of the graph-prediction OCSR module MolScribe, our approach uses deterministic geometric rules based on 2D molecular coordinates to infer R-group attachment points and generate topologically valid molecular representations. To evaluate this approach, we introduce two curated benchmarks: MIPub-2k (Molecular Images from Publication - Université Bourgogne Europe) comprising 2,128 literature-derived molecular image crops for OCSR evaluation and MIReactPub (Markush template-based Image of Reaction from Publication – Université Bourgogne Europe) comprising 450 reactions from 23 schemes containing positional Markush structures for reaction-level assessment. The complete pipeline achieves an end-to-end F1 score of 0.46, increasing to 0.54 for schemes without textual-dependent ambiguities. Slightly reorganizing figures doubles the F1 score, demonstrating the robustness of the underlying geometric resolution while highlighting the substantial impact of visual heterogeneity in published literature on end-to-end performance. By explicitly encoding chemical and spatial constraints, our proposed approach provides a transparent and reproducible framework for resolving complex molecular representations from literature images. The resulting open-source pipeline and benchmarks provide a specialized component that can be further integrated into automated workflows for large-scale chemical information extraction.