Reproduce “MolGrapher: Graph-based Visual Recognition of Chemical Structures” on OpenResearch

About this paper

Lucas Morin, Martin Danelljan, Maria Isabel Agea, Ahmed Nassar, Valery Weber, Ingmar Meijer, Peter Staar, Fisher Yu

MolGrapher introduces a graph-based deep learning system for Optical Chemical Structure Recognition, reconstructing chemical structures from images by classifying atoms and bonds within a supergraph. The system achieves state-of-the-art accuracy on various benchmarks, demonstrating robustness to complex molecules and diverse drawing styles using a novel synthetic data generation pipeline and introducing the USPTO-30K dataset.

Key insights

• Explicitly leveraging the molecular graph structure with a GNN, instead of sequence-based approaches, significantly improves accuracy and robustness, especially for complex chemical structures.
• Extensive multi-level synthetic data augmentation can effectively train high-performing deep learning models for OCSR, eliminating the need for vast amounts of real-world annotations.
• The direct prediction of a graph allows for the detection of invalid structures, providing an inherent mechanism to flag potential errors and enhance the trustworthiness of automated extraction.

Summary by alphaXiv.