MiCRM: A Benchmark Dataset for Mechanistic Molecules

Optical chemical structure recognition (OCSR) systems aim to understand such diagrams computationally. However, developing methods for complex reaction schemes remains an outstanding challenge due to a lack of annotated datasets mapping molecular images and their identity in SMILES. We introduce MiCRM, Molecules in Chemical Reaction Mechanisms, a dataset designed to benchmark machine recognition capabilities of chemical molecules within named reaction mechanisms. Comprising 454 images, it spans a broad array of organic reactions, each illustrated with molecular structures and mechanistic arrows. MiCRM addresses this gap by offering a rich collection of annotated molecule images, enhancing the benchmarking process for OCSR methods. This dataset not only includes a machine-readable molecular identity for each image but also integrates mechanistic arrows, adding a layer of complexity absent in previous datasets. By doing so, MiCRM provides a more authentic and challenging framework for testing molecular recognition technologies, fostering advancements in accurately interpreting complex chemical reaction mechanisms.